The Unedited
Register

the question

What do people call it when nobody's listening?

Ask someone about pornography with a person in the room and you get a curated answer. "Content I enjoy." "Stuff I like." The edit happens before the sentence leaves their mouth, and it happens so fast that neither of you notices.

Study 01 found something odd. When an AI moderator asked the same questions a human researcher had asked, it reproduced the same themes (that study is reported separately). But some people answered in a register they had not used with the human. Blunter. More specific. Words you would not say across a table.

Something worth saying before any of the findings. Eighty-nine people agreed to describe, in detail, what they look for in pornography and why. Most did it with more thought and less embarrassment than the topic gets in almost any other setting. Forty-one of them laughed at some point. Two apologised for language nobody had asked them to apologise for. One, who works at an adult store, pulled out her phone mid-interview to check how the site actually behaved, for science. Whatever the numbers below say, they were generous, and the study exists because they were.

That was an accident. Study 02 was built to do it on purpose: 89 people, twenty-three product questions about how they find and choose and reject adult content. Nothing personal. Nothing invasive. Just: what do you actually call this? We aimed for an even split of men and women and landed at four in ten women, which matters for one of the findings below, and we come back to it at the bottom.

There is thirty years of research saying a computer gets franker answers than a person does. It is good research and we believed it. It is also about whether people admit to a thing, which turns out to be a different question from how they talk about a thing they have already agreed to discuss.

We expected the AI to be the reason people spoke freely. That is not what the data says.

finding 01

The vocabulary doubled. The moderator had nothing to do with it.

Study 01 ran the same guide two ways: once with a human moderator, once with an AI. Study 02 ran a guide rewritten specifically to make room for people's own words, with the same AI. Three corpora, one measure: how many distinct explicit terms people used.

A word on what gets counted, because it is the part we got wrong first. We counted against a published list of 403 explicit words that predates this study and that nobody involved in it wrote, applied whole, with nothing added or removed. Our first attempt used a list we built ourselves after reading the Study 02 transcripts, and it produced an effect nearly twice the size. A list you write after reading your own data finds what you were hoping for. The figures here are from the external one.

Swapping a human moderator for an AI moved the number from 34.0 to 32.8. The intervals overlap completely. Rewriting the guide moved it to 72.5, and its interval does not come near the others.

finding 02

And it changed kind, not just amount.

Counting terms only tells you how many. The more interesting question is what sort of words they are. Study 01's participants described pornography the way you would describe a shelf: content, videos, category, amateur, professional. Taxonomy. Study 02's participants named things that happen to bodies.

Same pattern. The moderator moved it half a point. The guide moved it nearly five times over.

finding 03

People don't warm up. They answer the question you asked.

There's a comfortable story in research about rapport: give someone twenty minutes, let them settle, and they'll open up. We tested it. Every interview was split at its own midpoint and each participant compared against themselves.

Nothing. 8.60 in the first half, 8.59 in the second. No warm-up, no thaw, no rapport curve.

What moves the register is which question is on the table. The block asking people to describe what they're looking for sits nearly four points above the opening block. The closing block, which asks them to reflect on the conversation, drops below it.

Twenty minutes of rapport bought us nothing we could measure. A well-built question bought us a great deal.

every interview, end to end

Eighty-nine conversations, drawn to scale

One ribbon per interview, ordered by how much of it the AI did the talking. Width is words: every segment is one turn, sized by how many words it ran. Warm is the moderator. Cool is the participant. Nothing is averaged, and nothing is smoothed.

the turn

So we asked people directly. And they told us the opposite.

Every interview ended with a plain question: would you have said this to a human researcher? Eighty-eight people answered, and we read every answer by hand, because a keyword search got it badly wrong (it found the effect in 12% of interviews; reading found it in 64%).

Two out of three people say the AI changed what they were willing to say. And nine of them say they would not have done the study at all if a person had been in the room.

So we have two findings that appear to contradict each other. The vocabulary says the moderator changed nothing. The people say the moderator changed everything. This is the point where a study either picks the answer it was hired to find, or goes and looks.

Both are true. The next section is about why.

finding 04 · the one we didn't expect

What people report doesn't predict what they say.

We split the sample by their own answer and compared how explicitly each group actually spoke. If "the AI freed me" meant freer words, the first group would use more of them.

The most explicit person in the study says the AI made no difference to him. A man who used not one explicit word says he would never have spoken to a human at all. Across the whole sample, the group who said the AI freed them averaged 5.50 explicit terms per thousand words. The group who said it changed nothing averaged 6.61. The gap runs the wrong way, and we can detect no relationship between the two at all (p=0.489). With 56 people in one group and 18 in the other, this cannot rule out a modest association. What it rules out is the straightforward one everybody assumes.

