Q: Is the continuous growth of AI-generated content online a danger for journalism?

It is certainly a problem yes, in three ways: it makes reporting harder, it makes it harder to support journalism financially, and it makes it harder for audiences to trust your reporting.

It makes reporting harder because a journalist has to deal not only with a lot more misinformation but also a lot more information full stop.

It makes the business model more difficult because there’s more competition from AI-generated slop.

And that’s compounded by the fact that audiences are less likely to trust your reporting — either because they might dismiss it as AI generated, or because they might have already begun to believe AI misinformation that your reporting contradicts.

Q: Is the industry equipped to handle it?

I don’t think any industry is equipped to handle how AI is changing the information environment, and I think it will take some time for society to adjust to a world of unlimited cheap and unreliable information.

But journalism is probably in a better position than most industries to handle that change because it has rules and workflows that a lot of other industries do not.

For example, journalists are trained to research both sides of the story and seek a right of reply, and to find more than one source for key facts. A high value is put on speaking to real people and capturing raw footage of events. Reporters are encouraged to be sceptical and apply verification and factchecking skills, and their stories go through editors who are supposed to look at the story dispassionately. These are not qualities that all industries have.

Also, the industry has been exploring applications of artificial intelligence for over a decade, so there’s technical knowledge there too which many organisations will lack.

The biggest challenge is the collapse of the financial basis for journalism and I think there will have to come a point where more public or charitable funding is used to support it.

Q: When AI can fabricate convincing news content at scale, who is ultimately responsible for the integrity of what we read and watch online?

Primarily it’s whoever created that content. When the printing press came along someone might have asked who is responsible for these books that are being published? And of course we know the answer now is the authors and the publishers.

But technology companies are also responsible, because AI is not a neutral tool like a printing press.

ChatGPT is a collection of recipes for predicting text, and so OpenAI, which owns ChatGPT, has responsibility for the way it has written those recipes because it can change them. It has changed them, and does change them, regularly.

If you try to create racist or other offensive content in ChatGPT or Gemini or Claude, for example, it will generally refuse to do so. If you create images in Gemini a watermark is added, so it can be identified as AI generated. If you ask Claude for medical advice it will recommend consulting a health professional. These are editorial choices by the technology company.

So while individual crimes involving fabrication might be tackled by identifying and charging the creator, the ultimate responsibility for any systemic social harm will lie with technology companies who haven’t considered key risks and taken steps to prevent that. And there will need to be regulation and enforcement to ensure that happens.

Q: What does this mean for the public’s trust in journalism?

We have always had to earn that trust, and that has become more and more important as competition for the public’s attention has increased.

Part of that trust is about building understanding: not only explaining how we actually do journalism, and why we do it that way, but also listening to our audiences to understand what they need and why.

In order for the public to trust journalism they have to see how it is distinct from other sources of power, and holds that power to account. They have to see how it is on their side, rather than looking down on them.

Q: When fabricated content goes viral, what happens to people’s ability to make informed decisions?

I think we already had a problem with people being able to make informed decisions, as it became easier to find information that confirmed a person’s existing beliefs rather than information that challenged those.

Ultimately the more misinformation there is, the more likely it is that people are either making decisions based on flawed information, or unable to make decisions because they can’t be 100% sure a certain piece of information is true.

I think we are going to learn some quite painful lessons in the next few years because most people think they won’t fall for this stuff, and yet most people do. I am an intelligent person who trains people in verification and factchecking, and I have retweeted information that is not true.

There are two key things that we need to remember: firstly, that you will fall for misinformation at some point, and probably already have, that is certain.

Secondly, slow down and look for counter-evidence.

Q: For an ordinary person watching the news or scrolling their feed, what practical steps can they take right now to check the veracity of what they are seeing?

There’s a great piece of research in psychology that finds when we look at an image of a crowd of faces, we pay attention to the angry faces more: our eyes spend more time on those.

Our brains are designed to pay more attention to threats, and social media algorithms have learned this, so they prioritise material that triggers anger and fear because we spend more time looking at it, and are more likely to react to it.

So the first practical step people can take is to slow down. When we come across new information two parts of our brain processes it: the first part is fast and instinctive. It makes a decision whether to pay attention and what to do with the information. That’s the part that shares or likes or comments on a chat message or video.

Don’t share. Don’t like. Don’t react.

The key thing is to let the information move through to a second phase, the part of the brain that actually processes that information rationally.

Now, ask yourself if that update is triggering some core emotion. Is it confirming something you believed? Is it making you angry or afraid? Why? The short answer is: because it keeps you on that site or app for longer. But is it true?

The main way to check veracity isn’t to look for clues that something is fake. An image or video might be real but from a different time or place. Something might be factually true but misleading because it’s missing context. Trusted friends and authority figures will share this stuff, so don’t use that as a primary signal.

The main way to check veracity is to look for evidence that challenges it.

For example you can use a website called TinEye to see where an image has appeared before (you can also freeze frame videos and take screenshots). You can put the information into Google with the word ‘hoax’ or ‘factcheck’ to see if it’s already been debunked.

You can try to identify the original source, rather than whoever passed it on to you. That might give you clues about their independence, expertise or agenda.

Q: And people can use AI tools to verify as well?

Using AI to verify material is very dangerous. We’ve seen a big increase in the last few years of people using AI tools to check if an image or a video is real, and very often they get it wrong.

Part of the reason is that people ask the wrong question. If you ask AI “is this fake?” then you have to remember tools like ChatGPT and Grok are sycophantic, they are designed to please, so they will say “yes” even if they’re not certain. They are also overconfident.

The key thing with AI in general is to always ask a neutral question. Ask it what evidence there might be to help you identify whether something is true or false. Ask it what methods you can use.

AI is best used to challenge you and open up other ideas and techniques. Don’t use it to confirm things because it will just tell you what you want to hear.

Never rely on AI’s answer as the end of a process – it can only ever point you to next steps. Remember that it is not a factual tool, it is a language prediction tool. So it can predict the language of a piece of advice around verifying something, but it cannot know if something is real or fake, because that’s not what AI is designed to do.

From stories about political speech and song lyrics, to street names and social media chatter, data journalists now have a wide range of examples of text-as-data to draw inspiration and guidance from, while tools such as Pinpoint and NotebookLM are making text analysis easier than ever.

I compiled a list of over 200 pieces of data journalism where text or documents were used as sources. Quantification techniques ranged from counting the frequency of a single word and using Google’s ngram viewer, to machine learning and topic modelling.

Looking at those articles it’s clear that, once quantified, journalists tell the same stories about text as any other piece of data: using the seven most common angles.

But how those angles are used — and how often — is where it gets interesting…

Text-based data journalism most often uses an exploratory feature format

While most data stories about numerical data focused on scale or change, the data stories about text I looked at were overwhelmingly dominated by exploratory formats.

There are two obvious possible explanations for this. First, analysing text normally requires more time and skill than working with numerical data. That would be hard to justify for a simpler news article revealing scale or change.

Second, text is often rich and complex, lending itself more to exploration. The analysis itself will often require explanation too, especially if it involves classification.

The Pudding‘s scrollytell data feature on speeches in Parliament is a good example of an exploratory approach: text from over a million speeches is classified in two ways: by the gender of the speaker, and by the topic it relates to. It is quantified in terms of the percentage of time that speakers spend on each topic (code is shared in a GitHub repo).

The story explores the data through a number of themes: the economy, welfare, education, and so on. Although the focus is largely on variation between men and women, this is secondary to the primary exploratory angle.

The same thematic approach is adopted by South African data journalism website The Outlier in #SONA2022: How does it compare to the last 5 speeches by Ramaphosa? and by Sueddeutsche Zeitung in their feature on why Russians perceive the West as a threat.

Another way to structure an exploratory feature about text is by breaking the feature up around different questions. This is the approach taken by Quartz in The hidden structure of the Apple keynote (“Who’s on stage?”, “Who’s the funniest?” and “When is the unveil?”)

Stories about changes in language

Text can act as a proxy for cultural attitudes and fashions, or political focus, so it’s not surprising that many stories based on data analysis focus on changes in society or power that language can reveal.

Text from social media and forums can be analysed to answer questions about whether the mood of populations is changing (The war in Ukraine has made Russian social-media users glum) or if people are getting ruder (Foul-mouthed mothers are causing problems for Mumsnet).

Song lyrics can be analysed to identify trends in negativity and complexity while, over a longer timescale, book context can be analysed to conclude, as Jon Burn-Murdoch one Twitter thread, that “western society is shifting away from a culture of progress, and towards one of caution, worry and risk-aversion”.

Changes in political speech are an important indicator not only of the priorities of those in power, but also their relationships with each other and their role in setting the tone of national conversation.

