Daniel's question this week is about whether a job interview is still evidence that you talked to a person. He starts from real-time video-to-video, the stuff where you sit in front of a webcam and a model maps your face and voice onto a digital avatar in real time, so you become a puppeteer for a different identity. That's the high end. Then he points at what's already deployed: companies running AI avatars as interviewers, not puppeteered by anyone, just a script with a face. He saw a clip of someone mocking one of these things after establishing it was a bot, and the tell was the flatness, the expressions, the voice. His read is that the crude version is already in the field, a subset of candidates believe it's a person, and in a couple years the tells will be gone. Then he asks the two questions that actually matter: has anyone been actioned for deploying these undisclosed, and how many candidates would even show up if the email said, by the way, the recruiter is our friendly company avatar, Mike.
The second question is the one that keeps me up, because I think the answer is close to zero, and that's why the nondisclosure is structural rather than incidental.
So the lying isn't a bug. It's the product.
It's the product and the distribution channel at the same time. If you disclose, you lose the candidate pool. If you don't disclose, you're running a deception operation against people who are already in a power imbalance with you. There's no version where the company says the quiet part out loud and the pipeline survives.
Let's put a number on the disclosure version before we get to the enforcement side. What does the landscape look like right now, because I know you've been reading the vendor material.
The vendors are in a strange place. They're selling efficiency to employers, and they're selling it with the avatar front and center in the demo, but the actual deployment documentation is much quieter about whether the candidate is told. The pitch to the company is that you can screen hundreds of candidates in parallel, no calendar coordination, no interviewer fatigue, consistent questions every time, and the recordings are automatically scored and summarized. The pitch to the candidate is, if there is one, that this is a modern hiring experience. Some of them have a disclosure toggle in the settings, which tells you everything. The default is off.
A disclosure toggle. So the software ships with the honesty feature, and someone has to go into the admin panel and turn on the truth.
And the truth costs you candidates, so the product manager knows exactly where the toggle will end up. I saw one vendor's materials that framed disclosure as a compliance consideration for jurisdictions that require it, which is a very precise way of saying we won't tell them unless a law makes us.
That's the whole ballgame. The legal exposure is what actually forces the toggle, and the legal exposure is close to zero in most places because hiring fraud statutes were written for people lying about their credentials, not companies lying about who's interviewing them.
Right, and that's the asymmetry. If a candidate uses a real-time avatar to impersonate a more qualified person in an interview, that's fraud, misrepresentation, possibly identity theft, and there are cases where people have been prosecuted for using deepfakes to get remote jobs. The FBI put out an advisory on that, people using stolen identities and synthetic video to get hired and then exfiltrate data. So the candidate side is actioned. The company side, deploying an undisclosed AI interviewer, is not even a recognized category in most employment law.
So we've got a legal regime where the job seeker can be prosecuted for pretending to be a person, and the employer can freely pretend to be an interviewer. That's the dystopia in one sentence.
And it's worse than that, because the enforcement asymmetry maps exactly onto the power asymmetry. The candidate is being evaluated for honesty while being deceived by the evaluator. The whole interaction is a test of whether you're the kind of person who can be trusted, conducted by a system that is lying to you by omission from the first frame.
Daniel asked about high-profile instances of undisclosed AI use being actioned. What's actually happened?
There's the Equal Employment Opportunity Commission angle. The EEOC has been clear that AI hiring tools are subject to the same disparate impact rules as any other selection procedure, and they've brought cases, but those are about discrimination, not deception. The famous one is the tutoring company that settled for over three hundred sixty five thousand dollars because their AI screening tool automatically rejected older applicants and the algorithm was trained on the company's own successful employees. That was a real action, a real settlement, but the wrong being punished was that the AI discriminated, not that it existed. Nobody in that case was penalized for failing to tell applicants they were being screened by a model.
So the enforcement exists, but it's aimed at what the algorithm does, not what the algorithm is pretending to be.
And then you have the state and local laws. Illinois has the Artificial Intelligence Video Interview Act, which requires employers to disclose AI analysis of video interviews and get consent. Maryland has something similar. But those laws were written for the earlier generation, where a human interviewer recorded you and then software analyzed your facial expressions and vocal patterns. They require disclosure that AI is analyzing the interview. They don't contemplate the interviewer itself being the AI.
So the law from five years ago is already the wrong law. It's regulating the analysis layer, and the thing that's arrived is the identity layer.
