People don't talk to a machine the way they talk to a person.

The score was never the story

AI moderation reproduced the human study: all 14 themes, saturating early, at ~80% of the depth.

That number is the headline, and it holds. But a score flattens something the transcripts make loud: the human and the machine ran two different kinds of conversation, and people met each one in a different register. The findings matched. The texture did not.

Two moderators, two conversations

Each ribbon is one interview; segment width is how many words that turn ran. Warm is the moderator, slate is the participant. Two things read without a single word: the human interviews run more than twice as long, and their words split evenly between the two speakers. In the AI interviews the moderator carries more of the words, its reflective summaries taking about three words for every two the participant speaks.

The rhythm is measurable. Counting only the moderator's own speech, the two ran structurally different conversations.

The human asks fewer questions but runs a longer, near-balanced exchange. The AI asks slightly more, in half the words, and still out-talks its own participant. The depth follows: blind-coded Signal Depth (a 0-to-5 read of how much the participant gave) was 4.76 for the human against 3.82 for the AI, and it holds even per minute (the human draws about 1.4 times more words per minute, not just more time on the clock). On moderator effectiveness the two are tied, their confidence intervals overlapping. Two registers, not a ranking.

Two repertoires

The rhythm comes from different toolkits. The AI's basic unit is mirror-then-question: it reflects your last answer back, then asks the next scripted thing. Neat, even, and economical.

So the combination of long videos, reliable search, and the fact that it is completely free without asking for payment information makes you feel good about using the site…The AI's signature turn, before the next question · George arm

The human's repertoire is relational. She gives before she takes, reframes the script on the fly, and catches herself when a question leans. Same guide, two ways of running it.

Moderator-move coding over all 160 interviews. One human moderator, so "human effects" is this one researcher's craft, not humans in general.

Query vs conversation

To the machine, people talk like a search bar. To the human, they talk like a person.

Take every participant word, split it by which moderator heard it, and ask which are distinctive to each side (weighted log-odds). The two vocabularies cover the same ground. What differs is stance. The machine gets a task register: state the thing, get the thing. The human gets a relational one, casual and personal.

Say it two ways

The sturdiest proof holds the topic fixed. Here is the same subject, how they look for what they want, put to the machine and then to the human. When the topic cannot vary, the stance still flips.

Both are speech, transcribed. Because the subject is held fixed across the pair, the register difference here cannot be explained away by different topics or by a chattier transcript, the way a raw word count can.

Their style tracks the human's more than the machine's

People's word-style converges toward the human. Toward the machine, less.

Language Style Matching measures how closely two speakers use the little function words, the pronouns, articles, prepositions, that carry style rather than content. The higher the match, the more the two have fallen into step. We computed it for every interview, participant against moderator.

Why the machine got the unperformed version

We don't have to guess why. Participants say it themselves, unprompted, when the interview asks who they pictured on the other side. There is no one there to perform for.

You kinda just felt like a neutral third party. I didn’t feel, like, embarrassed or shy to talk to you about it ’cause you’re an AI.P014 · man · 25–34 · to AI (George)

Considering the interviewer is AI, it technically is not any of those things. It’s not human.P039 · woman · 35–44 · to AI (George)

This is the old computer-disinhibition effect, live. Strip the person, and people drop the impression management that a human conversation asks for. Mostly that reads as efficiency. For a few, it goes all the way.

Some go all the way

The blunt, anatomical register is a facet, not the headline. This is about intensity, not incidence: only about one interview in five ever goes explicit, and that share is roughly the same whichever moderator is listening. But when a participant does drop the euphemism, it tends to happen with the machine, and it's striking.

There’d be sucking and fucking and just anal and just anything you needed.P098 · man · 35–44 · to AI (Mia)

Does gender change how people talk?

The moderator's gender: no. The participant's gender: not how bluntly, but what about.

The study's other headline was gender, so we ran it through the language lens too. The machine's presented gender leaves register untouched: across the three AI personas the numbers barely move (moderator-to-participant ratio 1.55 to 1.62, hedging 18 to 20 per 1,000 words, intervals overlapping). Gender is as decorative for how people talk as it was for what they said.

Participant gender is a different story, but not where you would expect. Men and women do not reliably differ in how blunt or how hedged they are; those gaps sit inside the noise. They differ in what they reach for. Men talk in content and logistics. Women talk in feeling and representation.

With a person, there is a face to protect

Porn is a face-threatening topic. Talking to the human, people cushion it, hedges, softeners, permission-giving, euphemism; hedging and curation run 28.4 per 1,000 words with the human against 19.1 with the AI. Talking to the machine, there is less face in the room, so less of that work gets done. It is most visible when a participant turns the care on the moderator herself: in the last minute of one interview, they stop answering and start asking.

A participant fished; the human gave something back; the interview became a conversation between two people. That reciprocity is structurally impossible with an interviewer the participant has already classified as "not human." It's not a leading question. It's the thing a person can do that a query cannot.

What they were all actually saying

Under both registers, the content rhymed. Person after person described porn that isn't quite for them, and a quiet accommodation to it. A word kept surfacing.

I don’t always find what I’m looking for. I settle.P054 · to AI

A queer woman settling for the straight stuff. A man calling the bisexual category the stepchild. The machine and the human both heard it. They just heard it in different words.