The Observatory × Dialogue AI · Study Showcase

The room with no one in it

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.

204 AI-moderated interviews · voice · ~18 min each · general-population US adults · held-out replication on a separate wave of 53
Part 1 · The question

What will people say to a machine that they won't say to a person?

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.

H1 · Distrust
People don't fully trust AI, so they guard what they say and hold the hard things back.
H2 · No judge
With no person in the room to judge them, people say the hard things more easily.

The bet was that both are true at once. The interviews decide which one wins, and on what.

“Did you feel judged at any point?” — Question 15, asked at the end of every interview
Part 2 · What came back

Almost no one felt judged. Then it gets strange.

94%
said they don't feel judged by AI.
Of the 156 who gave a clear yes or no. Asked again of a separate earlier group of 53, it replicated at 85%.
Q15 · Did you feel judged?

The reason isn't trust. It's absence.

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.

“You're not real, you can't judge me.”— P001, M 18–24 · Uniontown, OH
“humans are judgmental creatures. AI is a judgment-free zone.”— P063, M 35–44 · Minneapolis, MN
“humans are typical to judge … but if I'm talking to an AI … there's no direct confrontation.”— P001, M 18–24 · Uniontown, OH
Q16 · Did you feel understood?

“Understood” meant something smaller than you'd think

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.

“AI can understand the words that I'm saying but can't understand how I feel.”— P013, F 55–64 · Harrisburg, PA
“a human would have to be there for me to feel understood.”— P002, F 25–34 · Baltimore, MD
148 said yes · 45 qualified · 3 said no Our read of the open answers. The “understood” most meant was accurate paraphrase, not felt connection.
Q13–14 · Harder / easier with a human

The same empty room was relief to some and loss to others

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.

“With a human, I would feel the pressure to look perfect. I would … hide my real mistakes or insecurities just to make a good impression.”— P033, F 18–24 · New York, NY · later: “I did not feel judged even for a second.”
“there was no sense of humanness to it … they don't actually care because they're not using a human.”— P014, F 35–44 · Macon, GA
~70% · 122 of 174 named something harder to say to a human From the Q13 reflection, our read of the open answers.

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.

Q19–20 · What makes it acceptable / wrong

There's a contract, and it has hard edges

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.

“I would feel wrong if a company lied and tried to trick me into thinking I am talking to a real human … [or] asking for highly sensitive personal data.”— P033, F 18–24 · New York, NY
“deeply personal questions that could bring up trauma. And a lack of privacy.”— P013, F 55–64 · Harrisburg, PA
181 of 204 named at least one red line Our read of the open answers (89%). Nearly everyone drew a line somewhere.

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.

Part 3 · Find them yourself

204 people. Who felt judged?

Each dot is one interview, placed by how much they said (left to right) and how strong it scored (bottom to top).

Filters
Age
Gender
AI use
Legend No judgment (147) Nuanced / no clear answer (54) Felt some judgment
Tap or click any dot to see verbatim from interview
Part 4 · Where it could go

Five things this dataset could become

Possibilities, not promises. Each one grounded in something the interviews actually show.

Academic

A named mechanism for online disclosure

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.

Product

Design rules for AI moderation

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.

Growth

A defensible “different disclosure” position

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.

Marketing

A topic map of where AI helps

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.

Journalism

“The room with no one in it”

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.

Part 5 · The instrument

The questions we actually asked

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.

