Everything that’s wrong with AI Interviews

Interviews are a human ritual, not a data harvest 

Interviews did not start life as a spreadsheet category. They started as a human moment. Two people in the same space, or at least the same conversation, trying to answer a very old question with very modern consequences: Do I trust you enough to build something with you. That trust is not generated by a score. It is assembled in real time, from tone, honesty, curiosity, and the subtle social physics of how a person shows up when it matters. Research on interviewing keeps circling back to a simple idea: interviews work best when they are structured, purposeful, and grounded in job relevant behaviours, not when they are treated as theatre or trivia. Plenty of the value is in the interaction itself, the clarifying questions, the follow ups, the moments where a candidate says something surprising and the room changes.  

A good interview is not just a company interrogating a candidate. It is also a candidate interviewing the company, even if nobody says that part out loud. When the conversation is real, candidates learn how decisions are made, how disagreement is handled, whether the team is thoughtful or frantic, and whether the workplace feels like a kitchen during dinner service or a library during finals. Those are not fluffy vibes. They are predictors of whether someone will last, grow, and be proud of what they are building. This is why people still insist on meeting the actual humans for roles where judgement, leadership, and accountability matter. In The United States and Canada, even public sector guidance on using automation in hiring leans heavily on transparency, explainability, and the hiring manager’s responsibility for the decision, precisely because hiring is not merely processing, it is consequential judgement.  

Now contrast that with the modern temptation: replace the messy humanity with a tidy machine experience. The pitch sounds soothing. No scheduling. No bias. No awkwardness. Just consistency. Yet what actually disappears first is not bias. It is context. The moment you remove the real conversation, you also remove the candidate’s ability to read the room and the company’s ability to understand how the candidate thinks when the script runs out. The interview is supposed to be the place where a résumé stops being a document and starts becoming a person. Turning that into a machine led monologue is like replacing a first date with a customer satisfaction survey. Technically, questions were asked. Socially, nothing happened.  

The business model problem: selling shortcuts as strategy 

A lot of “AI interview” products look like innovation until you squint. Then they look like a thin coat of paint on top of a general chat bot with a microphone. In the wider tech world, even investors and platform leaders have been publicly warning that entire categories of thin “wrapper” products are fragile, because they can be copied, under priced, or made obsolete when the underlying platform adds a feature. That is not a moral judgment. It is a durability problem. When the business is mostly a convenience layer, the incentive is not to build a careful hiring system. The incentive is to grow fast, sell faster, and let someone else deal with the wreckage, ideally in a different fiscal year.  

Hiring is an awkward place to bring that mentality. Talent is not a casual add on feature. It is the company. In fact, recent reporting and litigation around automated employment systems has made one thing painfully clear: you cannot outsource responsibility. Regulators and courts are increasingly treating algorithmic hiring as a serious risk area because it can scale harm quickly, invisibly, and with the plausible deniability of “the system did it.” Cases like Workday have highlighted just how contested and consequential these systems become when applicants allege that automated recommendations or screening are discriminatory, and when agencies argue that vendors can still be accountable for how their tools function in real hiring pipelines.  

The uncomfortable truth is that “we are using an AI interviewer” is often not a thoughtful strategy. It is a coping mechanism. Pipelines are flooded. Recruiters are exhausted. Managers want speed. Nobody wants to fund the boring work of training interviewers, maintaining structured scorecards, or building a real talent intelligence layer that cuts noise before the interview. So a bot shows up, wearing the costume of progress, and everyone pretends it is a plan. In this article we have listed down everything that is wrong with AI Interviews

The AI interview adds noise when hiring needs signal 

Here is the first big failure mode: the AI interview often creates more information than it removes. That is not help. That is paperwork with better marketing. 

Hiring already has an attention crisis. Eye tracking studies on résumé screening have shown that initial “fit or no fit” decisions can be made in seconds. The well known Ladders eye tracking research reported an average initial screen around 7.4 seconds. Think about that. We trained an entire generation of candidates to compress their adult lives into one or two pages because the reviewer is basically speed dating their work history.  