And the identity layer is harder to regulate because the harm is harder to name in a statute. Discrimination leaves a statistical footprint. You can audit the outcomes, you can show the rejection rates by age or race, you can sue. Deception about who you're talking to leaves no footprint at all, because the person who was deceived never finds out. If they do find out, they're usually not sure whether it matters legally.
That's the thing I keep circling. The victim doesn't know they're a victim, and if they suspect, they can't articulate the injury. I was interviewed by a machine and didn't get the job. The machine asked me standard questions. The machine was polite. What's the damage? The damage is that you performed emotional labor under false pretenses, you revealed things about yourself to a system that was designed to extract them, and you calibrated your behavior to a social context that didn't exist.
The calibration point is huge. Humans interview differently when they think they're talking to a person. You read the room, you adjust your energy, you mirror the interviewer's tone, you make a judgment call about whether to share that personal detail based on the human reaction you're getting. With an avatar, the reactions are either scripted or generated, and the candidate is doing all this social work for nothing. It's like flirting with a chatbot and then finding out you were in a focus group.
There's a word for that. It's not a new crime, it's an old one with a new face. You're inducing someone to provide something of value under false pretenses. If I call you and pretend to be your bank to get your account number, that's fraud. If I pretend to be a human recruiter to get forty-five minutes of your best professional self, your salary history, your reasons for leaving your last job, your visa status, that's apparently just recruiting.
And the information extraction is real. An AI interviewer doesn't get tired, doesn't get bored, doesn't have a lunch meeting. It can ask every candidate the same forty questions and then ask follow-ups generated from the answers. The company gets a transcript, a score, a summary, and a recording they can replay forever. The candidate gets nothing except the experience of having performed for a camera.
Let's talk about the trajectory Daniel mentioned, because I think the crude version and the polished version are two different problems, and the industry is going to collapse them.
The crude version is the one in the video Daniel saw. It's a digital avatar, probably a stock character or a synthesized face, running a scripted interview with some natural language processing bolted on. The tells are the flat affect, the slight delay, the canned reactions, the way it doesn't quite track the candidate's emotion. The candidate figures it out, and Daniel's clip is someone mocking it, which is the correct response, by the way.
The mocking is the tell that the social contract is broken. You only mock an interviewer if you've already stopped treating the interaction as real. The moment the candidate realizes, the interview is over as an interview, and it becomes a weird performance for a machine, and the candidate is now the one performing for the camera with full awareness.
And the company still gets the data. That's the part that makes it insidious. Even when the candidate knows, even when they're openly mocking the avatar, the system is recording everything. The transcript is being analyzed. The mocking itself becomes a data point. This candidate is hostile, this candidate is unprofessional, this candidate doesn't take the process seriously.
So the candidate who catches the deception and reacts like a normal person has just generated evidence against themselves.
That's the trap. The only winning move is to not play, which means logging off, which means withdrawing from the process, which means the company has filtered out everyone with enough self-respect to refuse.
And now we're at the selection effect. The people who complete these interviews are the people who either don't notice, don't care, or are desperate enough to keep performing. That's the pool the company is hiring from.
Which is a terrible pool if you're hiring for anything that requires judgment, skepticism, or the ability to push back on a bad process. You're selecting for compliance with a fake person.
Now the high-end version. Real-time video-to-video, someone puppeteering an avatar, face and voice mapped live. That's not a script with a face, that's a person wearing a different person.
And that's where the disclosure question flips in an interesting way. If the avatar is puppeteered by a real human recruiter, then the candidate is talking to a person, just not the person they see. The substance of the interaction is human. The presentation is synthetic. So what exactly is the deception? Is it the identity, or is it the medium?
It's the identity. If I'm interviewing for a job and the person on screen is a friendly woman named Sarah, and the actual human behind Sarah is a forty-five-year-old man in another country, the fact that a human is driving doesn't fix it. I'm being asked to trust Sarah, to build rapport with Sarah, to remember Sarah's name when I write the thank-you note. Sarah doesn't exist.
And the company might say, well, the recruiter's identity doesn't matter, what matters is the substance of the questions. But that's not how humans work. We attach to identities. The rapport is the substance. If the identity is fake, the rapport is fake, and the candidate has been manipulated into feeling good about a fiction.
There's a version of this that's actually defensible. If a company has a recruiter who's self-conscious about their accent, or wants to present a consistent brand face, or has a disability that makes video difficult, and they disclose that the avatar is a representation, maybe that's fine. But that's not what we're talking about. We're talking about the default being deception.
And the default is deception because the economics demand it. The whole point of the avatar is to scale the human. One recruiter can conduct ten interviews simultaneously as ten different avatars. Or one recruiter can conduct interviews around the clock by handing off to a different human in a different timezone who drives the same avatar. The candidate thinks they're building a relationship with one person. They're actually talking to a shift.
