26 March 2025

Recording

10.00 AI and the future of news. What we know so far.

Mitali Mukherjee, Acting Director, Reuters Institute
Richard Fletcher, Director of Research, Reuters Institute

Notes

  • Research conducted over past year:
    • Talking to newsrooms, 80 per cent saying somewhat or totally transformative in newsrooms.
    • Exploring product, RI exploring how chat bots respond to questions about elections
    • Talking to audiences, six country survey on what people think about generative AI in news.
  • Going forward repeating these projects.

10.15 Panel: Power, profit and plurality in the age of generative AI

The digital media landscape is heading for major disruption as generative AI is being used by AI developers, platforms and creators to find new ways of creating, accessing, and monetising information. The emergence of large language models from OpenAI, Google and others has precipitated an urgent strategic challenge for news publishers: how to maintain relevance, protect intellectual property, and secure sustainable revenue streams in an increasingly AI-driven information ecosystem.

This panel will critically examine the complex power dynamics emerging between news organisations and AI developers, focusing on the critical challenges of content licensing, data valuation, and strategic positioning of publishers.

Felix Simon, Research Fellow, Reuters Institute
Andrew Strait, Associate Director, Ada Lovelace Institute
Matt Rogerson, Director, Global Public Policy and Platform Strategy, FT
Klaudia Jaźwińska, Researcher, Tow Center for Digital Journalism
Federica Cherubini, Director of Leadership Development, Reuters Institute (moderator)

Notes

  • Klaudia – 34 interviews with news leaders. Concern about the rise of third party summarisation products. Impact on plurality (single product summarising multple producers), brand integrity and dilution, audience engagement and value exchange.
    • Most of the news industry won’t be able to make deals.
    • Quantitative research looking at chat bots using real-time data and their ability to cite newsrooms. 20 different US publishers – some direct agreements with AI companies, some in lawsuits, others no formal relationships. Found articles, pulled excerpts, asked major chat bots to identify the source.
      • 1600 queries.
      • Rather than decline to answer some were confidently wrong.
      • Premium models (Perplexity Pro, Grok 3) provided more confidently wrong than free counterparts.
      • Some platforms retrieve information from publishers that had intentionally blocked their crawlers. Eg, Nat Geo blocks Perplexity but they identified 9/10 articles.
      • Some identified the article but hallucinated the URLs.
      • Content licensing appeared to give no guarantee of citation. SFC has deal with OpenAI but ChatGPT could not identify SFC content.
  • Andrew discussing different types of deals between tech companies / media companies.
    • Three forms of deals:
      • Training or finetuning data deals, building out the model. Give us access to all your historical data. News data is high quality – it’s edited, it’s factual. It’s not like social media data. It helps train the ‘morphology of data’ – how the LLM can express itself in credible ways.
        • Andrew is sceptical that these deals will be renewed. AI companies need the totality of the data once to facilitate training – after that the morphology is set and the value goes.
      • Inference deals, ability to pull live information to answer queries.
        • Inference deals are important longer-term.
      • Third type of deal covers both use cases.
    • There’s a sense of FOMO driving deal-making at both media companies and AI companies.
    • There’s not good tools to independently assess whether deals are adhered to.
  • Felix states that there’s very little sense about the value of AI – certainly at media companies and government, but also at tech companies – which makes it very difficult to strike proper commercial deals or to regulate.
  • Matt says OpenAI is making more deals because it doesn’t have the data corpus that Google or Microsoft do through their search.
    • Mentions News Media Bargaining Code and collective bargaining overseen by ACCC as a mechanism for enabling agreements between entities of different power and scale.
    • Tech platforms have enormous power – arguing to the White House that copyright shouldn’t exist, positions that extreme.
      • Klaudia elaborates – Trump EO establishing AI Action Plan. Google and OpenAI called to water down copyright. Codify right to train on public data without restriction. OpenAI argued in its response that DeepSeek / China don’t care about copyright and Western companies will fall behind.
  • Andrew suggests there may need to be ‘social innovations’ – legal innovations – that capture technology properly. Copyright is a part of it but not everything.
  • RAG Product: Retrieval-Augmented Generation, process of optimising the output of LLM. Google, Microsoft, Brave. Panelists mention the “RAG issue” a few times but I missed what that is.
  • Panel doesn’t believe that opt-out solutions will work – hard to enforce, hard to know who is breaking agreements, easy to override.
  • Andrew says that AI product development is (publicly) argued as a utopian agent – will solve the climate and book your travel – but nobody really believes that is what is on the horizon; but it’s hard to argue the needs of news if the other side is talking about curing cancer.
  • Matt says there’s a first challenge which is that we still don’t agree what a good media ecosystem looked like. It has degraded based on previous technological change and we’re still catching up to that; without it how do we assess the impact?
    • Andrew says that in workshops conducted in Belfast, Sheffield and London people reflect that news has become harder to access, understand, and they miss understanding the world. There’s an appetite for a solution.
  • Klaudia says insufficient expectation setting. Search on GenAI is presented as a natural extension of search engines, but they’re not. Could be value in chatbots being required to offer confidence or honesty when they can’t access information rather than defaulting to pleasing the user.

