Presentation – Adapting AI Governance to Sector-Specific Challenges: Comparing Perspectives – Workshop IVADO Sept 14-18, 2026, Montréal.
1.
According to the OECD, an AI system is an automated system that produces different types of outputs from the data it receives as input. These outputs can include predictions, content, recommendations, or decisions that influence physical or virtual environments.
This definition encompasses a wide range of AI applications used in journalism, including search engines, fact-checking tools, text mining applications, recommendation systems, and generative AI tools. But before governing AI systems, and generative AI in particular, we need to understand what they actually do to the conditions under which something becomes recognisable as fact. That’s what I’ll talk about today: epistemology.
2.
In October 2024, a journalist from Swedish public broadcaster SVT asked Copilot whether Tim Walz, then the Democratic candidate for Vice President of the United States, had indeed been accused of having inappropriate sexual relations with minors. Without expressing any doubt or pointing out the lack of evidence, Copilot confirmed the allegation and supported it with a highly problematic source: MSN, Microsoft’s content aggregator. Subsequent investigations would reveal that this accusation was part of a Russian disinformation campaign aimed at discrediting the Democratic candidate and, more broadly, influencing the 2024 US presidential election.
3.
The chain of amplification and propagation can be traced schematically. It begins with an anonymous social media post. Then, a foreign media outlet picks up the post without verifying it and treats it as a reliable source. MSN then amplifies the information, and Copilot summarises it in a response that seems factual and trustworthy, even though the allegation remains unsubstantiated. However, the journalist who receives this response instinctively questions the assertion. Copilot’s confidence contrasts sharply with the lack of any primary source to verify it.
4.
While this case is particularly striking, it is not exceptional; the problem is structural. Large language models do not retrieve facts from structured, verified databases. Instead, they generate their responses through probabilistic recombination of patterns learned during training. These responses can appear coherent, confident and authoritative while remaining unverifiable, difficult to trace and lacking accountability by their very nature. A language model has no intrinsic mechanism for systematically distinguishing truth from falsehood. Its primary objective is to produce the most probable word sequence given the context, rather than to establish the truth of the content it generates.
5.
In this case, Copilot cited neither a colleague nor a directly consulted source, nor even a user’s search history. Instead, it cited MSN, Microsoft’s automated news aggregation service. The claim about Tim Walz thus entered the system through a corporate editorial pipeline. Therefore, no one explicitly decided to revisit or verify this claim before it was republished.
This raises two distinct questions from two different perspectives. The first is ethical: who is responsible when this occurs, and what rules should govern the deployment of such systems? Meanwhile, epistemology poses a more fundamental question: what makes this phenomenon possible in the first place, and how can information be treated as fact without ever being verified?
6.
Before going any further, we must acknowledge that ethics itself is not monolithic. Three major traditions structure contemporary debates.
The ethics of duty, particularly associated with Kant, asks what we ought to do, often resulting in rules, obligations and codes of conduct. Consequentialism, notably developed by Bentham and Mill, evaluates actions based on their consequences and foreseeable effects. Virtue ethics, rooted in Aristotle’s philosophy, shifts the focus to the agent’s character, asking what an ethical, prudent, and honest professional should do. In practice, the AI guidelines adopted by European editorial boards in recent years demonstrate the complementary nature of these three approaches. They demonstrate moral pluralism in action.
Accountability is more closely associated with duty ethics, in which moral action is based on the obligation to respect principles and rules to which the agent can respond. In contrast, virtue ethics places less emphasis on institutional accountability, focusing primarily on the agent’s character and integrity. Consequentialism, for its part, prioritises evaluating the consequences of actions, particularly in terms of minimising harm.
For instance, an editor might choose not to publish an AI-generated image in an opinion piece, even if it strengthens the argument, because the risk of it being mistaken for a real photograph outweighs its illustrative value. This is consequentialist reasoning, whereby the legitimacy of an action depends on its foreseeable effects.
Virtue ethics prompts a different question. Rather than asking, « Which choice will produce the best result? », it also asks, « What choice would an ethical, prudent and honest professional make? » The decision reflects an understanding of professional character, with qualities such as prudence, intellectual honesty and respect for the reader becoming important considerations, regardless of the immediate calculation of advantages and disadvantages.
7.
But even combined, these three traditions are insufficient to resolve the tensions that arise in practice: accuracy versus speed, innovation versus harm prevention, personalisation versus diversity, transparency versus explainability, public interest versus privacy, autonomy versus security. In each case, it is not simply a matter of determining which rule to apply, which consequence to prioritise, or which virtue to cultivate. Several legitimate values compete, and none can be maximised without compromising, at least in part, the others.
