AI on the job: does mentioning AI adequately inform the recipient?

Public authorities, health services and businesses can now attach an AI-assistance disclosure to their messages. The management question is specific: does this disclosure help recipients understand the scope of the message, or leave them to work out what was produced and checked?

What the literature establishes

George Akerlof (1970, The Quarterly Journal of Economics) formalises a problem of information asymmetry: when quality is difficult to observe, market exchange can deteriorate, including for high-quality offerings. His analysis concerns neither AI nor administrative communication, but establishes that the information available about quality can change how a market operates. By analogy, knowing how a message was produced does not necessarily mean knowing how reliable it is.

A disclosure is not necessarily an informative signal

Michael Spence (1973, The Quarterly Journal of Economics) shows, in a labour-market model, how an observable signal can convey information about a characteristic that cannot be observed directly. The mechanism depends, among other things, on the costs of signalling for different types of actors: the mere visibility of an indication does not guarantee its informational value. Applied cautiously to AI, this distinction encourages us to separate the label “written with AI” from information about the checks actually performed.

The premature conclusion

One might conclude that an AI disclosure is useless unless it guarantees quality; that would confuse transparency about the process with assurance about the outcome. Such a disclosure can meet a legitimate need for information without proving that the content has been validated. The question is therefore not whether to disclose everything or nothing, but what recipients can reasonably infer from the disclosure. (our executive and employee training programmes)

What these studies cannot establish

Both articles are theoretical contributions to economics, not experiments on how people receive AI-assisted messages. They cannot predict either the effect of a disclosure on trust or which wording a particular audience will find easiest to understand. Nor do they settle legal transparency requirements, which must be assessed separately in the relevant context.

A practical check in Berne

As part of SHR’s “AI on the job” programme, an exercise in Berne could compare two versions of a fictional message: one simply disclosing AI use, the other also specifying its role, the human checks actually performed and the service to contact. Cases would be adapted to federal administration, public health, telecommunications and precision manufacturing, without personal data or confidential information. Each participant would randomly receive just one version and identify what had been automated, what had been checked and whom to ask for clarification. The measure would be the proportion of correct answers on these three points, compared between versions within each context: a comprehension test, rather than a simple trust rating. To go further: explore the AI on the Job training in Bern, or browse our executive and employee training programmes in Switzerland.

In pictures: AI on the Job in Bern

• ai on the job
• Bern
• research
• training bern