Government policy has been moving in the same direction for the same reason: reduce the cognitive load. A recent Office of Personnel Management guidance document explicitly supported a two page resume limit to make review easier and more consistent. 

Now add the typical AI interview workflow. A candidate records answers, the tool produces a transcript, then it generates a summary, then it adds sentiment observations, then it produces a rating, then it produces “insights.” If your hiring team did not have time to read a third page of résumé, why would they have the time to read five pages of bot commentary about a recorded monologue. This is how organisations accidentally build a second bottleneck right on top of the first one. Even research on AI interview platforms that take transparency seriously still talks about the need to reduce overwhelm and avoid meaningless precision. In a widely cited case study of HireVue, the vendor used three broad bands rather than fine grained rankings, explicitly to avoid clients being overwhelmed by small score differences that may not be meaningful in practice.  

The deeper irony is that application overload is getting worse, not better, because job seekers have their own automation. A Canadian report on hiring trends cited a survey of thousands of businesses across multiple countries and found employers dealing with application overload linked to mass apply tactics and AI generated résumés. It reported that the time spent sorting irrelevant applications per role could reach over nine days. That is the context in which companies say “let’s add an AI interview.” It is like trying to fix a house fire by bringing in scented candles. More smoke, better fragrance. 

If an AI interviewer is used at all, its only defensible purpose is triage, not replacement. That means the tool should be deployed narrowly, for specific roles, with clear rules, and with an explicit design goal: reduce human workload, not create a new genre of reading. The moment the tool becomes a universal first step, it stops being a filter and starts being a factory that manufactures more output than anyone can responsibly interpret.  

A conversation with nobody: candidate trust and respect

The second failure mode is social, and it is brutal: many candidates do not want to talk to a bot. This is not just nostalgia. It is logic. A candidate is deciding whether the company values people or processes. When the company’s first act is to turn the candidate into an unattended recording session, the message is heard clearly, even if nobody intended to send it: we are too busy for you. Candidates often react as if they are being asked to audition for the privilege of being acknowledged. 

Real candidates have described these experiences in language that should make any employer wince. One job seeker described the voice of an automated interviewer as sounding like Siri and called the experience creepy. Another described joining what they expected would be a real interview only to find themselves alone, talking to themselves, without the normal human signals that help a person calibrate and feel respected.  

Survey data supports the discomfort. A survey conducted by Talker Research for Newsweek reported that 43 percent of respondents felt uncomfortable with AI conducting job interviews, with a much smaller share expressing comfort and a sizable group uncertain. Another broader survey effort from Consumer Reports reported that Americans were uncomfortable with the use of algorithms in high stakes decisions, including employers analyzing video job interviews.  

Candidate drop off data also points the same way. A candidate experience compilation from CareerPlug reported that a meaningful share of job seekers have abandoned applications that require one way video interviews, and that most job seekers still prefer an in person interview format. You do not need to romanticise the old world to respect this. Hiring is a two way evaluation. When candidates feel disrespected, they exit, and the people most able to exit are often the people you most want. 

This is why the “nobody wants to talk with an AI interviewer” argument cuts deeper than feelings. It is about market dynamics. If a company wants to hire top talent, it has to compete for that talent. In competitive labour markets, the interview is part of the pitch. Using a bot as the opening act is like inviting someone to dinner and greeting them with a vending machine. Efficient, yes. Also wildly missing the point. 

When stakes rise, the bot breaks: senior and complex roles

The third failure mode is practical competence. Senior roles are not complicated because they involve more keywords. They are complicated because they involve judgement under uncertainty. A senior hire can reshape culture, budgets, risk posture, and team psychology. The cost of a poor hire is not a slightly lower performance review. It can be missed strategy, broken trust, and an expensive clean up that takes longer than the original hiring process ever would have. 

So what happens when you run a senior candidate through an AI interviewer. 

First, you flatten the signal. Senior candidates do not answer questions like multiple choice tests. They ask clarifying questions. They challenge assumptions. They negotiate scope. They spot contradictions and explore them. They do not just perform, they probe. Many automated interview formats struggle with that because the format is designed for throughput, not depth. 