One Guardian analysis, for example, leads on How rightwing rhetoric has risen sharply in the UK parliament, while the award-winning USA Today feature Hope’ is out, ‘fight’ is in: Does tweeting divide Congress, or simply echo its division?” reports that “Language has become more divided and emotional” based on an analysis of more than 2.8 million tweets posted by members of Congress (the original article is no longer online but parts can be seen here).

More simply, the Press Association used the official record of Parliament to calculate that mentions of “Brexit” had gone From nine mentions a year to 9,000.

A more investigative use of change is provided by the Washington Post’s investigation into claims that “USAID’s IG removed critical details from public reports”. By obtaining draft versions of audits reporters were able to identify what changes were made before publication: “more than 400 negative references were removed from the audits between the draft and final versions”

Ranking the most used words and phrases

“Doncaster is the least sexist place in Britain when it comes to street names,” the Doncaster Free Press reported in 2020. The newspaper hadn’t done the analysis: a PR firm had scraped the names of over 200,000 streets and classified them along gender lines. They had also ranked the places with the biggest gap between male and female proportions, and the most common names (Victoria and John).

Ranking can be a quick way to get a story out of a corpus of text — if you can identify a pattern to extract data from. ABC News in Australia, for example, analysed over 1,000 job descriptions to reveal “some of the most common words and phrases advertisers used to attract interns”, while The size of things: an ngram experiment “used Google Books’ Ngram dataset to find the most popular size analogies in English books”.

It turns out that “size of a pea” was the most common.

Even exploratory features can focus on exploring different rankings: the Pudding feature Sing My Name moves through sections ranking the most popular names in songs, which songs have the most repeat mentions of the same name, and what songs contain the most names, among others.

The scale of an issue — revealed in language

Text analysis can allow journalists to reveal the scale of a problem which is not visible in more traditional statistics.

Most of the text-based stories I’ve worked on at the BBC fall into this category: there are no official statistics on the outcomes of police misconduct cases, but I was able to analyse police watchdog reports for a story that revealed “Half of police employees who committed gross misconduct were not dismissed”.

Along similar lines, classifying text in festival line-ups led to the story Music festivals: Only 13% of UK headliners in 2022 are female and declarations of interest were analysed to establish that “One in five MPs continue to employ a member of their family”.

Social media analysis might focus on the scale of a problem on particular platforms: 3 million tweets were analysed for the story Scale of abuse of politicians on Twitter revealed and similar techniques were used by Amnesty and by the Turing Institute to establish the scale of abuse of particular groups.

Scale is a useful fallback angle if you are quantifying text, because no one else will have analysed the text before. The Guardian’s investigation into extremist Facebook groups, for example, leads on the finding that the network “exposes hundreds of thousands of Britons to racist language, conspiracy and disinformation” (my emphasis), while USA Today analysed campaign rally speech transcripts for Trump used words like invasion, killer to discuss immigrants 500 times.

It can also be a useful approach for factchecking or putting a news event into context, as in Der Spiegel’s story How much violence is in Rammstein’s lyrics? following allegations of sexual assault against the band’s singer (the investigation was later dropped)

Women versus men and other variation stories

Variation stories rely on an expectation of fairness, equality or parity. This limits the opportunities for this angle, but it can work especially well where language reveals implicit biases in society that are not quantified anywhere else.

The LA Times‘s scrollytell There are more women than ever in Star Wars. Men still do most of the talking is just one of a number of pieces of data journalism looking at gender variation using text. Others include The Gender Divide in Star Wars Scripts, Meghan gets twice as many negative headlines as positive, analysis finds, The New York Times’s The Words Men and Women Use When They Write About Love, and The Pudding’s The physical traits that define men & women in literature:

“Do authors really mention particular body parts more for men than for women? Are women’s bodies described using different adjectives than those attributed to men?”

Variation along lines of race is revealed in Football commentary racially biased, study finds and HuffPost UK’s The Met Police Are More Likely To Publish Your Mugshot If You’re Black, which classified press releases on criminal sentencing based on mentions of ethnicity, and compared the proportions to full data on sentencing by ethnicity.

Examples of political and ideological variation can be found in The Economist’s Twitter’s algorithm does not seem to silence conservatives and Paris Match’s political speech comparison Macron-Le Pen: Their silences speak volumes.

The Washington Post’s Almost all news coverage of the Barcelona attack mentioned terrorism. Very little coverage of Charlottesville did: “Even before we did our study,” the story notes, “research showed disproportionately high media coverage of terrorism committed by Muslims — even though right-wing extremist groups have committed more attacks”.

Given that databases of media coverage (such as Nexis) provide an accessible source of text data, this feels like an area ripe for more analysis.

‘Meta’ data stories are all methodologies

It is increasingly rare to find ‘meta’ data journalism angles: stories about a lack of data, poor data, or ‘Get the data‘ articles that share data for others to analyse. And the exclusive nature of text data makes it even more unlikely.

But the complexity of quantifying text for analysis means that there is sometimes a need to tell the story-behind-the-story about the methodologies that were employed: How Sky News investigated X’s algorithm for political bias, along with The Guardian’s How we measured the rise of populist rhetoric, and Reading the post-riot posts: how we traced far-right radicalisation across 51,000 Facebook messages are typical examples, while Behind the Guardian’s analysis of 100 years of MPs’ language on immigration shows another approach.

Connections, similarities, and correlations revealed by text analysis

Relationship angles were the least common in the examples of text-based data journalism I found. The few stories that did adopt this focus, however, pointed to at least three possible types of relationships that could be used for stories: correlations, similarities and connections.

A correlation angle can be seen in The Economist’s Why beer snobs guzzle lagers they claim to dislike, which visualises the relationship between word frequencies and user ratings.

The example points to the types of sources that might provide material for similar relationship stories: where text appears alongside a numerical measure. 

Reviews are just one such category of text data. Social media updates (which appear alongside numbers of likes, shares or views, and timestamps) and cultural texts such as books, songs and scripts (figures on sales, streams and views) have the same qualities. 

An alternative approach is to quantify text yourself, and combine it with other data: The Guardian used this to establish a relationship between Donald Trump’s use of a teleprompter and populism, while Paris Match used it to establish a lack of relationship between search interest and candidates’ mentions of an issue.

Similarities formed the focus of the network graph-led A Map of Lexical Distances Between Europe’s Languages and Lexical Distance Among Languages of Europe 2015. The methodology points to how the ‘distance’ between words can be quantified in order to identify patterns.

Connections are the focus for Mathematicians mapped out every “Game of Thrones” relationship to find the main character, which quantifies text by classifying two characters being mentioned in the same sentence as a connection. Although this is a ranking story (the focus is on identifying the ‘main’ character), the analysis could equally have been used to tell a story about relationships, and the same method could be adapted for any text where entities (people, companies, locations) are mentioned. Paris Match, for example, used co-occurrence analysis to identify what words candidates associated with ‘system’.

Connections encoded in large language models provide another potential source: Economist data journalist, Sondre Solstad, uses this to look at the connotations of baby names.

The Epstein files provide further ideas for working with text: The Economist’s Inside Epstein’s Network, for example, also chooses a ranking angle revealing the “500 people who appear most frequently”, and while The Guardian headlined their story as “Jeffrey Epstein’s elite relationships visualised”, what is specifically visualised is the relationships between correspondence and time. What is missing from coverage generally are the clusters, cliques and bridges that can be generated by the Epstein Document Network Explorer.

Leads from analysis of text

The Epstein files also provide an example of how data-driven approaches to text can provide useful story leads. The New York Times, for example, used AI to help “identify clusters, patterns and themes,” although it may be that this is being underused. “The brunt of the work,” they admit, “is being done by a team of editors and reporters across beats and bureaus who have been preparing to dig into the files long before they were published.”

Once you have collected a corpus of text, exploratory analysis can surface potential stories you may not have considered.

When I worked on the BBC story DBS background checks mean NHS staff ‘paying to work’, for example, the angle came as a result of identifying phrases which appeared multiple times in NHS job ads (we then went on to use the data to establish the scale of this). And once I’d collected documents on police misconduct, topic modelling allowed me to see common themes across a large number of documents, any one of which could have been an avenue for further reporting.

I’d love to know of any examples where text analysis has been used to identify story leads. If you’ve been involved in a text analysis project, please let me know in the comments or on LinkedIn.

PEER is a mnemonic (based on a previous post) for remembering the following four types of source:

• Power
• Expertise
• ‍ Experience
• Representative

Each type of source brings something different to the story: voices of power primarily (but not solely) answer questions about action: what was or is being done, what should or would be done about a particular issue. These are easily the most commonly quoted sources in news reporting.

People with expertise can answer the “why” and “how” questions — and are often more likely to speak to journalists — while those with experience can verify or validate (put a human face to) events. Representatives can speak to the wider impact or significance of an issue, or represent community sentiment about it.