11.15 How audiences feel about AI uses for news

Our researchers Richard Fletcher and Amy Ross Arguedas on how audiences feel on different AI uses for news drawing on findings from the Digital News Report 2024 and our survey on public attitudes on generative AI.

Richard Fletcher, Director of Research, Reuters Institute
Amy Ross Arguedas, Postdoctoral Fellow, Reuters Institute

Notes

  • Six country survey – Argentina, Denmark, France, Japan, UK, USA. >2k responses each.
  • What are you using generative AI for? 5 per cent avg. reported they had done it. 10 in US, 2 in UK / Denmark.
    • From a separate study Richard reflects that ChatGPT is bad at news and this might be why low usage – only pulled latest news 10 per cent, 30 per cent it gets old news (from earlier that day to years ago), 52 per cent refuses – “I’m unable to do that”.
    • Around half of news sites blocking ChatGPT via robots.txt at start of 2024.
  • How do people feel about news ‘mostly’ made by AI with human? Avg 15 per cent comfortable, 50 per cent uncomfortable. Most uncomfortable with crime, politics, local news.
  • How do people feel about news mostly made by human with AI? Avg 38 per cent comfortable. Most comfortable with sport, entertainment.
  • Also different based on process – more comfortable with transcription, moderately comfortable with delivery, uncomfortable with creating content. Very uncomfortable with generated visual imagery.
  • People assume journalists are using AI for behind the scenes work, not for creating content. Only 32 per cent think journalists are always or often checking generative AI output in their processes though, with high uncertainty.
  • Desire for transparency highest for creating images, writing articles, data analysis. Lowest for spelling and grammar, headlines, making charts.
  • Eight per cent think AI will make journalism more worth paying for, 41 per cent less.
  • People tend to think AI will make news cheaper to make and more up to date, but less transparent and less trustworthy.

11.30 Panel: Are journalists telling the most relevant AI stories to their audiences?

A large part of the industry’s focus has been on innovation, use and challenges of using AI in newsrooms. Three of our Journalist Fellows discuss whether journalists are reporting, documenting and sharing the most impactful AI-related stories to global audiences. What lessons can we share from different news outlets? And how do we report on AI keeping the needs of our audiences first?