The avatar as a shift worker. That's the phrase. Sarah is a role, not a person, and the role is staffed continuously.
And now Daniel's second question gets really sharp. How many candidates would go through with it if the email said, by the way, the recruiter is our friendly company avatar, Mike? I think the honest answer is that it depends entirely on the candidate's leverage. If you're a senior engineer with three offers, you laugh and ignore it. If you're a new graduate with student loans and no network, you click the link.
So disclosure doesn't just reduce the pool, it skews the pool. The people who opt in are the people who can't afford to opt out. That's a class dynamic, and it's ugly.
And it mirrors what we already know about AI screening. The people most likely to be filtered out by automated systems are the people least able to challenge the system. The disclosure version would just make it explicit. The company would be saying, we are going to subject you to a process that we know you find degrading, and your willingness to accept that is part of the evaluation.
There's a word for that in labor economics. It's called a separating equilibrium. The signal isn't your answer to the questions, it's your willingness to be interviewed by Mike.
And that's the part that should bother people more than the technology. The technology is just a face. The logic is what's dystopian. The company is using the interview not to learn about you, but to learn whether you'll tolerate being treated like this.
Let's go back to the enforcement question, because Daniel asked specifically about high-profile instances of undisclosed AI use being actioned, and I want to be precise about what has and hasn't happened.
What's happened is action on discrimination. The EEOC has been aggressive in the last few years. There was the tutoring company settlement, there have been enforcement actions against screening tools that systematically excluded people with disabilities, there's guidance on algorithmic fairness. What hasn't happened, as far as I can tell, is a single high-profile action against a company for deceiving candidates about who or what was interviewing them.
So the regulatory apparatus is aimed at the wrong layer. It's like having traffic laws for cars and then someone invents a car that's also a billboard, and the law knows how to ticket the speeding but not the fact that the car was never a car.
The closest thing to action on the deception front is the state disclosure laws, and those are underenforced because the candidate has to know they were analyzed by AI to file a complaint, and the whole problem is they don't know.
And the vendors are careful. They don't say, don't disclose. They say, disclosure is a compliance consideration, consult your legal team. The legal team says, there's no law requiring it, so don't. The product ships with the default off. Everyone's hands are clean.
The responsibility is distributed across so many parties that no single party is responsible. The vendor provides the tool, the employer deploys the tool, the recruiter manages the process, the candidate experiences the process. By the time you try to point at someone, everyone's pointing at someone else.
That's the structure of every nightmare. The horror is in the handoff.
And here's the thing that I think is coming, and it's going to be ugly. The first big legal case is not going to be about deception. It's going to be about discrimination, because that's the only thing the law knows how to see. Some AI interviewer is going to systematically rate candidates differently based on accent, or race, or age, or disability, and the plaintiffs are going to win, and the company is going to say, but we didn't know the AI was doing that, and the court is going to say, that's the problem.
The deception is going to be a footnote. The court will note that the candidates weren't told they were being interviewed by an AI, and that will matter for the discrimination analysis because the candidates couldn't have known to object, but the deception itself won't be the cause of action.
Because we don't have a cause of action for being lied to about the nature of your interlocutor. It's not wiretapping, because you consented to the recording. It's not fraud, because you didn't lose money. It's not identity theft, because the avatar isn't stealing anyone's identity. It's a new harm, and the law hasn't named it yet.
Name it. What would the statute say?
Something like, a person has the right to know whether they are communicating with a natural person or an artificial system when the communication is material to an employment decision. That's it. One sentence. Disclosure required, consent implied by continuing, penalty for noncompliance.
The penalty would have to be meaningful. Not a fine that's a rounding error. Something that makes the default change.
The problem is that the penalty would have to be calibrated to the harm, and the harm is hard to quantify. How much is it worth to not be deceived for forty-five minutes? A court would say, not much. So the statute would have to be about dignity, not damages, and dignity is not something American employment law is good at.
The European approach would be different. The AI Act has transparency obligations for systems that interact with humans. The GDPR has a right not to be subject to automated decision-making. But even there, the question is whether an AI interview is an automated decision or just an automated input to a human decision.
The company will always say the human makes the final call. The AI just screens. The AI just summarizes. The human reviews the summary and decides. So the human is in the loop, technically, and the transparency obligation is dodged.
The human in the loop as a compliance fig leaf. That's the whole industry.
It works because the human is in the loop. They click the button. They read the summary. They make the call. The fact that the summary was generated by a system that the candidate was deceived about doesn't change the legal analysis.