Making each type of source explicit allows us to think about what those roles really mean — and identify less obvious ideas for sources with power, expertise, experience or representative qualities.

Power isn’t just political

The most obvious examples of sources with power are politicians. Their power comes primarily from the ability to make and shape laws and other rules, to spend money, and cut budgets. But by thinking about power in general, you can break power down into at least five sub-categories of source:

• Political power more broadly: not just those in government and opposition, but local government, police and crime commissioners, mayors and other elected and unelected officials
• Legal power: police, regulators, sports governing bodies and professional membership bodies all have the power to enforce rules
• Financial power: company bosses, investors, etc.
• Cultural power: celebrities, high profile experts, pundits, authors, former sportspeople, editors and media bosses, and anyone with a large following
• Mobilisation power: those able to mobilise large numbers of people include unions, the armed forces, campaign groups and charities, the media and influential individuals

With that list, you can start to think of a much wider range of potential people to approach on a story than the usual suspects.

Expertise isn’t just academic

The same applies to experts: while universities might be the most obvious place to look for expertise, it’s just one option:

• Academic expertise: those who research and teach in the field
• Non-academic research expertise: those who research topics outside of academia, such as statisticians, charity researchers and analysts, and authors.
• Professional expertise: those who hold authority on the basis of working in the field at a senior level, often for a long time
• The expertise of achievement: those who have won awards and/or broken records in the field (for example, an award-winning barista or former sportsperson)
• Experiential expertise: those with extensive experience of a process or issue (for example someone who has patented dozens of inventions)

The Guardian feature The best £20 you’ll ever spend: 16 affordable essentials the experts swear by provides a good example of the diversity of expertise journalists draw on. Writing in the newspaper’s Inside Saturday newsletter, Daisy Schofield explained how she identified sources for the piece:

“When it came to fashion, I immediately thought of Chioma Nnadi, British Vogue’s head of editorial content, who has impeccable taste. My pet-related expert had to be the dog behaviourist Louise Glazebrook. Her book, The Book Your Dog Wishes You Would Read, was a lifeline when I got my puppy.

“Kelly Holmes was another obvious choice. The double Olympic champion has recently turned her attention to home fitness – including for beginners – so I knew she would be able to advise on work-out equipment for the exercise-averse.”

Other sources included founders or CEOs of companies, more authors, and a prominent menswear designer.

Representatives are not always elected

Representatives take two broad forms: those who act as a voice for a wider group of people, and those who have an ‘ear to the ground’ and can summarise what is being said. Examples include:

• Unions
• Professional associations (not to be confused with regulators)
• Industry bodies and networks (e.g. Chambers of Commerce)
• Charities, campaign groups and pressure groups
• Owners, moderators and founders of groups and forums on social media and offline, and some specialist media
• Religious leaders, community leaders (the two are not the same thing)
• Elected representatives

There are subtle and importance differences between different types of representatives: a union, for example, can negotiate formally on behalf of its members, and union leaders are elected to their positions.

A Facebook group owner is in a very different position: they can give an indication of what members’ main concerns appear to be — but do not speak on their behalf. In some cases they may play more of the role of ‘expert’ on those concerns rather than that of formal representative, but this remains a useful alternative when a group has no formal representation at all.

Experience is hard to find

The hardest type of source for journalists to find is the person with some experience of the issue they are reporting on. These sources often answer one of two key questions in a story: “What happened (or what happens)?” and “What is it like to be affected by this issue?”

They include:

• Eye witnesses to a news event (e.g. an accident, natural disaster, attack, etc.)
• People affected by an issue (e.g. a medical condition, cuts, addiction, bereavement, etc.)
• People who have experienced something similar in the past (e.g. competing in a competition, a crime, etc.)

This requires some empathy both to imagine the types of people who might be affected, and how they might be reached.

A common strategy, for example, is to imagine organisations a person might come into contact with and approach those instead, asking if they can put you in touch with a case study.

Journalists will often approach charities that represent people in certain situations, or with particular medical conditions, for example; or they might speak to campaign groups and lawyers who specialise in particular types of cases or laws.

A more direct strategy is to find sources by looking on social media, specialist forums and groups, or crowdfunding pages. In this approach you are trying to imagine spaces where those people might share their experiences, gather to coordinate action, access support, raise money, or keep in touch with former colleagues.

More tips on finding people and making approaches can be found in How to Tell the Stories of People Behind Data Points.

Sources can occupy more than one category — but their role shouldn’t

You’ll notice that some individuals can appear in more than one category:

• An author can be the voice of expertise, or exercise cultural power (campaigning for change for example).
• A former footballer can describe what it is like to walk out at a cup final, or they can provide expert insights on the tactics of the finalists.
• A politician might be quoted because they have responsibility for a sector, or in their capacity as a representative of their constituency. They might be asked to give expert insights into a political process, or share their experiences of being abused on social media.

It is important, then, to clarify what role a source is playing in your story, and shape your questions accordingly. A politician can be the voice of power, or the voice of their constituents, for example, but it would probably feel confusing or lazy — or both — for them to asked questions that mean they play both roles within the same story.

You should obviously also consider the reliability of sources, and how they relate to each other: make sure to identify and seek out independent voices and sources that disagree with each other to improve the accuracy and objectivity of the story overall.

Audit your sources: PEER’D?

Once classified in this way, you can also use the PEER framework to audit the sources you tend to seek out when reporting, or quote.

Do you tend to mostly think of, or quote, people in positions of power, for example? Are representatives and case studies more of a blind spot?

More specifically, you might use an audit to identify what voices of power (political, cultural, financial, etc.) are loudest in your reporting, or what categories of expertise you have a habit of seeking out.

Are there types of sources you could seek out more which would improve your reporting or give you a wider range of story options? Sources you need to quote less often?

It is worth measuring the diversity of those sources in your audit, too (if it works for you, make the acronym PEER’D) — the 50:50 project, for example, already provides a way to “help production teams improve the range of their on-screen contributors in terms of gender, disability and ethnicity”, and could be integrated into this to identify if there is a particular lack of diversity when generating ideas for certain types of sources.

However you approach it, having a clear understanding of the types of sources you tend to draw on most can help you widen your range of sources, generate more story ideas, have a more diverse range of voices in your reporting, and make your journalism more original.

You can read more about the different roles that those sources play in Here’s a framework to help fill the ‘human gap’ in your story.

Being curious: how often is this happening? How much has it increased?

Curiosity is the first quality I identified in my series on the 7 habits of successful journalists — and FOI is a great way to hone that.

One good way to get started with FOI is to identify an event or problem that you’ve read about, and get curious about it: how many times is that event happening? How much is that problem costing? These are perfect questions for FOI.

Asking for those figures over a period of time also allows you to establish whether those events or costs are becoming more frequent or increasing.

This sort of request can work particularly well when prompted by the experiences of you and your contacts. If you discharged yourself from hospital because there were no staff available to do so, for example, how many other patients are having to do the same?

You might also just be curious about how many times particular things happen — police being disciplined, or lives being saved.

One thing to consider when asking for data on incidents or costs is how things are classified and counted. FOI requests about crimes involving clowns, for example, provided a string of quirky stories back in 2014, but as Richard Osley wrote at the time, the figures included crimes where someone was called a “clown”, had a clown tattoo or wore “clown-like shoes”. Locations and objects using the keyword were also included.

Official definitions of certain terms can be useful to know (or ask for). If you’re asking about frequent callers to the ambulance service, for example, it helps to know that they define that as “A patient aged 18 years or over, calling from a private dwelling [that] generates 5 or more incidents in a calendar month” — so it wouldn’t include younger callers or those calling from phone boxes.

Being sceptical: checking a claim

How can you develop scepticism as a journalist? Find an official statement and ask what evidence there is behind it.

Protests have proven particularly fruitful subject matter for this: official bodies often leave vital information out of public statements that can be unearthed through FOI.

For example police force injury figures from protests have often included minor injuries unrelated to the protestors themselves.

As well as asking for breakdowns of figures, you can also ask for internal correspondence and reports to compare those to the public story being told.

A good example of this is the East Bay Express‘s use of FOI to obtain internal documents recording discussions about a surveillance hub called the Domain Awareness Center (DAC). Publicly, the surveillance was pitched as aimed at tackling gang and street violence, but FOI revealed that privately this wasn’t the real objective:

“While the emails reveal a great deal about the DAC, they are also notable for what they do not talk about. Among the hundreds of messages sent and received by Oakland staffers and the city’s contractor team responsible for building the DAC, there is no mention of robberies, shootings, or the 138 homicides that took place during the period of time covered by the records. City staffers do not discuss any studies pertaining to the use of surveillance cameras in combating crime, nor do they discuss how the Domain Awareness System could help OPD with its longstanding problems with solving violent crimes. In more than 3,000 pages of emails, the terms “murder,” “homicide,” “assault,” “robbery,” and “theft” are never mentioned …

“City staffers repeatedly referred to political protests as a major reason for building the system.”