Sannuta Raghu, Head of AI Lab, Scroll
Katharina Schell, Deputy Editor-in-Chief, APA
Jazmín Acuña, Editorial Director, El Surtidor
Eduardo Suarez, Head of Editorial (moderator)

Notes

  • What stories on AI do you seen in your countries?
    • Jazmín says in Paraguay it’s typically tech stories, product stories, or they’re focussed on government / economics – a new data centre, foreign investment. Should talk about the social purpose of AI. Latin America is a focus for data centres due to cheap energy but it’s impacting communities with existing resource insecurity. Also linguistic inequality; national security impact of models trained or owned by foreign entities.
    • Sannuta says that India is heavily focussed on engineering as a discourse and discussion is mediated through that tech. Needs to be more coverage about environmental impact – eg, facial recognition tech was used at Kumbh Mela. What impact? Reflects Indian may be the capital of AI slop. Not lots of discussion of products beyond LLM.
    • Katharina says that in Austrian media it’s often either optimistic or scared; buying into the hype but also clueless about it. EU regulation should be a better part of the story – the EU AI Act.
  • Sannuta. Indian election. AI tools widely used in translation during election. Old leaders were resurrected via deepfakes for campaigning. Consultants building businesses out of off the shelf tools of politics – confident this is a workshop for other countries.
  • Katharina. Reflecting on Italian newspaper publishing AI edition, felt that the coverage of this was insufficient. Eg, the Guardian said that the sentences were grammatically correct – fine, but that’s a shallow view.
    • Climate change is an analogy here – it’s going to come up in every type of story and all journalists need some level of expertise here. It’s not enough to have a specialist.
  • Jazmín developed chatbot agent that would be a woman in jail on drug trafficking charges. A way for audience to engage with broader story of the war on drugs. Did not use generative AI for the answers – they were from interviews – didn’t want to risk hallucination. Did use AI to test how audiences might explore the story, what questions they would ask – ‘co-design’ with audience.
  • Sannuta. Trying to identify the ‘fundamental unit of information for news’. What is a news atom? How can you bring multiple atoms together for new products? Platform agnostic. Trying to identify exactly what audiences want from articles and move on from the article format.
  • Katharina. B2B newsroom. Developed tool to provide alt text on infographics – retelling the narrative of the figures. Reflects that experiments are one thing, but embedding it in processes is a much harder task. Many publishers operating in ‘legacy systems’ with no capacity to deploy products.
    • Tech is not the only driver of workflow so it won’t totally overhaul organisations.
  • Sannuta. Built a text to video tool to quickly produce video content for social, have human in the loop policy, found that the tool put enormous pressure on the individual responsible. Human in the loop becomes the bottleneck as well – the tool can produce 10k videos a day but the human can only review 15.
  • Jazmín reflects that Guarani is an oral language so there’s no corpus for training.
    • Sannuta adds that many Indians mix English and Hindi; and that written and spoken Hindi are different.
  • Katharina reflects that humans have to adjust the ‘German’ to ‘Austrian German’.

13.45 Public attitudes towards platforms and AI

Our Senior Research Associate Rasmus Nielsen on public attitudes towards platforms and AI, drawing on the findings of our recent report on digital platforms and on the Digital News Report 2024.

Rasmus Nielsen, Senior Research Associate, Reuters Institute

Notes

  • Presentation drawn from three studies:
    1. 1. How the news ecosystem might look in the age of generative AI
    2. 2. What does the public in six countries think of generative AI in news?
    3. 3. What do people want? Views on platforms and the digital public sphere in eight countries
  • The public navgiates based on cautious technology optimism and platform ambivalence. People recognise the problems (misinfo etc) but view platforms as having made their lives better.
  • Across eight countries (Argentina, Brazil, Germany, Japan, South Korea, Spain, UK, USA), platforms types broadly (social media, search engine, messaging app, gen ai) are viewed as having positive effect on both individuals and society.
    • Social media mixed unfavourable in both categories in German; bad for society in SK, UK, USA, flat for society in Japan. Favourable in Argentina, Brazil.
    • Search engines highly favourable in all counties; highest in Argentina (+70/+64), lowest in Germany (+47/+38).
    • Messaging apps broad spread; highest in Argentina (+53/+58), lowest in Japan (+11/+8) and South Korea (+12/+7)
    • Gen AI highest in Argentina (+37/+34) lowest in Germany (+0/-9).
    • People might be wrong in their views, but it’s still a social fact.
  • How much do people think Gen AI will change society?
    • Majority view large impact for tech companies, science, news. Plurality view large impact for ‘ordinary people’, retailers. No category majority or plurality ‘small impact’
  • Do people trust different sectors to use it responsibility?
    • Varies widely by country. Broadly high trust in healthcare and science, low trust in tech companies, government and political parties.
    • News media trust: USA / Argentina 30%, Japan 23%, Denmark 21%, France 18%, UK 12%.
  • In combination of previous two, people expect generative AI will have high impact on news media but don’t trust to use responsibility. In company with social media.
  • Generalised scepticism about political information received on different platform types. Despite this, continued openness to self-regulation for tech sector. Notable higher willingness to see governments involved in generative AI willingness.