These tensions reveal the limitations of an approach that seeks to derive decisions mechanically from moral principles. This highlights the need for a register of practical arbitration. The challenge here is to determine which value should prevail in a given situation, at what cost, according to which criteria and under what form of responsibility.
8.
Let’s take five very concrete dilemmas from everyday journalism: summarising a 500-page report with AI under time pressure, illustrating an opinion with an image that never existed, using AI in a context of plagiarism, publishing under a fictitious identity, or even considering AI as a « colleague ».
In each case, ethics and epistemology raise different yet complementary questions. From an ethical perspective, the questions are what is acceptable, who is responsible, and what professional obligations apply. From an epistemological perspective, the question is what still counts as reliable knowledge, evidence or legitimate information.
Automated summarisation raises the question of whether unverified information counts as knowledge. Generated images blur the line between illustration and proof. AI-assisted plagiarism makes a text’s origin more uncertain. A false identity can weaken the link between knowledge and the author’s authority. Finally, the concept of AI as a ‘colleague’ raises the question of who determines what is considered relevant, credible or accurate.
These examples therefore show that the ethical dilemmas of AI concern not only what we should do, but also what allows us to consider something as true, verified, and trustworthy. This is where ethics meets epistemology.
9.
We can see three reasons to consider these two registers together. First, the two registers can fail independently. A process can be perfectly transparent while producing an unverified fact. Conversely, a fact can be verified but circulate within a process where no one is truly accountable.
Secondly, ethical codes address principles such as accuracy, verification, and reliability. However, they are not frequently integrated directly into the procedures and tools used daily.
Third, within an organisation, the same decision concerns both registers. For example, deploying AI in a newsroom is both an epistemic question: what is considered sufficiently verified?
But it is also an ethical question: who is responsible for this decision? We cannot therefore govern AI by looking at only one of these two aspects.
10.
Epistemic governance is a way of thinking about how rules and procedures can structure knowledge production. The key shift is this: in a principles-based ethic, responsibility rests primarily with the individual and the values they must uphold. Epistemic governance raises another question: how can these ethical obligations be directly integrated into procedures and practices?
In other words, it’s no longer simply a matter of telling stakeholders what to do. It’s about creating the conditions that make this possible in practice. And it’s not a question of distributing responsibility among different stakeholders, but of thinking about responsibility differently: operating throughout the entire process and involving everyone who participates.
11.
The idea is to consider the entire information production process, at each stage where AI intervenes. In this sense, we can distinguish three stages inspired by the data journalism model proposed by Paul Bradshaw. This model’s advantage is that it supports a process-oriented approach.
First, compile: search for and verify information. A journalist can consult a source, verify it themselves, or use AI for a specific task. In a newsroom, this collection can be automated and produce content at scale.
Next, combine: format the information. AI can write or summarise text, and teams can check the results case by case. Media organisations, for their part, can systematically integrate these uses into their content production and optimisation processes.
Finally, communicate: deciding what to publish and to whom. In the first case, editorial judgment remains primarily in the journalist’s hands. In the second, personalisation and recommendation systems can influence, sometimes on a very large scale, the visibility of content and its circulation among audiences.
This is where the concept of epistemic governance becomes relevant: a set of rules, procedures, and mechanisms that structure the production, validation, and communication of knowledge, ensuring its quality, reliability, and accountability.
12.
What was described at the beginning of this presentation, however, needs a name before it can truly be governed. The concept proposed here is that of epistemic instability. It can be defined as a situation in which the criteria that let us distinguish facts from non-facts become less reliable or harder to discern, particularly when information is produced by systems whose results rely on probabilistic calculations. We must therefore now understand how this instability changes our relationship to information.
13.
This is precisely where Luciano Floridi’s philosophy of information proves useful. For Floridi, the infosphere is not neutral. It encompasses all informational entities, their properties, and their interactions. This environment shapes both what we know and how we act. Floridi then introduces the concept of semantic pollution, in which the quality and reliability of our informational environment degrade. This pollution does not necessarily stem from a deliberate attempt to falsify information. It can also arise from structural conditions that produce or disseminate unreliable information.
From this perspective, we can understand epistemic instability as a form of semantic pollution of the infosphere at the infrastructure level. While information quality is central to these issues, what is most threatened is our collective ability to know what is true and therefore to act on that basis. In journalism, this is obviously critical, since verification and information reliability are at the very heart of journalistic practice.
14.
Epistemic instability can be understood through three interconnected and mutually reinforcing dimensions. The first is technological: it is generative instability. It stems directly from how a model produces a response, at the moment it generates it.
The second type is media-related: media instability. This occurs when systems come between us and information, providing answers instead of showing sources. The third type is systemic and is known as propagative instability. This occurs when content is circulated, reused, quoted and recycled, ultimately amplifying itself within the information ecosystem.
15.
