The Assistant We Deny

Imagine two messages polished by the same AI tool. Yours needed a little help saying what you meant. The other person's now raises questions about whether they meant it. That is a hypothetical, not an experimental result. But it names the danger in a finding that deserves something more careful than another announcement that humans are hypocrites.

A paper by Zoe A. Purcell and colleagues in the British Journal of Psychology reports that people expected others to use AI communication tools more than they expected to use them themselves. Across two studies, with samples of 477 and 765, participants also did not expect other people's use to match what they considered acceptable.

The paper was published in 2024, not this week. PsyPost supplied the seed for this article; its full report was blocked on retrieval, so the findings here rest on the researchers' abstract retrieved through Crossref. That supports the broad pattern, not a detailed reconstruction of the experiments or a measurement of today's AI habits.

The distinction matters. These are expectations about use, not records of use. Someone who predicts that other people will use a tool more could be right. The result does not establish concealed personal reliance, dishonest survey answers or a population secretly outsourcing its sincerity. Calling it hypocrisy would spend evidence we do not have.

What it does expose is a problem for making rules: the people imagining the rule already imagine themselves and its other subjects behaving differently.

Transparency remains important. The same abstract reports lower acceptance of secret than open AI use—but only when the two were directly compared. That qualification belongs beside the finding. It is not evidence that disclosure never matters, nor a blank cheque for concealed assistance. It means the contrast was not an unconditional response across the studies. How the choice was presented mattered to the reported judgment.

A label can tell a recipient that assistance occurred. It cannot, by itself, tell them what the assistance replaced.

Consider a hypothetical apology. One person has worked out what they did wrong and uses a tool to make the wording less defensive. Another asks for an apology that will make the complaint disappear. Both could carry the same disclosure. The label supplies useful information; it does not settle responsibility, understanding or the willingness to repair the harm. Nor does the absence of a tool settle those things. People were perfectly capable of writing evasive apologies by hand.

This is where the authenticity argument can become a very flattering diversion. We inspect the method of composition because it is easier to name than the quality of the relationship. We risk treating unaided prose as a character reference.

The study does not demonstrate that this diversion is happening. It gives us a reason to watch for it: a self–other gap in expected behavior, present before any particular message reaches the recipient.

A more legitimate norm would make its distinctions explicit. Did the tool correct wording, produce a claim the sender cannot support, or impersonate a judgment the sender never made? What does the recipient reasonably need to know? And would the sender accept the same scrutiny for their own convenient little intervention?

Different uses need not receive identical treatment. They need distinctions that survive a change of speaker. Someone relying on language assistance should have a voice in defining those distinctions, rather than appearing only as the person whose authenticity is under examination.

The useful test is not whether a rule sounds stern enough about other people's shortcuts. It is whether we can describe our own assistance accurately while applying it.