Employing empathy: investigating private companies

Empathy is an act of imagination, not sympathy. And when it comes to non-public bodies, you’ll need to use this to imagine at what points they come into contact with public systems.

You cannot send an FOI request to a company or a charity because they are not subject to the FOI Act. But you can use FOI to get information about those organisations through the public bodies that they have to deal with.

Accusations relating to poor or unsafe working conditions, for example, can be investigated by asking local emergency services about callouts to addresses owned by a particular company. Prisons might be run by private companies but you can FOI the health board for data on incidents, as The Ferret did.

Many companies are overseen by regulators who are subject to FOI. Charities are regulated by the Charity Commission — but also, if they are receiving money from public bodies to provide services, contracts can be obtained through FOI along with any information they are required to provide. Requesting correspondence with public bodies is another strategy to consider.

In sport, games have to be policed and stadiums inspected by authorities. Police forces collect data on how many fans are at matches — making it possible to report stories on how many fans are staying away. Tax authorities investigate footballers’ finances; public bodies carry out doping tests and provide grants.

Put yourself in the company’s place and ask the following two questions:

1. What rules does this company have to follow (and which public bodies enforce those)?
2. Where does it get, or give, money from or to public bodies (including tax)?

Being creative: adapting to what you get (or don’t get)

In some cases you will not receive what you asked for in your FOI request because the organisation doesn’t collect that information.

This doesn’t mean that you don’t have a story: the fact that the organisation doesn’t collect that information might be a story in itself.

It is said that “what gets measured gets managed” — which also means that anything that’s not being measured is not being managed. If something is a problem and it isn’t being managed, then that’s newsworthy.

One set of FOI requests to health authorities, for example, “revealed how few have a grasp of their legal obligations” because they “could not provide details on the number or the location of section 140 beds in their region [and] could not say what the arrangements were for people who need a section 140 bed to actually receive one.”

Before you send an FOI request check which organisation is responsible for dealing with the events you are interested in. The Department for Education, for example, will not provide information on missing children because that is the responsibility of the National Crime Agency.

If you are sending requests to multiple bodies, such as all the police forces or councils, there’s also a good chance that some will fail to provide information. In this case the phrase “at least” is going to come in useful:

• Liberty Investigates‘ FOI story, for example, revealed that “At least 28 universities are now known to have launched disciplinary investigations against students and staff in connection with their Palestine activism since October 2023”. Lower in the story it adds that “Nearly 50 universities refused to respond”.
• The Guardian reported that “At least 511 people died in England [after the ambulance they called for took up to 15 hours to reach them]” (“Only three of England’s 10 regional ambulance services provided full-year figures for the last two years as requested.”)
• BBC Radio 5 Live‘s story Police pay out at least £22m to informants in five years was based on responses from 43 of the 45 territorial police forces, one of which only provided figures for three of those years.

More generally, be prepared to adapt your story to what the information says. If you were hoping to do a story on change, but the figures haven’t changed, adapt by changing the angle to focus on the current scale of the problem, or which areas have the biggest problems — or use interviews to find a story about concerns over a lack of improvement. Or if the figures are improving, use interviews to find out what’s working.

If the request is refused, look at stories about refusals to get ideas on how to tell a story about refusals or lack of transparency.

Being passionate: pick a topic you care about

A focus on “Freedom of Information law” and “public bodies” can be dry. But focus on music, sport, discrimination or education and you might start to see FOI differently.

The key is to focus on what makes you passionate.

The Guardian’s roundup of How students use FOI to investigate their universities is a great place to start for student journalists who are curious, angry or passionate about their own institution.

If music is your passion look at stories like DJ Mag’s story on police requests to remove UK drill content from TikTok, or The Register’s piece on the use of facial recognition at Download Festival.

In sport look at The Pitch Inspection’s stories on the lack of transparency around meetings with football bosses and Saudi officials, or the Daily Mail’s “grim list of Premier League match-day injuries you NEVER hear about!” Look at public bodies that operate in the sector, like Sport England, UK Anti Doping, or even the organising committee for a Commonwealth or Olympic Games.

Discrimination is a regular focus of FOI requests, from racist incidents in classrooms to how many officers found guilty of racism are sacked, from stalling on equal pay claims to the gap between top-paid male and female staff, from the “shocking disparity” faced by people with non-visible disabilities to the number of flights missed by disabled passengers. If you’re passionate about discrimination, FOI can be a powerful tool to shine a spotlight on it.

Being disciplined: make a plan to follow up on policies

One particularly clever use of FOI is to check what happens after a new policy has been announced, a new process introduced, or a new taskforce or group formed.

The impact won’t be immediate — so you’ll need to be organised in planning ahead for when evidence of its impact will be available to FOI.

Set a calendar reminder for a certain time after its introduction: it could be as short as six months (for a taskforce expected to meet monthly, for example) or as long as five years. You might aim to get the information in time for a key anniversary, for example (either of the announcement or an event that led to it) by asking for the information a month or two before it.

Once you’ve sent an FOI request, it’s also a good idea to set a reminder four weeks later to chase it up (the FOI tool WhatDoTheyKnow sends you one automatically).

And for multi-FOI projects many journalists will create a spreadsheet to keep track of all their requests and the status of each. FOI website Proactively Open have shared a template spreadsheet here.

Being disciplined and creative: use existing stories as recipes

Each of the stories linked above provides a recipe for your own FOI story, which can be adapted for your own purposes. Here are just five ways you can be more creative in coming up with ideas for FOI (you can develop discipline by generating ideas for each method systematically):

• Bring it up to date: can you ask for the same information, just for a more recent timescale?
• Change the location: can you ask for the same information for a different area? Locally? Nationally? Internationally? (Dozens of countries have FOI laws)
• Change the category: if the previous story was about car thefts, what about asking for the same information about bicycle thefts, or violent crime, or shoplifting?
• Change the body: if it was about police, why not ask for similar information from the fire or ambulance service?
• Change the angle: if the focus was scale, ask for information that allows you to focus on change, or ranking, and vice versa. Use the 7 angles listed here.

Being persistent: negotiate, appeal, demand

FOI can provide great opportunities to develop tenacity and persistence. If an organisation rejects your FOI, don’t treat that as the end of the process. Instead, see it as an opportunity practise key skills:

Can you negotiate? Consider why they have refused it, and how you or they might adjust so that those reasons no longer apply.

If the authority has refused your request because it would take too much time, for example, you can narrow the request to a shorter timeframe, or fewer questions. They are obliged under the FOI Act to help you do this, too, so you can ask them to help identify what comes under the time limit.

Similarly, if the refusal is because of privacy or confidentiality, you can ask for aggregate figures instead of individual incidents, and you can ask for material to be redacted (this should not be counted for the purposes of cost limits)

Can you demand? Some reasons for refusals (exemptions) are subject to a public interest test. You can ask them to provide evidence that they have conducted this – and one useful technique is to include this demand as part of your original request should they choose to refuse. The idea is that refusing your request should be harder than complying with it.

Can you appeal? All refusals can be appealed. At first you have to appeal internally. A surprising number of internal appeals are successful, but if it’s not then you would next appeal to the ICO, and even the courts. In most cases you’ll be arguing that either the public interest of releasing the information outweighs any private interests, or that an exemption has been incorrectly applied (e.g. it’s not a national security issue).

You’ll need to do some research into the exemptions that they’ve used – there’s plenty of material out there on these, and cases where they’ve been successfully (and not) appealed.

You can’t learn that? More ways to develop good habits

For other tips on developing curiosity, scepticism, persistence, discipline, creativity and passion, check out the posts in the series below:

A pirâmide invertida do jornalismo de dados mapeia o processo de utilização de dados na reportagem, desde a geração de ideias, passando pela limpeza, contextualização e combinação, até à comunicação. A fase final — a comunicação — apresenta uma série de opções: desde a visualização e sonificação até à personalização e ferramentas. Mas quais são as melhores práticas para cada uma?

(Também disponível em inglês, alemão e espanhol, russo e ucraniano).

1. Visualização

A visualização é normalmente a forma mais rápida de comunicar os resultados do jornalismo de dados: ferramentas gratuitas como Datawrapper e Flourish muitas vezes exigem apenas que você carregue os seus dados e escolha entre várias opções de visualização.

Quando utilizada corretamente, a visualização pode aumentar significativamente o engajamento — mas a velocidade não garante qualidade. O aumento do “chart junk” — elementos gráficos supérfluos que distraem do conteúdo — demonstra que a visualização é tão suscetível a efeitos superficiais quanto outras formas de jornalismo.

