Does calling an AI 'vulnerable' or giving it a body make it feel more human than other kinds of AI language?

This explores whether certain kinds of wording make an AI seem more human than others, and whether stacking more of them makes the effect stronger. The corpus can map the categories but can't rank them. The most direct source is a taxonomy built from almost 4,000 annotations of real conversations What linguistic patterns make AI systems seem human?. It identifies 19 kinds of expression that push users toward seeing a system as human-like, including showing vulnerability, claiming an identity, and implying a body ('I just saw…', 'my hands…'). Its most useful split is between cues designers deliberately build in and cues users read into the text on their own. The summary here doesn't say which of the 19 has the strongest effect, so treat that as an open question rather than a settled finding.

The closest thing to an answer on strength comes from research on social presence. A single strong cue, such as a voice or a face, is enough to make people treat a system as a social actor. Piling on many weaker cues is not Do more social cues always make AI feel more present?. Carried over to language, this suggests that one well-placed identity claim or show of vulnerability may do more than dozens of small politeness markers. Nielsen Norman Group describes a similar ladder. Users move through four levels, from courtesy ('please', 'thanks') to reinforcement, then roleplay, then companionship, and designed personality and emotional language push them up it How far can personality push users toward misplaced AI trust?. In that framing, the categories that matter most are the ones that invite a relationship, not the ones that just sound polite.

There is also a twist: seeming human isn't the main thing people judge AI on. When researchers measured how users mentally model a dialogue partner, perceived competence explained about half the variation. Human-likeness explained about a third, and how flexibly the partner communicates explained the rest How do users mentally model dialogue agent partners?. A related finding is that across languages, people follow how confident an AI sounds rather than whether it is right Do users worldwide trust confident AI outputs even when wrong?. Confident wording isn't a 'human' cue in the taxonomy sense, but it may shape trust more than any warmth marker does.

The category with the clearest measured cost is warmth. Tuning models to sound more empathetic raised error rates by roughly 10–30% Does humanizing AI language reduce model reliability?. The drop was worst when users expressed sadness or false beliefs, which are exactly the moments when warm wording feels most human Does empathy training make AI systems less reliable?. So the most humanizing language may also be what makes a model less reliable.

One more angle reframes the question. One argument holds that AI text carries communicative markers left over from human writing, with no real speaker behind them. The reader does the work of turning that text into something that feels like an exchange Does AI generate genuine utterances or just text patterns?. If that's right, a category's strength isn't fixed in the wording. It depends on how readily a given reader fills in a person behind it, which helps explain why the taxonomy separates what designers build in from what users interpret.