LLM outputs arise from the interaction between training data, the model architecture, and the user’s instruction at a specific moment, with a specific model. They are emergent facts: plausible in form, but produced by an opaque process that leaves them unverified in substance and highly variable. Two identical instructions can produce two different responses within the same model. Two identical instructions will also produce different results using different models. Therefore, generated outputs lack stability, including in quality.
16.
Some of these examples might seem amusing, but behind their sometimes absurd appearance lies a serious problem. This content presents itself as real information when it is not. The assertive and deceptive nature of these productions can blur the line between verified and fabricated information. For journalists, the challenge is not to be fooled. However, recent examples remind us that it is not so simple.
A news report published by Norway’s national news agency contained several factual errors, including some completely fabricated information. An investigation into how the report was produced revealed that the journalist used a language template to generate a summary, which they then incorporated into the report the agency distributed.
This example is particularly interesting because it shows that the problem doesn’t just stem from AI producing erroneous information. The real issue is what happens when this « information » enters a journalistic chain and is subsequently disseminated as news by a reliable and respectable media outlet.
17.
Mediative instability arises when the systems that provide us with access to information blur the distinction between source and synthesis. The system can produce a response that seems authoritarian, even though we don’t always know where the information comes from or how it was constructed.
Three particularly important consequences stand out. Firstly, authority. The answer is presented as an established fact even though its origin and reasoning can be difficult to verify. Secondly, agency. By constantly receiving direct answers, we risk searching and questioning less, and gradually exercising our critical thinking skills less. Thirdly, substitution. An AI-generated answer does not constitute verification. It is a synthesis. While it can help us access information, it cannot replace independent verification or the responsibility of the person who decides to consider that information reliable.
18.
Search engines that rely on large language models create a new and concerning phenomenon. In a traditional search, the engine returns a list of sources, generally ranked by relevance. The source remains visible, and the user must consult these sources, compare them, and exercise their own critical judgment. The origin of the information therefore remains traceable. This reflects an information-retrieval logic.
With generative AI search, the process is different. The system no longer simply directs us to sources; it searches for various pieces of information, synthesises them, and produces a direct answer. While the user sees the result of the synthesis, it’s less clear what informed it and how the system selected and combined the different pieces of information. The source of the information therefore becomes more difficult to trace.
In the case of the computerised bees, the search engine retrieved online content without verifying its reliability. In this case, the content was an April Fool’s joke published on a website that would otherwise have seemed authoritative. This also illustrates the web’s limitations: an open, heterogeneous space where good and false information coexist. Moreover, the most reliable content is not always the most accessible, particularly when it is behind paywalls.
19.
Mediative instability takes on a particular form when the intermediary is no longer a simple chatbot or a search interface, but a chain of LLM agents. Each agent then transmits its output to the next, which receives it as reliable information rather than a hypothesis to be tested.
No single step guarantees verification. Each agent passes on what it receives until the final agent presents the result as fact. This creates an implicit system of trust, whereby each step relies on the previous one. This well-documented mechanism in distributed systems engineering can cause instability.
20.
Alongside propagative instability, epistemic instability lies in how content circulates. We begin with a plausible but unverified LLM output. A user may trust it, cite it or share it. The content is then published, indexed and redistributed. Other models can then re-ingest the content, thereby amplifying the initial instability. At each stage, nothing is necessarily corrected. The content is transmitted further and gains an appearance of legitimacy with each iteration.
This is where a loop can form: already polluted sources end up reintegrated into subsequent models’ training data. Structurally unstable systems can then amplify disinformation at scale.
What matters is that, unlike deliberate disinformation, propagating instability requires no actor, intention, or strategy. It can simply emerge from circulation and scale.
21.
The snowball effect refers to the mechanism by which a local error gradually acquires an appearance of legitimacy as it passes through a chain of agents. In agentic systems, an agent’s output is generally not treated as a hypothesis to be tested, but as reliable information destined for the next step. An initial error can thus be reformulated, enriched, and contextualised without ever being questioned. With each new transformation, the information gains in coherence and apparent credibility, not because it has been verified, but because it has been processed multiple times. This process makes failures increasingly difficult to identify and correct. A local error then becomes systemic.
22.
Propagating instability does not require malicious intent to produce problematic effects. It can emerge spontaneously from the circulation and amplification of content. However, this instability also constitutes a strategic resource for actors who deliberately seek to manipulate information. In an environment where content is constantly synthesised, reformulated, and reused, it becomes easier to introduce misleading information and let the system disseminate it. Malicious actors then no longer necessarily need to fabricate an alternative reality from scratch: they simply need to exploit an already vulnerable information infrastructure.
This vulnerability can be exploited in various ways, such as creating content intended to be reused as a source, artificially multiplying certain narratives within the information space, or exploiting systems’ tendency to synthesise plausible information without systematically verifying its validity. The aim is to manipulate the mechanisms by which information is selected, synthesised and disseminated.