Existe uma longa tradição de princípios úteis da visualização impressa que continuam relevantes para gráficos, mapas e infográficos online:

• Concentre-se em poucos pontos de dados relevantes
• Evite efeitos 3D
• Garanta que o gráfico seja compreensível mesmo sem texto adicional

Muitas redações também desenvolveram as suas próprias diretrizes — desde o Visual Vocabulary do Financial Times e o Visual and Data Journalism Cookbook da BBC até ao Audiences Tableau Style Guide. Autores especializados como Dona Wong (The Wall Street Journal Guide to Information Graphics) e Alan Smith (How Charts Work) também oferecem dicas valiosas.

A visualização pode ser extremamente eficaz — mas é precisamente por isso que é importante agir de forma estratégica e, por exemplo, garantir que cada gráfico inclua um link para a sua fonte.

2. Narrar

Qualquer história de dados também pode ser contada com palavras. Para uma notícia simples, concentre-se no que os dados “revelam”: a dimensão de um problema; quem ou onde está melhor ou pior; desigualdade; relações; ou mudança.

Decida antecipadamente quais números são mais importantes para a sua história e use o menor número possível: o princípio do “menos é mais” aplica-se à escrita sobre dados tanto quanto à visualização.

Tente intercalar os parágrafos baseados em dados com secções que contenham citações ou informações de contexto. Isso evita a monotonia e pode fornecer contexto fundamental.

Boas histórias não nos dão apenas os factos do que está a acontecer, mas também por que acontece (através de especialistas) e por que é relevante (impacto humano). As seguintes perguntas ajudarão a fornecer contexto:

• O número representa uma grande proporção do total?
• É superior ao do ano passado ou comparado com outras localidades?
• Para que seria usado esse valor? Torne os números mais tangíveis convertendo-os numa base per capita ou diária.

Lembre-se de que números adicionais podem sempre ser disponibilizados aos leitores através de uma tabela, gráfico ou link para os dados completos.

Em alguns casos, os números podem ser omitidos completamente: os dados podem levá-lo a um estudo de caso que se torna a história em si…

3. Humanizar

Estudos de caso colocam um rosto humano nos números. Para o jornalismo audiovisual em particular, isso é essencial: histórias baseadas em dados em plataformas de vídeo e áudio normalmente começam com um rosto, ou uma voz no centro da história, não apenas para fornecer material audiovisual fundamental. Mas isso também pode funcionar bem online e nas redes sociais.

Quando a BBC Shared Data Unit trabalhou numa história sobre as falhas no tratamento de drogas para jovens, por exemplo, optou por iniciá-la com: “Três mulheres cujos filhos adolescentes morreram após se tornarem dependentes de drogas disseram à BBC que o sistema falhou com os seus filhos.”

Os dados podem ajudar a conduzir-nos a estudos de caso: ordenar uma folha de cálculo para encontrar a organização ou o local mais afetado por um problema pode mostrar-nos para onde direcionar as nossas investigações ou onde filmar.

4. Tornar social

A comunicação é um ato social, mas não são apenas as palavras e os elementos visuais que são sociais: os próprios dados também podem ser sociais. O Guardian demonstrou isso com particular sucesso ao cultivar uma comunidade ativa em torno do seu Data Blog (que registou maior engajamento do que o artigo médio do Guardian) e em torno da sua API. Também aproveitou as propriedades sociais da visualização com um grupo no Flickr.

Iniciativas de crowdsourcing destinadas a recolher dados também podem conferir uma dimensão social aos dados. O projeto de despesas dos deputados do Guardian e a tentativa de Charles Arthur de usar crowdsourcing para prever as especificações do iPad são exemplos disso. Mas há outros exemplos também — especialmente quando é difícil obter os dados de outra forma.

A narrativa social exige pensar em como capacitar e inspirar o público a contar as suas próprias histórias. Isso pode ser feito fornecendo uma plataforma, como o grupo Flickr do Guardian; pode ser fornecendo dados brutos; ou pode ser fornecendo uma forma de os leitores criarem e partilharem histórias personalizadas descobertas através da interatividade (veja abaixo). O 7 billion and you da BBC, por exemplo, permitiu aos utilizadores partilhar o seu próprio número nas redes sociais. Isso gerou milhares de histórias simples individuais, além da história interativa personalizada no site da BBC.

5. Sonificação/materialização

Em formatos de rádio ou podcast onde gráficos não são possíveis, os dados podem ser representados como som (sonificação) — o que muitas vezes torna o conteúdo mais fácil de compreender. Por exemplo:

• Planet Money: “U.S. Home Prices, Sung As Opera“
• The World: Música para ilustrar apagões numa reportagem
• Reveal: Estatísticas de sismos traduzidas em som

O podcast “Loud Numbers” usa esta técnica em todos os episódios, para histórias que vão desde a época de incêndios florestais no Canadá até à degustação de cerveja e à economia dos EUA. Também acolhe uma comunidade para “sonificadores”. Pode encontrar ferramentas para converter dados em música no Data Sonification Toolkit.

Alternativamente, os dados podem ser transformados — materializados — em algo tangível que possa ser experienciado fisicamente. As duas formas mais comuns são como objetos num espaço expositivo — por exemplo, uma pilha de grãos de arroz simbolizando um determinado número de pessoas — e objetos decorativos ou de vestuário, como uma pulseira ou trabalho de crochet.

6. Personalizar/utilizar

Os dados abrem todo o tipo de possibilidades em torno da interatividade em geral — e da personalização em particular.

Pelo menos quatro tipos de personalização podem ser identificados na narrativa de dados:

• Personalização demográfica (por exemplo, personalização geográfica ou financeira)
• Correspondência de crenças (por exemplo, políticas partidárias)
• Personalização de interesses (por exemplo, páginas iniciais personalizadas)
• Personalização de conhecimento (por exemplo, nível de detalhe, extensões de navegador e jogos)

A personalização demográfica permite que um utilizador veja como “pessoas como eu” são afetadas por um problema ou política. A personalização geográfica é o exemplo mais comum: pode ser fornecida através de mapas interativos — mas também pode ser feita através de menus dropdown ou caixas de pesquisa onde o utilizador é convidado a inserir o seu código postal ou outras informações geográficas para descobrir como uma determinada questão está a desenrolar-se na sua cidade. Dados de geolocalização ou de redes sociais também podem ser utilizados.

As calculadoras personalizam a narrativa em torno de outras variáveis demográficas: salário, estado civil e padrões de consumo, por exemplo, podem ser usados para contar uma história sobre como um novo orçamento afetará o utilizador. O formato de calculadora também foi usado para personalizar histórias sobre inflação, obesidade, os ganhos de estrelas desportivas e celebridades, e término de relacionamentos.

A correspondência de crenças é mais frequentemente utilizada durante eleições, quando muitas equipas de dados produzem questionários interativos para ajudar a identificar o partido cujas políticas correspondem às suas.

A personalização de interesses tende a ocorrer ao nível do site, em vez do nível da história, onde uma página inicial pode ser personalizada para se concentrar nos interesses do utilizador (como a sua equipa desportiva favorita).

Mas também é possível ao nível da história, fornecendo diferentes níveis de detalhe com base no grau de interesse do utilizador no assunto. No passado, o “jornalismo estruturado” tornou isso possível. A IA agora torna esta abordagem muito mais fácil do ponto de vista técnico.

A personalização de conhecimento é muito semelhante: envolve personalizar a informação com base no conhecimento existente de uma pessoa. Um questionário é um exemplo simples: no final, as lacunas no conhecimento do utilizador são identificadas e preenchidas. Ou os sites podem pedir ao utilizador que descreva o seu nível de conhecimento — ou usar o seu histórico de leitura — antes de apresentar uma versão da história adequada a esse nível.

Os jogos muitas vezes fornecem implicitamente este tipo de personalização, pois o conhecimento e as crenças do utilizador moldam as decisões que toma, que por sua vez moldam a história. Decisões sobre um orçamento, ou no papel de um motorista da Uber, ajudam a “preencher as lacunas” sobre como poderiam ser as experiências resultantes dessas escolhas.

As extensões de navegador e a realidade aumentada (RA) são exemplos mais raros: ambas adicionam uma camada de conhecimento ao mundo do utilizador. Quem navega na Amazon com a extensão do The Markup instalada, por exemplo, verá os produtos de marca própria da plataforma “revelados”. E quem usa uma ferramenta de RA verá uma camada virtual adicionada aos seus arredores físicos.

