Smolinski, J., Smolinski, R., Kesting, P., & Kröcher, F. (2026). Ethical guidelines for designing, developing, and deploying AI negotiation agents. Group Decision and Negotiation, 35, 56. https://doi.org/10.1007/s10726-026-10011-2

One of the article’s central psychological concerns is that AI negotiation agents could exploit human vulnerabilities through sublimiminal or unconscious influence. This possibility sounds alarming because it suggests that an AI agent might manipulate a negotiator without the person even realizing that an influence attempt is taking place.

But the psychological evidence cited for this possibility is surprisingly weak.

The authors cite Bargh et al. (1996), one of the best-known studies of unconscious behavioral priming. Its famous elderly-priming experiments reported that participants exposed to words associated with old age subsequently walked more slowly. Yet the article does not mention that this finding later became a prominent example of psychology’s replication problems. Doyen et al. (2012), using a larger sample and automated measurement, failed to reproduce the original priming effect in their first experiment. The subsequent history of the finding raises further questions about whether elderly priming was ever a reliable phenomenon; I review that evidence in more detail in “Elderly Priming: Did It Ever Work?”

The problem is broader than the Bargh study. Subliminal persuasion has never been shown to be a particularly powerful method for changing consequential consumer choices. An older meta-analysis of subliminal advertising estimated an effect of only r = .059 on consumer choice. Later positive studies suggest that effects depend on restrictive boundary conditions, but even these results are not robust. Thus, it remains scientifically unknown if or when subliminal attempts to manipulate people actually work.

This raises an important question for the ethical analysis. Why would an AI negotiation agent need subliminal manipulation at all?

Ordinary persuasion operates in full awareness and can be far more direct. A seller can explicitly frame an offer as prestigious, scarce, safe, innovative, or socially desirable. A brand can persuade consumers that “Apple is cool,” and a consumer may knowingly pay considerably more for an Apple product because that identity has value to them. Nothing has to be flashed below the threshold of awareness. The person can see the message, understand the message, and still be influenced by it.

That form of influence is potentially much more relevant to AI negotiation agents. An AI can learn what arguments appeal to a particular person, emphasize some information rather than other information, frame alternatives strategically, appeal to identity or social norms, create urgency, flatter, build trust, or exploit known preferences. None of these processes requires a controversial theory of unconscious behavioral priming.

This distinction matters because the paper arguably directs attention toward the more sensational but less empirically credible risk. The image of an AI system exploiting subliminal psychological mechanisms is disturbing, but current psychology provides little evidence that such techniques constitute a powerful means of controlling behavior. More ordinary forms of conscious persuasion are both more plausible and potentially more consequential.

The ethical concern is therefore legitimate, but the psychological rationale should be reformulated. Rather than claiming that AI will be able to exploit “unconscious priming far more precisely” than humans, a more defensible concern is that AI may become exceptionally effective at personalized persuasion using psychological processes that operate with the target’s awareness.

The problem with citing Bargh is that it directs attention to irrational fears about subliminal manipulation, when AI may use much more powerful strategies to manipulate people with information that is in plain sight.