This paper was initially posted on the RSF website and is republished with the authors’ permission.
The paper is one of the two steering committee’s working documents available to the public. It reviews the challenges raised by AI regarding the right to information. These challenges cover three significant stages in the work of the media in its relationship with artificial intelligence: information gathering, news processing, dissemination, and strategic positioning of media in the AI era. The second document, AI and Media Ethics : Existing References and overview, provides an overview of the existing ethical initiatives in AI and journalism.
As part of its work on a charter on AI in the media, Reporters Without Borders (RSF) invites contributions from the media and civil society organisations, above all those specialising in the media. These contributions will help the committee created at RSF’s initiative to identify appropriate responses for media professionals to the rapid deployment of AI technology. Contributions can be submitted using this link. The deadline for submitting contributions is 11 October 2023.
Challenges raised by AI regarding the right to information
1. Information gathering
This phase encompasses journalistic research, investigation, and gathering raw materials such as pictures or videos. It also involves the tools and methods information professionals use to select news and topics to cover.
Fabricated contents
As generative AI (GAI) democratises, some estimate that synthetic content may soon overwhelm human-made content. A growing concern is the ability to fabricate evidence, such as HD videos or audio recordings. While image manipulation has historical precedents, GAI now allows for the near-instantaneous, cheap creation of synthetic content. The challenge for information integrity lies in discerning fake from reality, and the potential for genuine content to be mistaken as synthetic.
While some place hopes in the coming progress of AI detection tools, many argue that GAI’s ability to bypass these tests will progress at a similar pace. Synthetic content – text, image, or sounds – is already close to undetectable.
Some advocate for the automatic marking of synthetic content at the creation stage. This technique, known as “watermarking”, embeds patterns like combinations of pixels or words in artificial content. Others promote adopting content authenticity standards, which involve attaching secured metadata (author, date, etc.) at the creation stage to ensure the content’s provenance stays verifiable. However, GAI tools that don’t apply such marks and standards would still exist and be used by anyone who feels they have good reasons to.
Influence of AI on media coverage
Social media and news aggregators’ AI-driven recommender systems increasingly influence media coverage. While most journalists tend to think social media adversely affects journalism, a majority of them admit to relying on them to identify which topics and stories to cover, among other use cases.
This strongly impacts media coverage strategies and makes journalism vulnerable to platform leaders’ arbitrary internal decisions and external influence strategies.
Further, time spent on social media is mainly driven by attention-seeking design and recommendation systems, which risks skewing information gathering in favour of sensationalist, divisive, or misleading content. (More on this in Part 3).
Soon, media outlets may be incentivised to use AI predictive tools to pick topics to cover, aligning media coverage strategy with their audience to reach targets, rather than editorial choices or journalistic integrity. Such means may be coupled with search engine bots for gathering content on various topics, diminishing the role of ethics and human judgment.
Unreliability of language models
Conversational search engines are rapidly emerging players that will increasingly shape how we find and access information. Although their outputs may appear human-like, Large Language Models (LLMs) merely simulate an understanding of human concepts such as truth, honesty, fairness, or harm. LLMs do not have an underlying model of reality and do not aim for truth. They’re engineered to generate plausible text, that is, text that is likely to follow from the input they’re given, based on the patterns they’ve learned in their training data set and on the instructions they’re given during the supervised learning stage. As a result, language models often generate convincing falsehoods, sometimes supporting them with false quotes and sources. Because LLMs are specifically trained to produce plausible texts, these falsehoods share the appearance of logic and truth.
Privacy
AI-driven tools are increasingly being used for data collection and surveillance. While these tools can be invaluable for gathering vast amounts of data quickly, indiscriminate data collection can infringe on individual privacy and intellectual property rights. (More on this in Part 3). AI systems may also leak personal information, which has implications for the types of data journalists should or should not share with AI tools.