Algumas destas formas têm sido subdesenvolvidas devido aos limites de tempo de desenvolvimento e de informação sobre os utilizadores, mas a IA generativa reduz as barreiras em ambos os casos: a vibe coding torna mais fácil do que nunca desenvolver páginas web interativas, dashboards e ferramentas; e os grandes modelos de linguagem podem agora “traduzir” facilmente informação com base no conhecimento, interesses ou demografia de um utilizador. O principal desafio ético da personalização é como equilibramos duas tensões concorrentes: envolver o público com o que é importante e garantir que as pessoas não fiquem isoladas das experiências de quem é diferente delas.

Se tiver exemplos interessantes de narrativa de dados para acrescentar aos listados acima, informe-me nos comentários ou no LinkedIn.

Traduzido com Claude Sonnet 4.5. Por favor, diga-me se encontrar algum erro nos comentários ou através do LinkedIn.

Os projetos de jornalismo de dados envolvem várias etapas, cada uma apresentando seus próprios desafios. Para ajudar a compreendê-las, criei o que chamei de ‘Pirâmide Invertida do Jornalismo de Dados’. Ela delineia as etapas que precisam ser consideradas à medida que a matéria avança desde a conceção inicial até a comunicação dos resultados, e como elas se relacionam entre si. Abaixo, explico cada etapa, identifico questões a considerar conforme o projeto avança e ofereço conselhos e dicas sobre como enfrentá-las.

(Também disponível em Inglês, Alemão, Espanhol, Finlandês, Russo e Ucraniano.)

Etapa 1: Conceber

O primeiro desafio que um jornalista enfrenta é conceber uma ideia viável para uma matéria baseada em dados.

Existem pelo menos sete maneiras pelas quais os jornalistas normalmente chegam a uma ideia:

1. A partir de uma divulgação pública de dados;
2. A partir do acesso a um conjunto de dados exclusivo;
3. Replicando ou adaptando uma matéria anterior baseada em dados;
4. Dando seguimento a uma notícia;
5. A partir de uma pergunta;
6. A partir de uma denúncia; ou
7. A partir da exploração.

Abordagens que começam fora dos dados (perguntas, denúncias e acompanhamento de notícias) são mais difíceis. Nessas situações, é provável que você precise adaptar a ideia em resposta aos dados que realmente existem (se é que existem), e ao que esses dados de facto medem.

Em Como fazer brainstorming de ideias de matérias sobre dados da COVID-19, descrevo algumas técnicas usadas para enfrentar esses desafios. Técnicas mais investigativas são exploradas em Como usar as ‘4 etapas da curiosidade’ como estrutura para investigações e Empatia como ferramenta investigativa: como mapear sistemas para gerar ideias de matérias.

Abordagens que começam com os dados envolvem um desafio diferente: identificar e escolher possíveis ângulos de matéria a partir dos dados disponíveis. Nessas abordagens, considere o seguinte:

• Que colunas os dados têm? (O que é medido)
• Que histórias poderiam ser contadas sobre cada uma dessas medidas?
• O que falta nos dados? (O que não está sendo medido)

Em Os ângulos mais usados por jornalistas para contar histórias com dados, descrevo como diferentes ângulos podem ser encontrados em diferentes medidas.

• Se houver uma coluna para datas, meses ou anos, por exemplo, então uma história pode ser contada sobre mudança;
• Se houver uma coluna para geografia ou categoria, uma história provavelmente pode ser contada sobre classificação ou variação.
• E quase todos os dados podem ser quantificados para contar uma história sobre a escala de algo.

Etapa 2: Compilar

O processo de compilação de dados pode ser rápido e simples — ou pode ser um grande projeto de jornalismo de dados por si só.

1. A compilação mais direta é quando os dados são publicados regularmente numa determinada página (ou enviados num comunicado à imprensa), e é daí que vem a ideia.
2. Você pode encontrar dados através de técnicas de pesquisa avançada
3. Alguns dados são publicados através de uma API, e você precisará escrever código para ‘puxar’ esses dados.
4. Você pode usar scraping — onde um script explora múltiplas páginas ou documentos para compilar dados destes numa planilha
5. Você pode converter dados não estruturados, como texto, em algo que possa ser analisado
6. Você pode compilar dados através de observação, inquéritos, formulários online ou crowdsourcing. O tempo e a despesa envolvidos em tal compilação significam que raramente é usado.

Esta etapa de compilação é especialmente importante não apenas porque torna o projeto de jornalismo de dados possível, mas porque você frequentemente precisará retornar a ela como parte das etapas subsequentes: os dados podem precisar ser compilados para limpar seus dados existentes, para contextualizá-los e combiná-los, e para melhor comunicar as descobertas.

Etapa 3: Limpar

Ter dados é apenas o começo. Ter confiança nas respostas que você obtém dos dados significa poder confiar na qualidade dos dados — e às vezes isso significa limpá-los.

A limpeza normalmente assume três formas: remover erros humanos; converter os dados num formato que possibilite fazer perguntas; ou garantir que as respostas obtidas sejam precisas.

Razões comuns para limpeza incluem:

• entradas duplicadas;
• entradas vazias;
• o uso de valores padrão onde nenhuma informação estava disponível;
• formatação incorreta (por exemplo, palavras em vez de números);
• entradas corrompidas ou entradas com código HTML;
• múltiplos nomes para a mesma coisa (por exemplo, BBC e B.B.C. e British Broadcasting Corporation);
• dados ausentes (por exemplo, distrito eleitoral);
• dados mistos na mesma coluna (por exemplo, categoria e subcategoria);
• ou dados na forma errada (por exemplo, trocar colunas e linhas).

Etapa 4: Contextualizar

Como qualquer fonte, os dados fornecem uma perspetiva sobre uma história e nem sempre podem ser confiáveis. Eles vêm com suas próprias histórias, vieses e objetivos.

Portanto, como qualquer fonte, você precisa fazer perguntas sobre eles: quem os recolheu, quando e com que propósito? Como foram recolhidos? (A metodologia). O que exatamente eles querem dizer com isso?

Você também pode precisar entender jargões, como códigos que representam categorias, classificações ou localizações, e terminologia especializada.

Estabelecer o contexto para um conjunto de dados pode levá-lo a compilar mais dados. Por exemplo, saber o número de crimes reportados numa cidade é interessante, mas só se torna significativo quando você coloca isso em contexto.

Esse contexto pode ser o tamanho da população, ou o número de policiais, ou os níveis de criminalidade cinco anos atrás. Pode ser o contexto de experiências ou perceções sobre o crime, taxas de condenação ou níveis de desemprego. Pode ser o contexto das definições de categorias de crimes.

Alguma literacia estatística básica é muito útil, então dedique um pouco de tempo a ler sobre o assunto. Como Mentir com Estatísticas é um clássico curto e fácil de ler na área, que é complementado pelo mais recente Como Fazer o Mundo Somar. E o livro de Ben Goldacre, Bad Science, prepara você para cobrir dados sobre saúde e ciência.

Adicione o podcast da BBC More or Less às suas subscrições para ouvi-los discutir histórias sobre dados, e confira The Tiger That Isn’t, dos apresentadores.

Etapa 5: Combinar

Boas histórias podem ser encontradas num único conjunto de dados, mas frequentemente você precisará combinar dois juntos. Afinal, dada a escolha entre uma matéria de fonte única e uma de múltiplas fontes, qual você preferiria?

Combinar dados é frequentemente parte de colocá-los em contexto: você regularmente vai querer combinar dados sobre o número de eventos em diferentes áreas com dados sobre as populações nessas áreas, por exemplo, para classificar a frequência per capita desses eventos. Dados sobre inflação terão que ser combinados com dados de gastos para ajustar valores de diferentes anos.

Às vezes você precisa combinar dados para obter a resposta a uma pergunta:

• Novos dados precisarão ser combinados com dados mais antigos para contar uma história sobre mudança.
• Dados sobre eventos precisarão ser combinados com dados populacionais para contar uma história sobre a escala por pessoa (ou por 10.000 pessoas, etc.)
• Dados sobre o desempenho de escolas ou hospitais precisarão ser combinados com dados sobre as localizações dessas instituições, para contar uma história classificando as áreas com os melhores ou piores desempenhos médios, ou simplesmente mostrar a extensão da variação.
• Para criar um mapa exploratório mostrando a distribuição de eventos, os jornalistas regularmente combinam um conjunto de dados com dados cartográficos. Exemplos incluem: Die Spur des Geldes: Prioritäten der EU-Migrationspolitik e Die Stadtflucht, enquanto Mein Viertel, eine Blase usa a mesma técnica para mostrar uma relação entre geografia e crenças.

