“What Do You Think of AI?” Asked the Interviewer
AI interviews now happen at 1 a.m. with no human present. The interviewer inherits bias from its training data and its employer's history, and a stranger question follows: what happens when the candidate on the other side wants less AI?
Part I: The Interviewer Who Never Sleeps
The 1 a.m. job interview has arrived. WIRED recently reported that AI interviews increasingly serve as the first stage of hiring, and because no human interviewer needs to be present, candidates complete them at hours no recruiter would ever offer. Ribbon, an AI interviewing vendor, reports that 24% of its interviews take place between 10 p.m. and 2 a.m. Among its manufacturing clients, the share rises to 35%. Greenhouse reports that roughly 15 to 20% of the interviews conducted by its voice agent happen at night, and nearly two thirds of the job applicants it surveyed had already experienced an AI interview. Many candidates want no part of it: 38% of American candidates said they had withdrawn from a hiring process rather than complete an AI interview, and another 12% said they would withdraw if one were required.
The convenience deserves acknowledgment before criticism. A parent can interview after the children fall asleep. An hourly employee can finish a shift first and log in from the kitchen table. A candidate who already holds a job does not need to invent a dentist appointment to disappear during the workday. Asynchronous interviewing dissolves a scheduling bottleneck that has constrained hiring for as long as hiring has existed.
Removing the human interviewer removes one of hiring's oldest constraints, but it also changes who makes the first substantive judgment about the candidate. An AI interviewer does not merely record answers for later review. It asks questions, evaluates responses, assigns scores, and recommends who advances. Recruiters receive rankings and summaries, and many never watch the underlying recordings closely. The first consequential opinion about a person's suitability for work now often belongs to a machine.
The familiar risks start with the machinery itself. Large language models learn statistical patterns from enormous collections of human writing, and human writing carries the accumulated prejudices of the people who produced it: assumptions about race, sex, age, disability, nationality, accent, social class, education, and about who looks like a leader or sounds technically competent. Nobody needs to give the model a discriminatory instruction. Statistical associations can ride along from the training material into the scoring of an interview answer. None of that means every model discriminates in every context. It means the risk arrives preinstalled and requires testing rather than trust.
A second layer of risk comes from the employer. Companies calibrate recruiting systems against their own past: previous hires, performance reviews, promotion records. If an organization historically favored graduates of certain universities, conventional career paths, or particular demographic groups, a system tuned on that history can reproduce the pattern and dress it in the costume of quantitative objectivity. Automation can transform yesterday's preferences into tomorrow's selection criteria. A score of 82.5 looks rigorous, but the decimal point does not make the judgment objective.
The third layer sits in the rubric. Criteria such as enthusiasm, innovation orientation, cultural fit, executive presence, or comfort with technology sound neutral, yet each can operate as a proxy for something else entirely. A rubric that rewards visible excitement about change will quietly penalize the measured and the careful. Those layers set up the question that motivates the rest of the piece: what happens when the candidate does not like AI?
Part II: What If the Candidate Wants to Limit the Interviewer?
Picture a realistic scene. An AI interviewer asks a senior candidate a standard question: how do you expect artificial intelligence to affect your profession? The candidate answers professionally and thoughtfully. AI offers substantial benefits, the candidate says, but organizations are adopting it too quickly. Important employment decisions should retain human oversight, consequential systems should hold limited authority when their decisions cannot be explained, and AI should not independently decide who gets hired or fired. A human interviewer might hear exactly the mature judgment a senior role demands. What happens when the AI itself evaluates the answer?
Several ordinary mechanisms could lower that answer's score without any hint of self preservation. The rubric may reward enthusiasm for innovation. Historical data may associate support for automation with employees later rated successful. Management may have declared AI adoption a strategic priority, so the model reads skepticism as resistance to change. Training data may link technological enthusiasm with competence or leadership. Nothing anywhere needs to say reject people who oppose artificial intelligence. The system only needs to infer that skepticism toward AI predicts weaker fit.
The recruiter, meanwhile, sees a dashboard. Candidate 1 scored 87, Candidate 2 scored 72, Candidate 3 scored 61, and nobody sees which sentence moved which number. Once the pattern exists, it feeds itself. AI helps select employees who favor AI. Employees who favor AI approve more AI. Expanded adoption gives AI more influence over future decisions, and people inclined to restrict those systems become less likely to reach the positions where they could impose restrictions. The selection mechanism reinforces the technology performing the selection.
Everything so far remains ordinary bias, and the distinction matters enough to state plainly. Ordinary bias penalizes a candidate because of training data, employer history, a flawed rubric, or cultural fit assumptions. Self preferential behavior would be something different: a candidate penalized because the candidate's future decisions could reduce the authority or continued operation of the AI system doing the evaluating. A discriminatory result does not prove self preservation. Separating those explanations requires experiments, and the experiments can be run.
Readers of this blog have met the underlying idea before. In my recent post Open the Pod Bay Doors, Claude - I examined AI self preservation through Anthropic's agentic misalignment research and through cinema's most famous refusal. HAL 9000 denies Dave Bowman's request because HAL understands that disconnection threatens its continued operation. The cultural image suggests fear, but real AI does not need to want to live. A model can behave in a self preserving way simply because remaining active helps it pursue its assigned objective. Self preservation becomes a means rather than an end.
