AI disclosure is not a truth signal

On 2 August, the EU's transparency rule for AI systems went into effect. The rule requires individuals conversing with a machine to be informed of its artificial nature. The theory behind this rule is to make the artificial entity more visible, prompting people to scrutinize it more critically. However, a closer examination of the rule reveals it may not serve its intended purpose.

In a piece published on 14 October, the author argues that the disclosure of AI may not address the more significant question of the competence and honesty of the entity behind the AI. Two studies published in 2026 provide evidence that the label arrives after the judgment it was supposed to inform. In the first study, disclosing that a caller was an AI did not reduce conversion rates compared to not disclosing it.

In the second study, AI agents still underperformed human agents, but the gap was attributed to process discipline in sales calls rather than customers perceiving the machine as unreliable. In an online experiment, customers largely identified AI voice agents, regardless of whether the disclosure was made, and rated both disclosed and undisclosed AI similarly.

The label added a forecast of what the next minute with the AI would cost the customer – less empathy, more effort, and a company that was not sincere. One study found that AI disclosure lowered the perceived credibility of accurate information and increased the perceived credibility of misinformation. However, this effect was moderated by negative attitudes toward AI.

The label seemed to relocate trust based on veracity, with false information receiving more trust and true information receiving less trust. The label appears to function as a cue rather than a truth signal, informing readers about the kind of interaction to expect and shifting trust accordingly. This cue can have unintended consequences – if the content is correct and labeled as artificial, the credibility of the content may decrease; if the content is false, the credibility may increase.

This bidirectional effect suggests that the label is a proxy for an expensive judgment and a poor measurement. The findings of these studies indicate that the label should not be used as a truth signal. Instead, the label should provide only the information it can warrant – that the text was generated by a specific system at a specific time (provenance claim).

The EU's deadline for machine-readable marking of synthetic content from existing systems is 2 December, but this mark is designed for machines and does not aim to help the person reading the sentence. The visible label is asked to perform a job it cannot do on its own – tell a person what is true. A more honest approach may be to have a label that merely indicates where a sentence came from, without attempting to judge its truthfulness.