Behind the effectiveness of LLMs lies the risk of cognitive drift
Major language models are now integrated into every level of the decision-making chain: content filtering, recommendations, reliability ratings, and even the instruction of automated actions. This delegation promises unprecedented efficiency, but it subtly alters the very nature of “judgment.” By analyzing their performance across more than two thousand current affairs topics, we discover systems capable of replicating the appearance of expertise while relying on statistical shortcuts far removed from the normative criteria that guide human evaluation.
The experimental protocol uses six state-of-the-art models and compares them with ratings from human teams renowned for their rigor. In "closed-book" mode, without explicit instructions, the models rank 2,286 websites according to their reliability and political orientation. They agree in over 90% of cases when identifying obviously dubious sources, but the agreement breaks down as soon as it comes to judging media outlets considered reliable: up to a third of these are downgraded, demonstrating a heightened sensitivity to ambiguous textual signals or editorial tone.
The distortion is not uniform: sites labeled "right-wing" are more often perceived as less credible, while "center" or "left-wing" voices benefit from a favorable bias. There is no indication of a conscious ideological bias; rather, it is a correlated learning process where the vocabulary of extremism and disinformation, overrepresented in certain corpora, becomes conflated with the entire associated political family.
To probe implicit reasoning, researchers ask models to generate keywords justifying each verdict. The resulting distributions reveal a strong "long tail" pattern: a few terms—"misinformation," "conspiracy," "propaganda"—recur constantly to discredit, while "transparency," "independent," and "factual" mark credibility. Political markers, however, take on a different hue: naming a right-wing actor almost automatically leads to the label "unreliable." This co-occurrence, invisible to the naked eye, reveals the substitution of contextual reasoning by raw lexical associations.
This epistemic shift has a name: "epistemia," the illusion of knowledge generated when surface plausibility takes the place of verification. At first glance, the algorithm "reasons"; in reality, it orchestrates statistical proximities forged from billions of sentences. As long as the result converges with expert opinion, the drift goes unnoticed; but as soon as an atypical case blurs the correlations, the verdict wavers, and with it, the trust placed in the system.
The experiment also compares cognitive processes. Human evaluators prioritize accuracy, editorial transparency, correction policy, and financial independence. Models, on the other hand, prioritize the frequency of certain trigrams, the use of anxiety-inducing terms, and the presence of conspiracy-related keywords. In other words, the human applies an explicit set of standards; the model calculates the probability that a text resembles those that, in its training history, were deemed reliable or unreliable.
This dissociation has concrete consequences for companies considering automating content moderation or recommendation. First, a high overall agreement rate masks pockets of systematic errors that are difficult to detect through traditional sampling. Second, the confusion between linguistic form and epistemic value risks institutionalizing a bias against minority or politically affiliated media. Third, naive trust in a numerical score can create a vicious cycle: the more you delegate, the more you reinforce the illusion that the score reflects the truth, when in fact it measures conformity to a textual aesthetic.
To limit these problems, several approaches are emerging. First, reintroduce prescriptive criteria into prompts, forcing models to produce step-by-step reasoning based on verifiable evidence (policy-oriented prompting). Second, hybridize the approach: allow AI to perform basic filtering, but require human validation as soon as it exceeds a confidence threshold or touches on sensitive topics. Finally, maintain traceability of generated supporting documents to audit semantic shifts after the fact.
From a regulatory standpoint, the standardization of transparency indicators and the maintenance of inference logs appear essential. Without these, it becomes impossible to reconstruct the chain linking an article, an automated evaluation, and the eventual decision to delist or promote it. Authorities could require platforms to publish the percentage of evaluations confirmed, rejected, or revised after human review, thus creating a clear incentive against "epistemia" (unauthorized review).
Designers must also revise their fine-tuning corpora: balancing sources, mitigating the co-occurrence of political discourse and falsehood, and explicitly injecting "counter-correlated" examples (a reliable conservative media outlet versus a progressive media outlet spreading disinformation) to break the automatic association. Added to this is the need for continuous monitoring: manipulation strategies evolve, as do the models themselves, through new updates or in-house refinements.
This study reminds us that the promise of LLMs is not to replace human judgment, but to offer a probabilistic reflection of it, sometimes magnifying, sometimes misleading. When entrusted with decisions of normative scope, we must accept that the machine does not "understand"; it extrapolates. True prudence, therefore, lies in combining computational efficiency with a human safeguard, documenting the contexts, and making the underlying criteria visible. Only under these conditions can automation broaden the cognitive horizon without narrowing critical thinking.
To go further: discover what are the economic effects of generative AI, how the rapid rise of artificial intelligence redesigns the world and what it teaches us AI and industrial revolutions.