Second, you amplify bias risks in subtle ways. If your AI interview relies on speech recognition, transcription, or language scoring, you inherit known problems with accent and non standard speech. A socio legal study by Natalie Sheard discussed how language based assessment can disadvantage candidates, including through documented word error rates that can rise substantially for non native English speakers with certain accents. Broader reporting on speech recognition has similarly highlighted that systems can have meaningfully higher error rates for diverse speech and dialect patterns, which matters when speech becomes an input into hiring judgement.

Third, you create a respect problem that is amplified at senior levels. A senior candidate is not merely applying. They are evaluating whether the leadership team is serious. Sending them to a bot can feel like being invited to someone’s home and then being told the housekeeper will do the bonding while the hosts are “just finishing something.” People remember that. 

Public sector guidance in The United States and Canada is unusually blunt about the danger here. Public Service Commission of Canada guidance warns that using AI to assess candidates, including automated scoring of interviews, is risky if the tool is opaque or complex, and it explicitly describes a scenario where an AI tool conducts video interviews and scores verbal and non verbal cues without transparent mechanisms as an inappropriate use case. This is not anti technology rhetoric. It is adult supervision. 

The paradox is that the more complex the role, the more tempting it is to automate, because the decision is hard and the stakes are high. Yet those two facts make automation less appropriate as a substitute. Complex hiring needs better structure, better evidence, and better interviewer capability. It does not need a chatbot in a blazer. 

The dark corners: black boxes, cheating, and hallucinations

This is where AI interview systems go from awkward to dangerous. Not always. Not every time. But often enough that any serious employer needs to stop treating it like a cute experiment. 

The black box problem starts with a basic governance issue: if the system produces a score or recommendation, can you explain it to a candidate, to leadership, or to a regulator. Many organisations cannot. That is why laws and policies are increasingly focused on transparency, notice, and audits. 

In New York City, Local Law 144 created obligations around bias audits and candidate notification for automated employment decision tools. An audit report from the New York State Comptroller summarised requirements such as conducting a bias audit and notifying candidates about the use and data collection of these tools. When governments write laws about your product category, it is a hint that society has noticed your industry is playing with matches near a gas station. 

Even audits can become theatre if they are limited, not independent enough, or framed as public relations. An analysis from Brookings Institution discussed how algorithm audits can range from meaningful accountability to PR, noting that some vendor commissioned audits examined only documentation, did not independently analyse models or data, and were not representative of the vendor’s full set of tools.  

Now layer on cheating. Recruiters are not the only ones with automation. Candidates have it too, and it is getting sophisticated. There are at least three kinds of “cheating” behaviour now colliding in hiring. 

One is assistance during interviews, including tools marketed to quietly listen and feed suggested answers in real time. Major employers have responded. Amazon reportedly issued guidance to recruiters that using generative AI tools during interviews can be grounds for disqualification, framing it as an unfair advantage that interferes with assessing authentic skill.  

The second is impersonation and deepfake style identity fraud. This is no longer a sci fi anecdote. Survey data from Resume Genius reported that 17 percent of hiring managers had encountered candidates using deepfake technology to alter video interviews. The issue is serious enough that Checkr surveyed thousands of managers and reported that only 19 percent were extremely confident their processes would catch a fraudulent applicant, while a notable share reported financial losses over 50,000 dollars and even over 100,000 dollars due to hiring or identity fraud.  

The third is organised infiltration risk, particularly in remote technical hiring. Palo Alto Networks and its Unit 42 have documented the use of synthetic identities and real time deepfakes as part of remote work fraud schemes, including patterns like repeated virtual backgrounds across personas. Separate reporting on automated interview systems in university admissions has also shown that deepfake attempts are detectable in the wild, even if currently limited in proportion.  

An AI interviewer makes these risks easier to scale. A human interviewer can demand a spontaneous clarification, ask for a live screen share, request an on the spot reasoning walk through, or pick up on the tiny hesitations that often accompany deception. An AI interviewer is usually just waiting for the next audio segment. It is not suspicious. It is polite. It will happily accept nonsense at machine speed. 