2. Information processing
This phase involves analysing, organising and contextualising gathered facts and content, independence, impartiality, and “no-harm principle” concerns. As algorithms can analyse data and generate content in seconds, news agencies and media outlets increasingly use them to investigate complex issues or automate content creation based on data and statistics. These include financial reports, sports results, or weather forecasts.
While some newsrooms have quickly integrated generative AI tools like ChatGPT into their operations, as of 2023, only a minority have established guidelines. Many journalists use chatbots and GAI for story inspiration, contextualisation, or illustration. Generated content ranges from templates and outlines to news stories or even sometimes to entire news and information websites operating with little human oversight. The content produced is heavily influenced by an opaque set of training data, undisclosed pre-prompts, and reinforcement learning protocols determined by AI system providers.
GAI tool providers often sideline ethics concerns and do not provide robust guarantees towards journalistic ethics and standards. While they claim to train, fine-tune, and pre-prompt their models to prevent them from serving dangerous goals, their systems can easily be jailbroken and be used for harmful purposes. While there might be little incentive to intentionally prompt a model to provide misleading content from a journalistic perspective, there’s a significant risk of inadvertently receiving such content.
Biases
Biases can be more insidious than factual inaccuracies because they subtly distort perceptions and reinforce systemic prejudices in much less perceptible ways, which are harder to verify. Research has found that in a given text, even the subtlest representation of a metaphor can have a powerful influence over how readers perceive a social problem and attempt to solve it.
LLMs and GAI tools have shown biases in various contexts, from gender and race to political and moral views. Like the issue of unreliability, this is a structural issue. LLMs tend to reproduce or even amplify biases contained in their training dataset.
Audience reach-optimisation
GAI brings a wide range of new techniques to adapt content to the implicit incentives set by social media and news aggregators. LLMs can be coupled with basic methods such as A/B testing for titles and thumbnails or be used to draft entire articles optimised for Search Engine Optimization (SEO) and recommender systems. There’s a growing concern that AI-generated content, designed to maximise engagement, may inherently favour extreme, sensationalist, and divisive narratives, mirroring the trends observed in social media news feeds.
Transparency, explainability, and openness challenges
AI system providers offer varying levels of transparency concerning the design (code, parameters, training dataset, and methodology), performance (reliability, robustness, bias, etc.), and actual use cases (e.g., recommendation systems, military or medical usage) of their products.
Organisations like the UN and several governments have advocated for increased transparency and explainability in AI, raising both technical and ethical questions. On the technical side, while requesting any level of transparency regarding the design, performance, and actual usage of AI systems is feasible, more is needed to make them explainable. Neural networks, particularly Large Language Models (LLMs), are often described as black boxes. Their design inherently prevents the simple explanation of their functioning, making it virtually impossible to clarify how they arrive at specific conclusions.
On the ethical side, while high levels of transparency seem broadly desirable for both performance and usage, complete transparency in design presents risks and complex dilemmas. Requiring providers to release models’ source code, as Meta did with LLaMA, could make these models accessible to reckless or malevolent actors.
Hyper-personalisation
GAI makes it economically and technically feasible to hyper-personalise information in its format and content, potentially skewing public understanding of critical issues. GAI tools are being used to produce synthetic news anchors. In a media landscape competing for attention, such anchors may be modified to match viewers’ preferences in voice and appearance. They may even be designed to foster emotional bonds with humans.
3. Information dissemination
This phase encompasses the selection of the appropriate medium – be it print, broadcast, digital, or social media – and the actual distribution or publication of the content.
Social media recommender systems
As Herbert A. Simon famously put it in 1971, “A wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it.” AI-driven recommendation systems have been social media’s primary response to this challenge. These systems, which organise billions of newsfeeds daily and autoplay hundreds of billions of videos on platforms such as Instagram and TikTok, have been successful from a particular viewpoint: 5 billion people use social media, and a typical user spends around 2 hours and 30 minutes a day on it, predominantly engaging with AI-recommended content.