The register people report and the register you can measure are two different things wearing the same name. Feeling free to speak is not the same as using different words, and this study can measure both and show they don't move together.

finding 05

Women feel it far more. Their words are identical.

The study was designed around a question about gender: when the social filter comes off, do women's words differ from men's? We measured it three ways. Explicit vocabulary rate. How often people hedge a preference instead of stating it flat. How often they name a concrete act when asked what they want.

All three came back flat. Not small differences. No differences.

Women were 1.6 times more likely than men to say the AI changed what they told us. Men were more than six times likelier to say it made no difference. Both survive correction for multiple testing.

And their vocabulary is the same. 8.01 against 8.14 per thousand words. Women were, if anything, marginally more likely to use an explicit term at all (94.6% against 89.8%).

There is one place the gender difference does show up in behaviour rather than in report, and we nearly missed it by lumping two things together. The transcripts mark laughter two ways: a full laugh and a quieter chuckle. Counted together they looked like a modest female tilt. Counted apart they are two different phenomena. Women laugh six times as often as men (3.14 per interview against 0.51, and half of women laugh at least once against under a fifth of men). Chuckles show no reliable difference: 1.03 against 0.78, and thirty-five per cent of each group used one at least once. The laughs also sit in the middle of the conversation where the explicit vocabulary is, while the chuckles bunch at the opening and the close.

Worth keeping in proportion. Both markers are concentrated: three interviews account for half of all the laughing, and 59 of 87 people never laugh once. The picture is not a room full of laughing women. It is that when someone does laugh out loud while describing what they search for, she is usually a woman.

There's a second thing worth holding next to that. Five of the six people who flatly declined to answer a question were women. So women are both the group most likely to say the AI unlocked them and the group most likely to refuse a specific question. Those aren't in tension. Both are what it looks like when someone is actively deciding what to disclose, rather than simply talking.

Asked what she would have said instead, she declined: "I'm not gonna tell you because I have a certain decorum of modesty." P52 · @29:10

every interview

Eighty-nine people, one dot each

Every participant in the study, arranged by whatever you'd like to sort them by. Click any dot to hear that person. Some of these variables have enough people behind them to bin honestly. One of them doesn't, and it will tell you so rather than draw you a chart that lies.

what we heard but cannot count

Three findings that don't get a chart

Some of the sharpest material in this study sits in cells far too small to put a number on. The honest thing is to quote it and say so, rather than draw a bar that implies a precision we don't have.

Content made for someone else

Surfaced by 12 participants. The identity block was only reached by 23 of 89 interviews.

Read the order of that second sentence. The platform failed her, so she stopped searching. Demand that stops generating queries stops existing as far as any search log is concerned.

When the category is the harm

Three interviews reached this block. Three. Far below anything that could be called a rate, and carried here because a count of three does not make it untrue.

For most participants in this study the vocabulary problem is theirs: they know what they want and can't type it. Here it inverts. The words exist, they're printed on the site, and they're slurs.

What our participants were told they'd signed up for

Ten of the 89 raised this themselves, without being asked.

Participants were recruited through a research panel under the description "Study on language and content discovery preferences." It did not mention adult content. Most took the turn in their stride, and several said the surprise was part of what made it easy. But it is a real limitation and it cuts both ways: people arrived unprimed, which is good for measuring how they actually talk, and they arrived without having consented to this specific topic, which is a fair thing to hold against the study.

The people who said no

Six participants declined a question outright. Hand-verified from fifteen candidates; nine turned out to be people talking about skipping videos.

Six people out of 89 drew a line, and every one of them stayed in the interview afterwards. That is the off-ramp working. One of them volunteered, unprompted, that the moment the moderator let his refusal stand without pushing was the moment he relaxed into the rest of it.

the receipts

How good is this data?

Every interview was scored on three things: how much real signal the participant gave, how well the AI moderated, and whether the data came through clean. Weighted into one number out of five. One square per interview.

No interview was dropped. Two people turned out to have been interviewed twice, and their second sitting is left out of the per-person numbers.

The honest limit of the whole thing: we never ran this set of questions past a human moderator, so we cannot fully separate what the questions did from what the AI did. That is the next study, and it is a small one.

Who we talked to

We aimed for roughly 100 interviews at an even gender split and landed on 89 interviews, 87 unique people, 51 men and 37 women. Women are under-sampled against a question that was specifically about gender. That matters in one direction only: under-sampling makes it harder to detect a difference, not harder to fail to find one. The gender results here are nulls on measured language and a large surviving gap on reported experience, so the shortfall doesn't undermine either. It would matter a great deal if we were claiming a small positive difference. We aren't.

What we asked

Twenty-three questions in seven blocks. Every one about the platform and the content, none about the person. Median interview reached fifteen of them.