Then we arrive at hallucinations, the most under appreciated hazard in the AI interview story. You can build an AI interviewer that only asks fixed questions you wrote, like a remote controlled puppet. That can reduce risk, but it also reduces usefulness and turns the system into a glorified form. Or you can let it generate follow ups, interpret answers, summarise, and score. That is where hallucinations move from amusing to catastrophic. 

The research world is explicit that hallucination behaviour is not a rare edge case in modern large chatbots. A major survey in the Association for Computing Machinery literature summarised how hallucination remains a persistent problem, and it reviewed benchmarks and detection methods developed precisely because the issue is structurally difficult. Very recent work continues to focus on factuality evaluation precisely because hallucination is still a live problem into 2026, with open access reviews published in early 2026 cataloguing the ongoing challenges and the limits of existing fact checking and evaluation methods.  

Even in more operational evaluation settings, hallucination shows up when systems should say “I do not know” but instead produce confident fiction. Independent benchmarking reported in late 2025 found extreme hallucination behaviour in uncertainty scenarios for at least one major model, illustrating that “confident nonsense” is not a solved problem. And industry evaluation projects that track hallucination rates across models exist because the rates are measurable and non trivial even when the system is summarising a provided document.  

In an interview context, hallucinations can show up in several ugly forms. The interviewer can generate a misleading follow up, misinterpret a candidate’s answer, fabricate a skill inference, or produce a polished summary that sounds authoritative but subtly distorts what was actually said. Then a human, already overloaded, reads the summary instead of watching the video, and the hallucination becomes organisational “memory.” That is how bad decisions get made with great confidence. 

When you combine black box scoring, candidate cheating, and hallucinations, you get a system that can be wrong in ways that are hard to detect, hard to contest, and hard to explain. Hiring is not a joke. In a small or mid sized company, one wrong hire can break a team. In a larger company, one flawed pipeline can quietly discriminate at scale and attract the kind of scrutiny that shows up with subpoenas and headlines. 

A better ending: TP360 and the return of adult supervision 

The goal should not be to declare technology evil. The goal should be to stop using technology like an irresponsible person with a shiny new toy. 

There is a difference between replacing interviews and augmenting them. The best path forward is not “AI interview.” It is interview intelligence that makes humans better at interviewing, while preserving exactly what interviews are for: human judgement, mutual evaluation, and culture level truth telling. 

TP360’s own writing has already put its finger on the core hypocrisy of the AI interview trend. Plenty of platforms generate mountains of interview output and then hand it back to the same humans who supposedly did not have time in the first place. The fix is not to generate more. The fix is to build a system that reduces noise before the interview and increases quality during the interview. 

The most interesting part of TP360 is not the idea of automating the conversation. It is the idea of re engineering the pipeline so the conversation becomes worth having. TP360 positions itself as a talent acquisition platform that reduces cognitive overload through a structured, multi parameter assessment, expands the searchable talent pool through broad profile access, and then provides guided interview intelligence so the human interviewer is more prepared, more consistent, and less reliant on gut feel.  

That approach aligns with the direction of serious governance. NIST’s risk management framework is built around context, accountability, and managing real world harms rather than worshipping novelty. Canadian public sector guidance insists that hiring managers remain accountable, that decisions must be explainable, and that tools should not be used in opaque ways that candidates cannot understand or contest. Even Ontario’s legal landscape is drifting toward transparency, with job posting disclosure requirements coming into effect in 2026 for certain employers using AI in screening or selection.  

So here is what winning looks like. 

Winning is a hiring process that uses technology to reduce the grunt work, cut through application noise, and support structured evidence, while refusing to outsource human respect. Winning is a pipeline that does not pretend that a bot can replace a hiring manager, the same way a calculator does not replace an engineer. Winning is a platform that treats interviews as a high trust human moment, then surrounds that moment with the right safeguards: transparency, clear scoring logic, defensible criteria, identity verification, and a design that assumes people will attempt to cheat because the incentives exist. 

If TP360 becomes the standard, the “AI interview” becomes a relic, like faxing your résumé or pretending that “tell me about yourself” is a screening technique. Interviewing becomes what it always should have been: a well prepared, evidence guided conversation between humans, where technology does the boring parts quietly, and humans do the human part proudly.