Em cada etapa: Questionar

Questionar envolve analisar seus dados para obter respostas — mas também ocorre em cada etapa:

• Ao conceber ideias, questione seus vieses e pontos cegos. De quem são as vozes que faltam no processo de desenvolvimento da ideia? Você está obtendo ideias apenas de conjuntos de dados públicos — você também poderia considerar dados não publicados ou compilar dados você mesmo? Você está focando no setor público mas não no setor privado? Você tende a gerar tipos particulares de ângulos de matéria mais do que outros (como mudança em vez de escala)?
• Ao compilar dados, questione quão autorizada é a fonte e quão confiável é sua metodologia. Isso afetará como você comunica os resultados e que compilação, limpeza e contextualização adicionais precisam ser feitas.
• Faça perguntas sobre que contexto pode ser necessário: populações são frequentemente necessárias para colocar o número de eventos no contexto de quantos eventos houve por pessoa. O contexto demográfico permite que você faça perguntas sobre relações potenciais com privação, idade e outros fatores. Dados históricos permitem que você coloque dados recentes no contexto de anos anteriores. Dados sobre dinheiro precisarão do contexto da inflação: qual seria o equivalente hoje de determinado valor de cinco anos atrás?
• Faça perguntas sobre quais dados você está combinando: você deve usar a população inteira ou uma faixa etária específica ou outra demografia que se relacione com a matéria? Você deve usar a inflação geral ou focar num tipo particular de bens ou serviços?

Questionar pode evitar o viés de confirmação e desenvolver ceticismo.

Contexto

Como qualquer fonte, os dados fornecem uma perspetiva sobre uma história e nem sempre podem ser confiáveis. Eles vêm com suas próprias histórias, vieses e objetivos.

Portanto, como qualquer fonte, você precisa fazer perguntas sobre eles: quem os recolheu, quando e com que propósito? Como foram recolhidos? (A metodologia). O que exatamente eles querem dizer com isso?

Você também pode precisar entender jargões, como códigos que representam categorias, classificações ou localizações, e terminologia especializada.

Estabelecer o contexto para um conjunto de dados pode levá-lo a compilar mais dados. Por exemplo, saber o número de crimes reportados numa cidade é interessante, mas só se torna significativo quando você coloca isso em contexto.

Esse contexto pode ser o tamanho da população, ou o número de policiais, ou os níveis de criminalidade cinco anos atrás. Pode ser o contexto de experiências ou perceções sobre o crime, taxas de condenação ou níveis de desemprego. Pode ser o contexto das definições de categorias de crimes.

Alguma literacia estatística básica é muito útil, então dedique um pouco de tempo a ler sobre o assunto. Como Mentir com Estatísticas é um clássico curto e fácil de ler na área, que é complementado pelo mais recente Como Fazer o Mundo Somar. E o livro de Ben Goldacre, Bad Science, prepara você para cobrir dados sobre saúde e ciência.

Adicione o podcast da BBC More or Less às suas subscrições para ouvi-los discutir histórias sobre dados, e confira The Tiger That Isn’t, dos apresentadores.

Etapa 6: Comunicar

A visualização é a forma mais óbvia de comunicar uma história que você encontrou nos dados – mas não é a única. Outras incluem:

• Usar apenas narrativas textuais
• Usar estudos de caso de pessoas reais para dar vida aos dados
• Usar interatividade para ajudar o utilizador a ver como pode ser pessoalmente afetado
• Transformar dados em sons ou objetos físicos para fornecer outras formas de interagir com os dados
• Criar ferramentas que capacitem a audiência a usar os dados de forma prática

Como há muito mais a dizer sobre isso (veja o diagrama abaixo), escrevi um post separado sobre o tema. Aguardo com expectativa o vosso feedback sobre a pirâmide!

Traduzido com Claude Sonnet 4.5. Por favor, diga-me se encontrar algum erro nos comentários ou através do LinkedIn.

Too often discussion around using AI is “either/or” — an assumption that you either use AI for a task, or do it yourself. But there’s another option: do both.

“Parallel prompting“* is the term I use for this: while you perform a task manually, you also get the AI to perform the same task algorithmically.

For example, you might brainstorm ideas for a story while asking ChatGPT to do the same. Or you might look for potential leads in a company report — and upload it to NotebookLM to perform the same task. You might draft an FOI request but get Claude to draft one too, or get Copilot to rewrite the intro to a story while you attempt the same thing.

Then you compare the results.

The best of both worlds

There are a number of advantages to parallel prompting:

• You get the ‘best of both worlds’: AI can draw on a vast amount of indiscriminate information; you draw on a narrower but more selective knowledge base specific to your context
• It reinforces and expands knowledge: drafting a prompt encourages you to ‘teach’ the AI what you know, a common way to reinforce knowledge. It might also encourage you to learn more about something in order to design a better prompt.
• It forces externalisation and review: drafting a prompt also encourages you to break down a problem and consider your criteria for a successful result. Comparing AI responses with your own results encourages you to be more critical about both.
• Human bias and AI bias can both be addressed: AI is biased in different ways to humans, who have cognitive biases. Comparing results allows you to address both.
• It improves prompt design skills: the more prompts you design, the better you get — you’re skilling up, not deskilling.

Of course, parallel prompting means that you’re not necessarily saving time by getting AI to do a task for you — at least in the short term.

But it can save time in the long term.

For example, an FOI request drafted this way may lead to fewer problems further down the line. Or interview questions generated using parallel prompting might save time chasing up extra information later. You might do some data analysis at the same time as asking an AI tool to perform the same analysis — highlighting mistakes that you (or AI) tend to make, or better methods you can use in future.

In other words, it’s a learning process through being exposed to new ideas and methods — precisely the opposite of the deskilling that can result from poor use of AI.

*Parallel prompting has also been used to refer to using multiple prompts with different personas at the same time. That’s cool too, but not how I’m using the term.

The checklist — M.A.T.T.E.R. — covers six things to consider: Meaning, ‘Affect’ (emotion), Type of format, Temporality, Ethics and relevance.

“It offers a fast, structured way to assess what a meme does, why it works, and whether it actually matters,” Bösch writes, encouraging reporters “to slow down, separate feeling from meaning, assess impact and ethical risk, and make a conscious editorial decision” before embedding or covering a meme.

The MATTER checklist

Meaning is all about checking if you understand the background to the meme, who is likely to understand it, and “which parts of its meaning are coded, dogwhistled, or only visible to in-groups.”

Affect asks “what emotion does the meme trigger first? Is it earnest, ironic, cynical, or ambiguous?”

Type involves looking at the format of the meme, if AI is involved, and to what extent it’s native to a particular platform.

Temporality requires you to consider how long the meme has existed and whether its meaning has changed in that time, as well as where it is in its lifecycle (for example, is it likely to trigger widespread recognition, or boredom)

Ethical considerations include any use for propaganda purposes and agenda-setting, the involvement of vulnerable people, trauma or harassment, and the potential to cause harm.

Relevance involves relating the meme to classic values around newsworthiness, asking the question why it matters now. It might be because of who is circulating the meme, or the impact that it is having. Relevance also relates to the audience and whether it is circulating among that audience and likely to affect them.

Marcus says he has tested the approach “on a mix of older and current memes” and is inviting others to try it out. You can contact him with feedback at m@marcus-boesch.de

A desire for friction

One person worried about the loss of friction that AI represented: using AI tools increased the risk that they would complete tasks too easily, without the friction that would help them to learn.

Low friction is a recurring challenge in teaching, learning and journalism: as access to information — and each other — has become easier through search engines, mobile phones and email/social media, expectations and behaviours have changed.

This lack of friction means less time for strategic choices about what information to use, and how to source it, or who to approach, and how.

It can lead to sloppiness and impatience, and reduce opportunities for developing problem-solving skills and persistence.

Done right, however, using AI can actually reintroduce friction into a process: this is what happened when I introduced an AI diary into my teaching last year, leading to an increase in reflection and problem-solving.

Using hybrid destination-prompting can also design friction into the process.

Concern over boundaries

Another attendee was concerned about crossing the boundaries that we set for students. “How,” he argued, “can we ask them not to use AI if we use it ourselves?”

One challenge here is that it is also difficult to enforce a boundary if you don’t know when it’s been crossed. This is especially difficult if you’ve not used AI yourself and developed a feel for its output.

Another challenge is that those boundaries differ between news organisations and universities — and are constantly shifting.

Part of the role of journalism study is to critically interrogate those boundaries and the arguments around them in order to know where and when they should be enforced.

For example, through experimenting with AI we might identify that the most important boundaries are determined by risk, or truthfulness (e.g. being up to date) or bias. We might also identify good and bad practice, or where boundaries are not clearly defined.

A desire to work hard

One person enjoyed working hard on projects, and did not want AI to do that work for her.

There are two concepts tied up in this sentiment: job satisfaction, and effort.

AI can increase the risk that less effort is expended on our work. But this is just one choice. Another is to invest the effort elsewhere.

The inventions of other labour-saving devices provide useful analogies here. Once the washing machine was invented, for example, a choice opened up: we could continue to work hard to wash clothes by hand, exercising certain muscles and maintaining that skill — or we could use the time to exercise other muscles or develop other skills.