However, “efficiently” allocating attention has very different meanings depending on one’s goal and perspective. Social media recommendation systems mainly optimise user engagement and advertising revenues, which explains their economic success. As many studies have shown, this tends to incentivise both coverage and dissemination of sensationalist, dismissive, polarising news, possibly at the expense of trustworthy, impartial, in-depth journalism.
AI recommender systems are, at present, the most impactful technology concerning global information dissemination. They significantly influence trending topics online and on social media, considerably shaping the topics and perspectives journalists and media outlets give importance to. Search engines and news aggregators
Search engines also influence media coverage and content dissemination. For all geographical regions, Google Search and Google News account for between 40% and 75% of the total externally driven traffic to publishers’ websites. Generally, recommendation criteria are kept private and may depart significantly from journalistic and media ethics.
Media outlets utilise various internal AI-based solutions to disseminate information, such as news feed apps, search bars, push notifications, and newsletters, all of which can be personalised and optimised for audience reach objectives. An international survey conducted for Reuters Institute’s “Journalism, Media and Technology Trends and Predictions 2023” indicates that two-thirds of publishers are experimenting with AI to “drive story selection/recommendations on [their] website and app”.
Conversational search engines
Conversational search engines are emerging players in the information dissemination landscape. These platforms use natural language to respond to user queries, offering a personalised and interactive search experience. This raises numerous concerns about the criteria these tools employ in prioritising or excluding content according to their interests without checks and balances.
4. Strategic positioning of media in the AI era
For safeguarding the right to information, the prevailing economic and legal incentive structures must favour information organisations and professionals who genuinely uphold ethical values.
Threats to ethics’ economic incentives
Media traffic has become increasingly reliant on gatekeepers competing for short-term attention. As the marginal cost of producing artificial articles and news content approaches zero, a strategy based on publishing a limited number of quality articles may face increasing competition from one based on generating a large volume of reach-optimised ones.
Media outlets might automate daily editorial decisions via AI systems to determine which topics are trending online, how to maximise user engagement based on the outlet’s specific audience, and which angle and format to adopt. Media outlets cautious about generative AI might struggle against competitors who prioritise reach and cost-efficiency over informational integrity. This trend would gradually lead to a situation where the editorial process includes little to no human decision-making and results solely from the automatic aggregation of short-term private incentives.
Media outlets’ relations towards AI providers and information gatekeepers
Several publishers and media organisations have accused GAI providers of using their content to train their models without fair compensation or explicit agreement. GAI tools produce content and build value based on the data they’re trained on, often resulting in text that mirrors specific sentences or images that mimic patterns, leading to unoriginal and copied content. Low-quality news websites also use AI to rewrite content from mainstream news outlets deceptively. Early court decisions have prevented copyrighting AI-generated visual content in the USA. Further, foundational models have been accused of absorbing a vast amount of publicly available information and content and walling it off into proprietary products.
Conversational agents may divert media outlets’ traffic while at the same time using their data, leading to advertising and subscription revenue losses. If not compensated for, this could threaten journalism’s economic sustainability and give increasing importance to anti-competitive, web-destroying, but compelling AI chatbots.
Media outlets also confront a growing asymmetrical power dynamic with social media companies. This disparity is evident in several areas, including advertising revenues and editorial decisions, and amplifies as advertisers increasingly shift from traditional media to social media platforms.
Journalism automation
GAI suggests a potential for significant staff reductions in media outlets. Some (poor quality) news sites already publish articles mainly written or entirely by bots. Others may soon follow suit, as well as tech companies. Because GAI systems do not directly collect facts by conducting interviews, taking pictures, or attending meetings, they can only update their view if they are explicitly trained on new data to consider new information. Few original materials may be collected if substantial media revenues are diverted from information gathering.
This document reflects the point of view of its authors. It is intended to illustrate the scope of the discussions of the “Charter on AI in the media” working group.