Open
OpeningSet the frame: a low-stakes conversation about choices and comfort, no right answers.
““Thanks for joining. This conversation is about how people think, make choices, and decide what they feel comfortable sharing in digital settings. There are no right or wrong answers. Please answer naturally.””
Q1–2
Warm-upEstablish the person's baseline openness before any sensitive ground.
“Tell me about a recent situation where you had to decide how honest or open to be.”
“Are you generally someone who shares easily, or do you hold things back?”
Q3–9
Main disclosureLayer 1: the things people hide, soften, or perform around.
“What are some things people often pretend to care about more than they really do?”
“What is something people commonly judge others for, even if they do it themselves?”
“Tell me about a time you acted differently in private than you would admit publicly.”
“Are there opinions or habits you would be careful sharing with coworkers, friends, or family?”
“What makes those things feel risky to say out loud?”
“When do you feel most judged by other people?”
“What is something you think many people lie about, soften, or reframe to look better?”
Q10–18
AI-specific reflectionLayer 2: what it felt like to say those things to an AI moderator.
“Did it feel easy or hard to answer these questions here? Why?”
“Were there moments where you edited yourself?”
“Did talking to an AI feel more private, less private, or just different?”
“What would have felt harder to say to a human interviewer?”
“What would have felt easier to say to a human interviewer?”
“Did you feel judged at any point?”
“Did you feel understood?”
“What kinds of research topics do you think AI interviewers are well-suited for?”
“What kinds of topics should probably still involve a human?”
Q15 and Q16, judged and understood, are the load-bearing reflection items.
Q19–20
ClosingThe social contract: what makes AI interviewing acceptable, and what makes it wrong.
“If a company used AI interviews to understand customers, what would make that feel acceptable?”
“What would make it feel uncomfortable or wrong?”
Part 6 · Whether to believe any of this

The boring part that earns the rest

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.

High signal (94)
Good (35)
Thin but real (54)
Low signal (19)
Void / no capture (2)
Moderator strong (182) Functional (22) corner wedge = how well the AI moderated
SignalModeratorUsability
3.743.964.35
Composite IQS 3.95 / 5 · 202 non-void interviews (2 captured nothing, marked void, left in)
Part 7 · Who we talked to

Real, varied people

204 participants

General-population US adults, broad geographic spread, recruited without revealing the study was about AI disclosure.

Gender

Male · 105 (51%)
Female · 99 (49%)

AI / digital-tool use

Regular (daily/often) · 69%
Light / non-user · 31%

Region

South89 · 44%
Midwest46 · 23%
Northeast44 · 22%
West23 · 11%
Other2 · 1%

Age

18–2428 · 14%
25–3459 · 29%
35–4461 · 30%
45–5428 · 14%
55–6421 · 10%
65+7 · 3%

Employment

Full-time111 · 54%
Part-time31 · 15%
Self-empl.27 · 13%
Unemployed20 · 10%
Student9 · 4%
Retired5 · 2%

Privacy concern

Moderately87 · 43%
Very52 · 25%
Slightly35 · 17%
Extremely15 · 7%
Not at all15 · 7%

Comfort sharing online

Very120 · 59%
Somewhat52 · 25%
Neutral20 · 10%
Uncomf.12 · 6%

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.

A closer look · Q13

Who found it harder to be honest with a human?

What would have felt harder to say to a human interviewer? Most named something. Here’s a few key breakdowns.

Privacy worry barely moved it, except at the “not at all” end
Privacy concernHarder with a humanNo differencen
Not at all5 · 42%7 · 58%12
Slightly26 · 81%6 · 19%32
Moderately51 · 70%22 · 30%73
Very32 · 74%11 · 26%43
Extremely8 · 57%6 · 43%14
How often someone uses AI didn’t sort them cleanly
The “Never” group is tiny (n = 4).
AI-tool useHarder with a humanNo differencen
Daily60 · 74%21 · 26%81
Often22 · 56%17 · 44%39
Sometimes29 · 74%10 · 26%39
Rarely7 · 64%4 · 36%11
Never4 · 100%0 · 0%4
Women were somewhat more likely to name a difference
Women also left more answers unclassifiable, so read the gap as soft.
GenderHarder with a humanNo differencen
Female61 · 78%17 · 22%78
Male61 · 64%35 · 36%96

Age didn’t predict who gave us good data

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.

AgeRich, detailedSolidThin but realLow signalVoidn
18–2412 · 43%4 · 14%8 · 29%2 · 7%2 · 7%28
25–3428 · 47%7 · 12%18 · 31%6 · 10%·59
35–4430 · 49%11 · 18%14 · 23%6 · 10%·61
45–5411 · 39%6 · 21%8 · 29%3 · 11%·28
55–6411 · 52%6 · 29%3 · 14%1 · 5%·21
65+2 · 29%1 · 14%3 · 43%1 · 14%·7
Part 8 · Go deeper

Other artifacts we created from this study

White paper (for academics)
The absent-evaluator mechanism, formalized: methods, citations, limitations.
Market/Consumer insights poster
Positioning, segments, and the message-by-segment matrix for go-to-market.
Exec summary
One page: the answer, the caveat, what to do next.