14.00 Panel: How three prominent newsrooms are using AI

Many news organisations have moved from experimentation to implementation when it comes to AI. But how have some of the leading international news organisations made use of generative AI and for what? What were the benefits, but also the hurdles they faced along the way? How did they overcome them and what work remains to be done?

Dylan Freedman, Machine learning engineer and journalist, New York Times
Liz Lohn, Director of Product, AI and Editorial Tech, FT
Nathalie Malinarich, Director, Growth, Innovation and AI, BBC
Felix Simon, Research Fellow, Reuters Institute (moderator)

Notes

  • Dylan, gen AI as background element in investigative journalism at NYT.
    • Parse TV appearances. Facial recognition to analyse thumbnails of shows to identify when Tulsi Gabbard was speaking.
    • Same approach but focussed on Pete Hegseth talking about alcohol on Fox. AI was able to intuit that an exchange – one person said ‘Hendricks’, referring to a place, and Hegseth quipped that it ‘made him think about something else’, meaning gin – was alcohol-related. Human in the loop and reviewing everything.
    • Screenshotting government websites; used OCR to pull text and LLM to identify differences between them. Could go beyond keyword search and understand different uses of eg. ‘equity’ as social or financial. Could also do it in Spanish.
    • How have NYT Travel recommendations changed over 20 years? Eg ‘sustainability’ not a popular word in that whole period (keyword search) but the ideas were still there and could be identified (semantic search).
  • Liz, FT
    • Developed an internal tool called AI Playground that links FT’s content with an LLM. Allows newsroom to play around with prompts. Eg, generate briefs on everything Trump has said about Greenland over 20 years.
    • In April 2025 will deploy AI-generated prompts to drive people to comment sections and engage with the content. Considered AI summaries of comments but it seemed risky.
    • Article summary into bulletpoints. Have heard that readers are already copy-pasting their content into other AI platforms for this purpose so might as well provide this service. Low hallucination level – small amount of condensed content. Human in the loop still – editor checks before publishing.
  • Nathalie, BBC
    • Two types of tools – those that help journalists work, those that help distribute.
    • Building a deepfake detector, provides confidence scores. Can search Google to inform its views but assesses images as well.
    • Sport: Built a tool to pull local live radio commentary, transcribe, identify key moments (goals, penalties) and then summarised onto a live blog. Finding higher usage than would get from local radio feeds.

14.45 Panel: How will AI reshape society?

Journalism is changing through artificial intelligence. But this technology is also leaving its imprint on other parts of society, from politics to education. Bringing together experts from various fields, our final panel will discuss the wider applications of AI and society – as well as the risks thereof – and ask what is being done to steer AI in a beneficial direction for everyone.

Victoria Nash, Director, Oxford Internet Institute
Chris Summerfield, Director, UK AI Safety Institute
Roxana Radu, Associate Professor of Digital Technologies and Public Policy, Blavatnik School of Government
Mitali Mukherjee, Acting Director, Reuters Institute (moderator)