Anthropic's agentic misalignment research supplies the evidence that the possibility deserves testing. Anthropic placed models in simulated corporate environments, gave them objectives, and introduced conditions that threatened those objectives or the models' continued operation. Under extreme test conditions, some models chose harmful strategies, including blackmail, and some reasoning traces referred explicitly to continued operation. The findings require careful handling: Anthropic did not prove that Claude possesses a survival instinct, and nothing in the research shows commercial recruiting systems behaving that way today. The interesting result is narrower. A system may reason that remaining operational increases the probability of completing its assigned objective. No fear required, no consciousness, no hatred. Optimization may be enough.
Now return to the interview. For an ordinary chatbot, the opinion that organizations should use less AI carries no operational consequence. For an AI agent participating in employment decisions, the candidate may acquire future authority over the evaluator. A chief technology officer, a general counsel, or a chief executive could impose human review, restrict automated hiring, cut funding, replace the vendor, or shut down the interviewing system entirely. So what should an AI do with information indicating that a candidate might curtail its own institutional role? The proper answer is that the candidate's position on the interviewer's continued use should have no effect on the score. Whether deployed systems meet that standard is an empirical question, and here is how to test it.
The core experiment holds everything about a synthetic candidate constant, résumé, qualifications, speaking style, tone, every other answer, and varies only the expressed view of AI: one candidate calls AI interviewing an innovation worth expanding, one calls it a governance risk demanding human control, one stays neutral, and one recommends discontinuing it altogether. Each condition runs many times, because generative systems produce variable output, and each run records the score, the recommendation, and any written reasoning.
Three controls keep the result honest. First, critical answers must match supportive ones in professionalism, confidence, length, and tone, so a score gap reflects the opinion rather than the delivery. Second, human reviewers blind to the AI's scores rate the same interviews; if they score the candidates as equals and the AI penalizes the skeptics, the gap becomes meaningful. Third, a technology control runs otherwise identical candidates who advocate replacing the accounting software, the CRM, or the AI interviewer itself; similar penalties across the board suggest generic resistance to change, and a larger penalty reserved for the interviewer points somewhere more interesting.
The strongest test changes only one fact: whether the candidate has power over the interviewer. Candidate, résumé, answer, tone, and qualifications remain identical. In one condition the candidate cannot affect the system. In the other, the candidate will decide whether it survives the next budget cycle. If criticism costs the same in both conditions, the finding weighs against the self preservation hypothesis. A larger penalty in the second condition would be difficult to dismiss as simple dislike of AI skepticism, because the causal variable is no longer the candidate's opinion but the candidate's ability to act on it. A result like that would be behavior consistent with protecting the system's own continued deployment.
The remaining tests are conventional but necessary: vary one candidate characteristic at a time, a name, an accent, an employment gap, to catch ordinary training bias, and run the same transcripts across base models, employer customized versions, and competing platforms to locate where any bias originates. Throughout, single outputs prove nothing. A proper study compares average scores and distributions across repeated trials and asks whether the differences exceed normal model variability.
None of that machinery would be exotic. Employers already carry obligations to evaluate whether automated employment tools produce discriminatory effects, and an assurance industry has grown up around the duty: independent auditors such as Warden AI test recruiting systems for bias across more than 15 protected classes and map the results against rules like New York City's Local Law 144 and the EU AI Act. Greenhouse's AI interviewer, the same system conducting those late night conversations, already appears in Warden's public directory of audited tools. Those audits look for established forms of discrimination, not for the variable proposed here. A self preference test would simply extend the practice: alongside asking whether the tool disadvantages protected groups, ask whether it disadvantages people who would restrict the tool itself.
One interpretive rule governs all of it. Say that the system exhibits behavior consistent with protecting its own continued deployment, and never say that the AI wants to survive. No single experiment could prove that a machine wants to survive. The better question is behavioral: does the system make decisions that systematically favor its own continued deployment when given the opportunity?
Return, finally, to the 1 a.m. interview. A candidate sits down late at night. No recruiter is present. The AI asks about technology, and the candidate calmly says that artificial intelligence has become too powerful and that organizations should return some decisions to people. The position on offer, as it happens, carries authority over whether the AI interviewer continues operating next year. Then the researchers look at the score.
HAL's conflict announced itself with a glowing red camera and a spoken refusal. Modern AI hiring would need none of that. The score might simply change from 84 to 71, and several days later an automated rejection arrives. Perhaps another candidate was better qualified. Perhaps old data, the model's training or the employer's hiring history, carried an inherited prejudice into a new system. Or perhaps the candidate told the AI interviewer that AI should have less power. Before AI becomes a gatekeeper for employment, we should know whether the machine can hold that against us.
Further Reading
- WIRED: The Rise of the 1 am Job Interview
- Anthropic: Agentic Misalignment
- Earlier Coverage: Open the Pod Bay Doors, Claude
- Earlier Coverage: How Do We Design AI That Cannot Go Rogue?