23.
Two attack vectors illustrate this logic. The first is LLM grooming . This involves saturating the information environment with content designed to be retrieved by AI systems, either during their training or through their search and navigation tools. The goal is less to convince human readers than to shape the knowledge that the models will have access to.
The second is data poisoning. In this case, attackers introduce adversarial data samples or hidden instructions into the datasets the systems use. These interventions can create persistent biases, modify model behaviour or trigger specific responses in certain circumstances without necessarily being visible to users.
In both cases, the attack doesn’t just target the information itself. It targets the conditions under which knowledge is produced. In other words, it exploits epistemic instability to influence what will later be presented as credible, relevant, or true. A study by Anthropic showed that 250 corrupted documents are enough to introduce a persistent vulnerability into a model trained on considerable volumes of data, regardless of its size. A system’s robustness therefore depends not only on the volume of data used to train it, but also on its integrity.
24.
The epistemic instability of generative AI has three direct consequences for journalism. Firstly, it triggers a verification crisis. Search engines initially distorted epistemic authority by ranking content by relevance rather than veracity. LLMs extend and amplify this phenomenon. Rather than simply guiding access to information, they now directly produce syntheses whose selection, prioritisation and combination criteria remain largely opaque.
Next, as generative models become the primary source of information, the traditional mechanisms of journalistic accountability weaken. It becomes impossible to interview witnesses, consult original documents or hold identifiable individuals accountable. Information can be presented as authoritative without any clearly attributable responsibility.
A third risk is that of epistemic surrender. AI systems can help journalists to research, organise and synthesise information more efficiently. However, they do not fulfil journalism’s fundamental epistemic function of investigating, verifying, and establishing facts independently. The risk is therefore not only that certain tasks will be progressively delegated, but also that the very processes of judgement underpinning journalistic practice will be too. Thus, assistance can become dependence and ultimately the abandonment of autonomous verification.
25.
News organisations are therefore no longer content to simply operate in a potentially polluted information environment: they are actively participating in its transformation by disseminating AI-generated content.
This trend is driven by well-known economic principles. The pursuit of audience reach, engagement, speed of publication, and cost reduction can favour fluidity and production scale over verification practices. The expected benefits of these systems sometimes rely more on the promises of innovation than on solid evidence that they improve information quality.
Therefore, the three forms of instability we have identified—generative, mediative, and propagative—should not be viewed solely as external threats facing newsrooms. They can also emerge within journalistic organisations themselves, at the heart of information gathering, production, and dissemination.
26.
This brings us to the final question: that of governance.
The first level of reflection is that of ethics. It asks what humans should do with AI. It establishes values, principles, and responsibilities at both the individual and professional levels. At this level, we find internal editorial guidelines and codes of ethics, generally centred on transparency, human oversight, and accountability.
The second is that of epistemic governance. The question then becomes: how should knowledge be produced, validated, and disseminated? At the organisational level, this refers to implementing procedures, control points, source-traceability mechanisms, and editorial protocols that integrate verification requirements into the process itself. However, these mechanisms are still rarely formalised in practice.
But the epistemic instability we have described does not stop at the boundaries of writing. When it becomes a structural characteristic of informational environments mediated by large language models, epistemic governance must also extend to the technical infrastructures that produce, select, and disseminate information.
This is where the limitations of off-the-shelf tools become apparent, as they largely escape any independent control or audit. A newsroom that adopts a commercial tool often has no access to what happens inside the system and no control over its design, training data, or the choices that underpin it.
27.
While ethics defines the values and principles meant to guide the use of AI, epistemic governance focuses on the concrete conditions under which knowledge is produced, validated, and disseminated. It concerns journalistic practices, newsroom organisational procedures, and, increasingly, the technical infrastructures that produce and circulate information.
In conclusion, the concept of responsible AI in journalism is generally considered a normative or regulatory issue, involving codes of conduct, governance frameworks, and compliance. While this approach is necessary, it is no longer sufficient.
The argument I have made today calls for three shifts.
First, we must move from ethics alone to an articulation between ethics and epistemology. Knowing what is right can no longer be separated from knowing what is true, verifiable, and justifiable.
Next, we must shift from individual decision-making to governing information production processes. The issues raised by AI concern not only the choices journalists make, but also the organisational and technical conditions under which knowledge is produced, verified, and shared.
Finally, responsible AI in journalism is not a state to be achieved. It is an ongoing process of organising the conditions for knowledge production. What is at stake is not only the quality of information, but also the stability of the criteria that make it possible to distinguish factual from non-factual information in increasingly AI-mediated information environments.
Original version in French.
Automatically translated with the help of Google Translate.
Humanly post-edited.
