X

GenAI, the rise of epistemic noise and the risk of epistemic surrender

Laurence Dierickx

2026-11-19

Generative AI (GenAI) technologies are known for their potential to generate falsehoods, either intentionally or unintentionally. From the mass production of fake news to the generation of convincing content that can fool users, GenAI is constantly blurring the lines between truth and fiction. We can call this informational noise, relating to their ability to generate mis- and disinformation at speed and/or at scale. At the same time, it affects the criteria we use to recognise a claim or piece of content as factual, which we can call epistemic noise.

This presentation was prepared for a seminar organised by the TRANSFO research group at the Université Libre de Bruxelles (ULB) in Charleroi on 19 November 2026.

In October 2024, a journalist from the Swedish public broadcaster SVT asked Copilot whether Tim Walz, who was then the Democratic candidate for US vice-president, had been accused of having inappropriate relations with a minor. Copilot confirmed this without hesitation, citing MSN, Microsoft’s news aggregator. The allegation had never been substantiated. It originated from a disinformation campaign attributed to a pro-Kremlin network known for disseminating false testimonies and deepfakes. The story began life as an anonymous social media post, was picked up without verification by a foreign news outlet, amplified by MSN, and finally summarised by Copilot as if it were an established fact.

This example illustrates a structural problem. Large language models (LLMs) do not retrieve primarily information from structured databases, nor are they built to produce verified information. Producing text that seems plausible do not make thems factual: providing an answer is how they are designed to work. We can understand this feature through two complementary dynamics, which follow distinct logics.

As a strategy, informational noise targets content such as falsehoods, synthetic material and amplified posts. It assumes there is someone behind it with an intention, often as part of an organised campaign. Informational noise as an inherent feature of the LLM manifests as any kind of semantic noise, including hallucinations. In both cases, information can be verified. However, epistemic noise has deeper implications as it affects the criteria we use to distinguish fact from fiction.

From informational to epistemic noise

GenAI makes fake content cheaper, faster and easier to produce, distribute and target. It also floods news channels more quickly than fact-checkers can respond. As technology advances, synthetic media are becoming increasingly difficult to identify, and there is no single detection tool that can cover every technique and provide 100% certainty. New image compression techniques based on neural networks can fool detectors by making real images appear to be AI-generated, and the same goes for text detection. These techniques confuse polished or predictable human writing with AI because they measure statistical predictability, not authorship. The latter is made mandatory by the EU Act but is easily bypassed. In addition, there is a growing body of online services offering to ‘humanise’ AI text, rewriting it so that it passes AI detectors. A whole market has grown up to make undetectable what other tools try to detect.

Tracing sources and verifying facts is therefore the most pragmatic response, as is labelling generated content.

LLMs-based search engines inherit all these flaws, even if the way they function should be distinguished: rather than generating text from scratch, they act as real-time intermediaries, flattening diverse web sources into a single, authoritative voice. This is not without having any impact on journalism.

In October 2025, the EBU and the BBC coordinated a large study on the integrity of information in AI assistants. Twenty-two public service broadcasters in 18 countries and 14 languages assessed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity on current affairs. They found that 45 per cent of the answers contained at least one significant issue, whatever the language or country. Nearly a third had serious sourcing problems, and one in five contained major factual errors or outdated information.

My MSCA LEAD AI Project at the University of Bergen, broadens this scope by studying how it affects information integrity and journalism. Preliminary results already show that journalism is a minority source in search engine’s reports and that playing with languages can lead to different answers to a same question. Biais in the queries are rarely challenged, leading to reproducing the biais in the answer.

Towards an epistemic surrender?

Since LLMs cannot distinguish between authentic and manipulated sources, malicious actors can exploit this instability. There is no need to hack the system: feeding it the right material is enough. A form of model collapse can occur when a user quotes or shares a model’s output that is plausible but unverified. Once republished and recirculated, this output feeds back into the model and the instability grows with each cycle.

While informational noise calls for verification, epistemic noise calls for rebuilding the conditions that allow us to establish what is true.

But what happens when GenAI permeates journalistic content? When journalists admit using LLMs to produce their paper without disclosing it, when they say their role is no longer to write but to collect information, when they get caught publishing inaccurate generated content, when they say, “asking the machine the same way they would ask colleagues”? Can we consider that we are facing a growing movement of epistemic surrender?

What is at stake is human judgement, and with it our capacity to doubt. Speed leaves little room for deliberation and reflexivity, both of which are central to verification. Epistemic surrender begins when we delegate judgement itself and not just tasks; assistance then turns into dependence. Being the author of a text means understanding the reasoning behind it, not just the outcome.

The first draft of this paper was crafted in French, then automatically translated with Google, then humanly polished before using Grammarly and Deepl.write to improve the language quality.