So using AI doesn’t mean you don’t work hard — you will probably just work hard on something else, such as sourcing and conducting interviews, verification and factchecking, writing and editing — or even learning AI-based skills such as prompt design.

And this is where job satisfaction plays a role.

A reflection by one of my students in the conclusion to their AI diary last year was a particularly good example of this: “After I’d become comfortable with the AI,” she wrote, “it was no longer enjoyable to use because I wasn’t mentally stimulated by my work.“

As a result of this, along other concerns, she made a conscious critical decision to use it less.

Equally, someone may find the stimulation provided by their work fades to a degree where they prefer to use AI to free up time for them to focus on more challenging parts of reporting.

A desire for expertise

“I want to be an expert” said another person. The implication here was that you can only become an expert in something if you don’t take short cuts, i.e. use AI.

This is something I’ve written about in my posts on journey and destination prompts.

Yes, AI can offer a tempting shortcut, and research suggests when it is used in this way we retain less information and have lower cognitive activity (the same applies to a lesser degree to using search engines).

But that isn’t the only way that AI can be used.

Journey prompting — asking for advice on a process, such as learning — can be used to help design a more effective strategy to becoming an expert, for example.

Role play prompts are a particular example of this: creating a mentor can help support the process of building expertise in an area.

And AI’s translation abilities — not just between languages, but between specialist vocabularies and levels of literacy — are particularly suited to getting to grips with a field where jargon and obtuse documentation can act as barriers to learning.

It is important to remember that AI is actually just an interface to information (its training material). And other interfaces — reading a book or being taught by an expert — are also ‘shortcuts’ to developing expertise (compared to, for example, practice, trial and error).

In fact, almost all learning is a trade-off between experience and efficiency.

And as part of that trade-off we need to recognise the limitations of each approach that we are using: books are helpful but not as effective as experiencing a process ourselves; experts are more interactive but we may find them hard to understand, or their experience may be limited; AI is extremely flexible, but the quality and reliability is variable and needs further checking and exploration.

Privacy

“What is it learning about me?” was the final objection and a reminder that surveillance capitalism — where you become the product — is an important consideration when using these tools.

There are ways that you can reduce the degree to which a large language model ‘learns’ from your activity, such as turning off memory and paying for pro versions of AI tools, but this is certainly an area to keep an eye on in the coming years alongside the data we provide in our search behaviour, browsers and phones.

Last month I wrote about destination and journey prompts, and the strategy of designing AI prompts to avoid deskilling. In some situations a third, hybrid approach can also be useful. In this post I explain how such hybrid destination-journey prompting works in practice, and where it might be most appropriate.

The industry has a destination habit

The idea of ‘journey’ and ‘destination’ prompts is to distinguish between prompts that develop skills and strategies (the journey), and prompts that focus on a simpler endpoint that saves time by ‘skipping to the end’ (the destination).

Using AI for research, for example, tends to involve a destination prompt that asks for an answer to a question. Summarising large or multiple documents, generating images or other content, lead identification, and translation are other examples.

In other words, the most common applications of large language models (LLMs).

These applications might appear lazy, but there are various reasons why we might use AI for these tasks:

• Empowerment: It might be the only way we can do that task (e.g. we don’t have budget for a translator or illustrator; we don’t have time to read hundreds of documents)
• Serendipity: It might spot things that we missed (e.g. lead identification alongside our own idea generation)
• Prioritisation: It might free up time for more valuable activity (e.g. drafting content frees up time for research)
• Efficiency: It might be cheaper or easier to use AI than not doing so.

Those are the opportunities — but there are also risks…

The risks of skipping to the end

There are two risks in particular to consider when using a destination prompt: that the quality of the work will be unacceptable (including but not limited to hallucination); and that automating or delegating that work to an LLM will have a negative impact on you or your colleagues, such as deskilling or reduced critical thinking.

When incorporating AI into a workflow, then, it is important to take steps to manage those risks — just as should happen when choosing to delegate similar tasks to an intern, freelancer, or junior member of the team.

Because delegating is precisely what is happening here — and as soon as you begin delegating you become not just a researcher or writer: you become a manager. [UPDATE: Ethan Mollick explores this in depth in his January 27 newsletter on management as an AI superpower]

One Microsoft study describes it as task stewardship: instead of your time being occupied only with research and writing, it now involves commissioning (i.e. designing prompts), answering questions and providing feedback, and selecting, checking and editing the results for quality and accuracy.

A further risk to consider, then, is the simple one of job satisfaction, and where you feel your efforts are most rewarding.

Enter the hybrid destination-journey prompt

Hybrid destination-journey prompting involves augmenting a basic destination prompt with some consideration of your own role in the process. That might be your skills development, critical thinking, or simply steps that might be taken to ‘manage’ the response.

Below is a simple example of a hybrid prompt, which asks a factual question (destination prompting) but adds a question focused on building verification skills (journey prompting).

Who is the head of English at Birmingham City University? Explain how I could verify that information outside of ChatGPT. 

Regardless of the response, even a very basic hybrid prompt such as this ensures that the user considers the risks involved in using AI, encouraging critical thinking, editorial independence and quality control, and reducing the risk of deskilling.

Here is another example, this time for summarising a document:

I am a journalist doing background research on a story on [topic]. Summarise the document attached. 
Highlight any points I should verify directly in the source or through additional references. 
Suggest three questions I should ask myself before relying on this summary in my work.

The importance of critical prompt design

If ‘writing an effective commissioning brief’ is a skill we need to develop in managing AI, we should also be designing a more detailed prompt that considers the limitations of the tool we are working with — one that incorporates critical prompt design techniques to manage AI’s tendencies towards sycophancy, greediness and hallucination.

Here is an example of a better designed hybrid prompt which does that:

Is it possible to identify the head of English at Birmingham City University? Explain how confident you are in the response and why. 
If the answer is uncertain or unverifiable, say so explicitly. 
Explain how I could verify that information outside of ChatGPT. 
Once you have answered those questions, do not offer to check any further websites.

That prompt adds a number of prompt design techniques:

• An open question rather than a direct order (to reduce the risk of hallucination): Is it possible to identify
• Self-criticism prompting (confidence estimation): Explain how confident you are in the response
• Uncertainty prompting: If the answer is uncertain or unverifiable, say so explicitly
• Negative prompting (to prevent greediness, or scope creep): do not offer to check any further websites

A prompt summarising a document should be similarly designed to counter-balance the gullibility bias of LLMs

Build enough knowledge to judge the responses

A growing body of research on the impact of AI usage on critical thinking suggests that AI usage can have a negative impact unless we have, or build, crucial subject knowledge and an awareness of AI’s flaws.

One (frequently misquoted) Microsoft study identifies two key factors that shape critical thinking (defined as including “analysis, synthesis, and evaluation”): a person’s confidence in AI, and their confidence in performing the task that they are asking the AI to perform.

If a person has high confidence in their own ability to perform a task, and low confidence in AI’s ability, they are likely to use more critical thinking in the task

But if a person has low confidence in their own ability to perform a task, and high confidence in AI, they are likely to use less critical thinking.

The takeaway is clear: we should ideally build knowledge and experience of a task ourself before asking AI to perform it.

If that is not possible, we should be especially rigorous in finding out and understanding what qualities are important in assessing the results of such tasks.

Either way, we should also be conscious that AI suffers from a number of flaws when performing tasks which will need to be checked for.

There are no shortcuts

For example, if we were asking an LLM to draft a news article we should really have experience of writing enough news articles ourselves to be confident of judging the results.

If we don’t then we have the responsibility to at least understand certain things to judge it effectively: the principles of news article structure, grammar and spelling, law and ethics (e.g. accuracy, objectivity), and the style requirements of the target publication are just some that we might list. We might also seek help from someone who does have the experience we lack.

If we don’t have experience or knowledge in an area, use of AI is particularly high risk because we have no way of judging whether it has produced something usable or not, regardless of how impressive it seems (if lawyers repeatedly fall for it, so will you).

Even with experience and knowledge, you will also need to bear in mind AI’s key weaknesses:

• AI’s verbosity leads to a tendency to overwrite;
• Hallucination creates issues around accuracy;
• The training of LLMs on a range of material outside of your context means they are likely to suffer from flaws in style that will need significant adaptation for your country, industry, and target publication.

An MIT study adds another consideration: if we use a muscle more, it becomes stronger; but if we use it less, it becomes weaker.

So, the more we spend our time on new ‘muscles’ like prompt design, interviewing and verification, the stronger those ‘muscles’ should become — but using AI costs us the opportunity to develop other muscles, whether that’s research, summarisation, or writing.

That balance between an opportunity and a cost is something to consider carefully, and a risk to manage too.