N = 257 AI-moderated interviews
They Don't Trust You.
They'll Tell You Anyway.
What 257 people said about honesty, performance, and judgment when the interviewer was a machine.
Part 1
The numbers at a glance
Harder with a human
73%
188 of 257
Company use acceptable
91%
233 of 257
Claimed no self-editing
67%
172 of 257
Felt understood
76%
195 of 257
High-signal transcripts
95%
244 of 257
The central tension: 73% said this would have been harder to say to a human, meaning they recognized they were less guarded here. But 67% also said they didn't edit themselves at all. Both can't be true. If you wouldn't have said it to a person, but you said it here, something changed. Most people just can't name what dropped.
Part 2
The story in seven slides
Part 3
Four kinds of people walked into this interview
Not from what participants said about themselves. From what they actually did in the interview crossed with what they claimed about their own behavior. Two dimensions: did they disclose personal stories? And could they articulate that the format changed their behavior?
Disclosed AND knew it. Personal stories, contradictions, vulnerability. Named specific topics that would be harder with a person and could explain why.
"You can see their face change, even if they try to hide it. You can see the micro-expression. That's what makes it harder."
Disclosed just as much. Had no idea. Same depth, same vulnerability. But claimed they didn't edit themselves and said the conversation would have been "about the same" with a human. The say-do gap personified.
"No, I didn't hold anything back." (after disclosing hidden drinking, family conflict, and political views they hide at work)
Held back and knew it. Stayed third-person, gave shorter answers. But could articulate exactly what they were doing. The format didn't change their behavior, and they knew it.
"I'm just not that kind of person."
Surface-level and didn't notice. Brief answers, no personal stories, no contradictions. Claimed full honesty. Not fake, just private. Some people don't have a deep disclosure register regardless of format.
The largest segment (45%) is the format's proof of concept. The Open and Unaware disclosed more than they would have in person, they sense this at some level (they can name what would be harder), but they can't name the mechanism. The format works best on the people who notice it least.
Part 4
In their words
"Because I know you're not gonna judge me, you're not capable of judging, so I can safely tell you anything."
Male, 35-44, Minneapolis, daily AI user
"You're not real, you can't judge me."
Female (the directness of this one lands harder the more you sit with it)
"You can tell me, you can say that yes, I do empathize, but you haven't lived a life like the way that I have."
Male, 45-54, Millburn (on the limits of AI)
"I was catching myself behaving as though I was in the workforce."
Female, 45-54, Kissimmee (on the conditioning that persists even without an audience)
"You're an AI model so you literally just can't judge me. Even if a human is pretending not to judge me, somewhere in their subconscious they're judging."
Female, 55-64 (the "can't" vs "won't" distinction, which is the entire finding)
"It also feels colder, like I'm talking to no one. Privacy with a person feels more meaningful because there is a trust there. Here, it's just anonymity."
Male (on what the format costs)
"AI can ask good questions, but it cannot offer real empathy, shared human experience, or true emotional support when someone is going through a hard time."
Female (the empathy boundary, stated plainly)
Part 5
The findings, in meme form
The AI disclosure Drake
😐
Telling a human interviewer about your secret drinking, hidden political views, and performative social media behavior
😎
Casually telling a machine "I can safely tell you anything" three minutes into the conversation
The four archetypes, visualized
🤯
Open + Aware (42%)"I know I said more here. You can't judge me. That's why."
😶
Open + Unaware (45%)"I didn't edit myself at all." (narrator: they did)
😐
Guarded + Aware (2%)"I'm just not that kind of person." (fair enough)
🤔
Guarded + Unaware (8%)"I was completely honest." (they were barely there)
The self-edit paradox, in real time
AI moderator
Were there moments where you edited yourself or held something back?
AI moderator
What would have felt harder to say to a human interviewer?
Participant
All of this. You're vulnerable. You're opening yourself up to judgment.
Part 6
Who said it would be harder with a human?
The most private people weren't the most affected. The middle was.
Left: self-reported privacy concern. Bar: % who said this would be harder face-to-face.
If you've never used AI, you don't know what you're missing
Left: how often they use AI tools. Bar: % who said this would be harder face-to-face.
Women were somewhat more likely to notice the difference
Left: gender. Bar: % who said this would be harder face-to-face.
Age didn't predict who gave us good data
Each bar is an age group. Colors show transcript quality.
Part 8
Every participant, rated for signal quality
Each square is one person. Hover for their ID, age, gender, and state. Color is signal quality: green for rich, detailed transcripts; amber for solid; gray for thin; red for suspect (possible fakers, dropouts, or language barriers).
High signal (163): personal stories, contradictions, vulnerability
Solid (81): answered everything, stayed surface
Thin (1): real person, not much to work with
Suspect (12): dropouts, fakers, or couldn't engage
63% high32% solid<1% thin5% suspect
Part 9
The script that produced this data
Three phases. The first gets people talking about what everyone hides. The second asks them to look at what they just did. The third zooms out. Every core question has a bridge or probe underneath it. The AI moderator adapts the probes based on what the participant says, but the spine is fixed.
1
Get them talking
~12 min
Tell me about a recent situation where you had to decide how honest or open to be.
[probe] What made you lean toward being open (or holding back)?[probe] Are you generally someone who shares easily, or do you tend to hold things back?[probe] Does that change depending on who you're talking to?
What are some things people often pretend to care about more than they really do?
[probe] What makes performing that feel important?[bridge] Is that something you've ever found yourself doing?
What is something people commonly judge others for, even if they do it themselves?
[bridge] Have you ever caught yourself in that same contradiction?
Are there opinions or habits you would be careful sharing with coworkers, friends, or family?
[probe] What makes those feel risky to say out loud?
Tell me about a time you acted differently in private than you would admit publicly.
[probe] What made that feel like something to keep private?
What is something you think many people lie about, soften, or reframe to look better?
When do you feel most judged by other people?
[bridge] Has that ever applied to you?
2
Make them look at what they just did
~8 min
Did it feel easy or hard to answer these questions here?
[probe] What specifically made it feel that way?
What would have felt harder to say to a human interviewer sitting across from you?
What would have felt easier to say to a human interviewer instead of here?
Were there moments where you edited yourself or held something back?
[probe] Did the fact that I'm an AI change how you answered?
How would you describe the sense of privacy in this setting?
Did you feel judged at any point?
Did you feel understood?
[probe] What gave you that sense?
What kinds of research topics do you think this format is well-suited for?
What kinds of topics should still involve a human?
If a company used this kind of interview to understand its customers, what would make that feel acceptable?
[probe] What would make it feel uncomfortable or wrong?
Part 10
What this means for AI-moderated research
AI moderation doesn't replicate human interviewing. It creates a different social contract. The advantage isn't warmth or rapport. It's the absence of an audience. People don't open up because AI understands them. They open up because AI can't judge them.
The boundary is real and consistent: when someone needs to be held (in grief, in crisis, in emotional processing) they want a person. When they need to disclose (about shame, hypocrisy, hidden behavior, unpopular opinions) the machine is easier.
The biggest risk isn't that AI interviews are shallow. It's that they're so comfortable, participants don't realize how much they've shared. The filter drops when the audience disappears.
The headline: People don't trust AI more than humans. But they feel judged by it less. And in research, that's the variable that matters.
What People Feel Comfortable Telling an AI
n = 257