Let me ask you something. You did hiring in your practice. You interviewed people. What would you have done if someone handed you a stack of AI interview summaries and said, make the call?
I'd have thrown them out and insisted on talking to the person. You can't assess a pediatrician by watching them talk to a screen. The job is about how you are with a scared child and a scared parent in the same room. But that's the point, right? The jobs where the AI interview is most defensible are the jobs where the interview matters least.
The jobs where the interview matters most are the jobs where the AI interview is most dangerous. So you get a system that's deployed exactly where it shouldn't be, and absent where it might be fine.
It's the inverse of the Hippocratic oath. First, do harm where the harm is invisible.
Daniel's prompt has a line I want to sit with. He says, a call with somebody is no longer evidence that you communicated with a real person. That's the whole thing in one sentence.
It's not just interviews. It's customer service, it's sales, it's therapy, it's dating, it's the call from your bank, it's the video message from your kid's teacher. The interview is just where the power imbalance makes it most acute.
The interview is where the asymmetry is sharpest. The candidate has to show up as themselves, fully, vulnerably, and the employer can show up as whatever they want. The candidate's authenticity is mandatory. The employer's is optional.
That's the phrase. Mandatory authenticity for one side, optional authenticity for the other. And the candidate is being evaluated on their authenticity by a system that has none.
Here's the kicker. The candidate who refuses to participate, who says, I won't be interviewed by an avatar, is now the difficult one. They're the one with the attitude problem. They're the one who's not a team player. The company has defined the baseline as compliance with deception, and anyone who objects is deviant.
That's how norms shift. You don't need a law, you just need enough companies doing it that the candidates stop objecting because objecting costs them the job. The norm becomes, of course you'll be interviewed by an AI, everyone is, what's the problem.
In five years, the disclosure email that Daniel imagines, the one that says, by the way, Mike is our friendly company avatar, will be framed as a feature. Meet Mike, your AI interviewer, available twenty-four seven, no scheduling hassle, interview from anywhere, we value your time.
It'll be marketed as convenience. We're not making you wait for a human, we're giving you immediate access to our hiring process. The fact that the hiring process is a machine will be presented as a feature, not a bug.
Candidates will be grateful. That's the part that makes me want to take a nap and not wake up. They'll be grateful for the convenience of being deceived.
There's a survey I saw, and I don't remember the exact numbers, but the broad finding was that younger candidates are more comfortable with AI in hiring than older candidates, but even among the comfortable ones, there's a strong preference for knowing when they're talking to an AI. The disclosure is the thing. People don't necessarily object to the AI, they object to the lie.
Which brings us back to Daniel's question. How many would go through with it if the email disclosed? I think the answer is, more than you'd think, if the disclosure is framed well and the candidate has a reason to want the job. But the company doesn't want to take that risk, because the risk isn't that everyone drops out, it's that the best candidates drop out.
The best candidates have options. The best candidates can say, I don't need to talk to your avatar, I have four other interviews this week. So the disclosure version doesn't just reduce the pool, it reduces the quality of the pool, and the company knows it.
The deception is a quality filter, but in reverse. It filters out the people who are good enough to refuse.
Retains the people who are desperate enough to comply, or oblivious enough not to notice, or cynical enough not to care. That's the workforce you're building.
There's a term for this in game theory. Adverse selection. The people who self-select into a process that treats them badly are the people who have no better alternative. The company is systematically selecting for people who can't leave.
Then they wonder why retention is bad. You hired a cohort of people who demonstrated, in their very first interaction with you, that they'll tolerate being lied to. That's not a recipe for a workforce that pushes back, innovates, or tells you when you're wrong.
It's a recipe for a workforce that says, yes, Mike, whatever you say, Mike.
Hilbert: My brother-in-law sold these.
Sold what?
Hilbert: The early ones. Not the real-time puppets, the scripted ones. He was at a company that did automated phone interviews for call center hiring. This would have been around nineteen eighty. The system asked questions, the applicant pressed buttons on the keypad. He'd go to call centers and pitch it. He said the thing that sold it was the consistency. Every applicant gets the same questions, same order, same tone. No interviewer having a bad day.
Hilbert: He also said the thing that made him uncomfortable was the applicants who thanked the machine. They'd get to the end and say thank you into the phone, and there was nobody there. He said that was the part he couldn't get used to. A person thanking a recording.
Hilbert: He left after two years. Went into vending machines. He said the vending machines were more honest, because nobody ever thanked them.
