We asked 204 people to talk about the things they hide, soften, or perform around: money, shame, the opinions they keep from coworkers. Then we asked them what it felt like to say it to a machine. Almost no one felt judged. The interesting part is why.
People edit themselves. They soften the embarrassing part, manage the impression, give the version that survives a face looking back. We wanted to know what happens when the face is gone. So we ran 204 interviews with an AI moderator and, at the end of each, asked the person what it had felt like to talk to it: easier than a person, harder, or just different.
Going in, two forces pulled against each other.
The bet was that both are true at once. The interviews decide which one wins, and on what.
People didn't say the AI was kind, or safe, or on their side. They said it wasn't there. There was no one in the room to form an opinion. That's a different mechanism than trust, and that's the key finding, in a nutshell.
When people felt understood, they mostly meant the AI had accurately paraphrased them back. The words matched. For a vocal minority, that wasn't enough. Understanding required a person on the other side, and a paraphrase only made the absence louder.
This is the split the study refuses to fake a number on. For some, no face meant no pressure to perform. They could finally stop managing the impression. For others, no face meant no warmth, no reaction, no point. We heard both, clearly and often. We do not claim how many fall on each side.
We heard relief and we heard loss. The balance between them is left unsized on purpose. The honest move when the signal is real but the count isn't clean.
The comfort isn't unconditional. People drew bright lines: don't pretend to be human, don't fish for sensitive data, stay on the subject you said you'd ask about. Cross those and the room stops feeling safe.
What this lets us do: size the one thing that's clean (the 94% not-judged, with a confidence interval and a replication), and map the topics and stances around it without faking proportions. A floor on the pattern.
Each dot is one interview, placed by how much they said (left to right) and how strong it scored (bottom to top).
Possibilities, not promises. Each one grounded in something the interviews actually show.
The “absent evaluator”. Comfort that comes from no one being there to form an opinion is a cleaner account than “trust” or generic online disinhibition, and participants articulate it themselves. With a replication already in hand, it's a publishable contribution to the CSCW / HCI literature on self-disclosure.
Possibility: would need a human-moderated control arm to fully close.
The contract has explicit edges: never impersonate a human, don't extract sensitive data, keep paraphrase honest, hold steady on stigmatized topics so a cut-off doesn't read as a slight. Those are shippable guardrails for any AI-interview product.
Possibility: derived from stated boundaries, not from A/B tested designs.
Not “faster interviews” and not “understands you better” (which several participants actively reject). The honest, ownable claim is a judgment-free environment for the topics people manage their image around: money, shame, judging others.
Possibility: the comfortable, AI-friendly sample means this is a ceiling read, not a population estimate.
Participants name where the buffer pays off (sensitive money, bodily, shame, judging-others content) and where it costs (anything needing felt warmth). That's the raw material for a comfort map that tells buyers when to reach for AI moderation and when not to.
Possibility: the topic clusters are described, not sized.
A human-interest piece on what people confess to machines, and why the relief is so often about the absence of a face rather than the presence of a confidant. The verbatims carry it; the self-judgment turn gives it an ending.
Possibility: quotes are exact and attributed by P-number only.
The full discussion guide, verbatim. Two layers: things people hide, then a reflection on what it felt like to say them here. Tap to open each section.
Before any finding, the data itself. Every interview was scored on a quality index: how much real signal the person gave, how well the moderator did its job, how usable the transcript is. One square per interview.
| Signal | Moderator | Usability |
|---|---|---|
| 3.74 | 3.96 | 4.35 |
General-population US adults, broad geographic spread, recruited without revealing the study was about AI disclosure.
One honest caveat carried from the methods: this sample skews comfortable. 69% regular AI users, 84% comfortable sharing online, 72% research-panel veterans. The plan wanted a 50/50 split and explicit skeptics. So the 94% is best read as a ceiling on how judgment-free AI moderation feels. A floor on the pattern, not a population estimate.
What would have felt harder to say to a human interviewer? Most named something. Here’s a few key breakdowns.
| Privacy concern | Harder with a human | No difference | n |
|---|---|---|---|
| Not at all | 5 · 42% | 7 · 58% | 12 |
| Slightly | 26 · 81% | 6 · 19% | 32 |
| Moderately | 51 · 70% | 22 · 30% | 73 |
| Very | 32 · 74% | 11 · 26% | 43 |
| Extremely | 8 · 57% | 6 · 43% | 14 |
| AI-tool use | Harder with a human | No difference | n |
|---|---|---|---|
| Daily | 60 · 74% | 21 · 26% | 81 |
| Often | 22 · 56% | 17 · 44% | 39 |
| Sometimes | 29 · 74% | 10 · 26% | 39 |
| Rarely | 7 · 64% | 4 · 36% | 11 |
| Never | 4 · 100% | 0 · 0% | 4 |
| Gender | Harder with a human | No difference | n |
|---|---|---|---|
| Female | 61 · 78% | 17 · 22% | 78 |
| Male | 61 · 64% | 35 · 36% | 96 |
Rich, detailed transcripts show up in every age band; the share doesn’t fall off with age. Cells are count and share of that age group; color marks interview quality.
| Age | Rich, detailed | Solid | Thin but real | Low signal | Void | n |
|---|---|---|---|---|---|---|
| 18–24 | 12 · 43% | 4 · 14% | 8 · 29% | 2 · 7% | 2 · 7% | 28 |
| 25–34 | 28 · 47% | 7 · 12% | 18 · 31% | 6 · 10% | · | 59 |
| 35–44 | 30 · 49% | 11 · 18% | 14 · 23% | 6 · 10% | · | 61 |
| 45–54 | 11 · 39% | 6 · 21% | 8 · 29% | 3 · 11% | · | 28 |
| 55–64 | 11 · 52% | 6 · 29% | 3 · 14% | 1 · 5% | · | 21 |
| 65+ | 2 · 29% | 1 · 14% | 3 · 43% | 1 · 14% | · | 7 |