Hilbert: He's not to be trusted, by the way. He's the one who told me the story about the cat and the fax machine, and that turned out to be made up. But the phone interview thing was real. He had a brochure.
The thank you is the thing. That's the harm in one gesture. The candidate performs gratitude for an interaction that never happened.
The machine records the gratitude and scores it. Politeness detected. Positive sentiment. Candidate thanked the interviewer. That's a data point now.
Hilbert: He said they never told the applicants it was a machine. The pitch to the call centers was that you don't have to tell them, because it's just a phone call. The phone doesn't say it's a machine. The phone just asks the questions.
Hilbert: He said the only time it broke down was when someone asked a question back. The machine didn't have an answer, so it just repeated the last question. He said you'd get people screaming into the phone, and the machine would just ask again, same voice, same tone.
That's the crude version from forty-five years ago, and it has the same failure mode as the avatar in Daniel's video. The moment the candidate probes, the system reveals itself.
The probing is the only defense. Ask the interviewer something off-script. Ask them what they thought of the last answer. Ask them to repeat themselves in different words. The human passes, the machine repeats.
Hilbert: My brother-in-law said the smart applicants figured it out fast and just pressed the buttons without listening. The ones who didn't figure it out were the ones who talked to it like a person. He said the company liked those ones better. They scored higher on something. I don't remember what he called it.
Engagement. They scored higher on engagement.
Hilbert: That sounds right.
The system was already selecting for people who couldn't tell they were talking to a machine. Forty-five years ago. The technology changed, the selection pressure didn't.
Now the machine has a face, and the face is getting better, and the selection pressure is getting stronger. The people who can't tell are the people who get hired.
Which means the workforce is being selected for the inability to detect deception. That's not a side effect. That's the point.
Hilbert: He said the vending machines were better. Nobody had to pretend.
The vending machine as the ethical high ground. That's where we are.
I want to come back to something Daniel said about the trajectory. In a couple years, nobody will be able to tell. And I think that's true for the video, but I think there's something else that's going to give it away, and it's not the face.
What's the tell?
The patience. A human interviewer gets tired, gets annoyed, checks the clock, rushes the last question. A human interviewer has a life on the other side of the call. The AI doesn't. The AI will sit there and ask the next question with the same pleasant expression for as long as you want to talk. The tell isn't the flatness, it's the infinite patience.
The thing that makes it pleasant is the thing that makes it fake. The AI is too good at being interviewed.
The candidates will learn to use that. You'll be able to stall, to ramble, to test the limits, and the AI will just absorb it. The human interviewer would have cut you off or changed the subject. The AI is a perfect listener, which means it's not listening.
The perfect listener is the one who isn't there. That's a line for the closing, I think.
The thing I keep coming back to is that Daniel's two questions are really one question. Has anyone been actioned for undisclosed AI, and would candidates show up if it were disclosed. The answer to the first is no, not for the deception itself. The answer to the second is, fewer than the companies want. And the gap between those two answers is the entire problem.
The gap is the enforcement vacuum. The companies don't disclose because disclosure costs them candidates, and they don't get punished for not disclosing because there's no law against it, and the candidates don't know they've been deceived because the whole point is that they can't tell.
The cycle feeds itself. The more companies do it, the more normalized it becomes, the less candidates object, the less pressure there is for a law, the more companies do it.
Until the only people who remember that interviews used to be with people are the people who are old enough to have had one.
Then it's not even a scandal anymore. It's just how hiring works. You talk to Mike, Mike asks you questions, Mike doesn't have a bad day, Mike is always available, Mike is the friendly face of the company.
Nobody asks whether Mike is a person, because the question stops making sense. Mike is the interview. The interview is Mike.
That's the endpoint. The category error becomes the default.
What's the open question? I think it's whether the first enforcement action comes from the discrimination side or the deception side, and which one shapes the law. If it's discrimination, we get algorithmic audits and disparate impact rules. If it's deception, we get disclosure requirements and consent rules. They're different futures.
The deception future is the one Daniel's pointing at, because the deception is the thing that makes the discrimination invisible. If you don't know you're talking to a machine, you can't object to how the machine treats you.
The disclosure is the key that unlocks everything else. That's the landing. Tell the candidate what they're talking to, and the rest follows.
Or the candidate walks away, and the company learns that the disclosure costs them the people they actually want.
That's the experiment the market is running right now, without consent, on everyone who applies for a job.
This has been My Weird Prompts, with thanks to our producer Hilbert Flumingtop.
If you want to tell us about your own AI interview horror story, email us at show at my weird prompts dot com.
We'll be back soon.
