We Read Every Negative AI Interview Review. Most Had Nothing to Do With AI.

In Part 1, we published the headline number: 83% of candidates rated their AI phone screen Excellent or Good. Classet has conducted hundreds of thousands of AI interviews across hundreds of organizations, and we have reviewed the candidate feedback closely. The most common unprompted comment was that it felt like talking to a real person.

But we promised to come back for the other 17%. The Poor and Fair ratings. The ones that LinkedIn commenters would screenshot and say "See? Candidates hate this."

So we did what nobody in the AI hiring debate seems to do. We read every single negative response. Categorized every root cause. And the story those reviews tell is not the one most people expect.

The rating breakdown, in full

Here is the complete distribution from continuous post-interview feedback collected over 30 days:

About 13% rated the experience Poor. Another 4% rated it Fair. That is roughly one in six respondents in the negative-to-mixed range.

Those numbers are real. We are not going to minimize them or explain them away with a hand wave. But when you pull those responses apart and focus on the ones where candidates actually told us what went wrong, the reasons behind the ratings tell a very different story than "candidates hate AI interviews."

The #1 root cause: device and network failures

We filtered to just the negative and mixed ratings where candidates left written feedback explaining what went wrong. Here is what they actually said:

Device and network failures: 70% of negative feedback with comments. This was the dominant theme by a wide margin. Microphones that did not pick up audio. Calls on bad cell signal. Background noise drowning out the conversation. Bluetooth headphones disconnecting mid-interview. Candidates calling from environments where any phone call, human or AI, would have gone badly.

Pacing and interruption issues: ~13%. Joy moved to the next question before the candidate finished answering. This is legitimate product feedback and the kind of signal that actually helps improve the experience.

Wanted a redo: ~10%. These candidates had their call drop or hit a technical issue and asked to interview again. Read that twice. They did not walk away frustrated. They wanted another chance. That is an engagement signal.

Language mismatch: ~7%. The interview was configured in the wrong language for the candidate. That is an employer setup issue. As we showed in Part 1, when language is configured correctly, candidates respond positively. Seven Spanish-language candidates gave organic positive feedback without any bilingual prompting.

The adjusted positive rate: 87%

A significant share of Poor ratings came from candidates who explicitly described device or connectivity failures in their written comments. Their hardware failed them before the interview had a chance to work.

When you remove those responses, cases where any phone-based interaction would have gone the same way, the positive rate moves from 83% to 87%.

We are transparent about this adjustment because the methodology matters. We did not remove every negative response. We only removed the ones where the candidate themselves described a device or network problem as the reason for the poor experience.

That is a conservative adjustment. The real number of device-related failures is probably higher, because not every candidate who had a mic problem wrote a comment explaining it.

What hiring leaders get wrong about negative AI interview reviews

There is a pattern in how hiring leaders process AI screening objections. It goes like this:

Someone shares a negative data point. "13% rated it Poor." The room assumes that means 13% of candidates were offended by talking to a machine. The conversation shifts to "candidates are not ready for this." And the evaluation stalls.

But that is not what 13% means. The data shows that the most common negative experience had nothing to do with AI at all. It had to do with the same problems that plague every phone-based interaction: bad signal, broken microphones, and noisy environments.

If a candidate calls a human recruiter from a bus with a dead mic, that phone screen goes badly too. Nobody blames the recruiter. But when the same thing happens with an AI phone screen, it becomes evidence that AI interviewing does not work.

That framing error is costing companies time. Teams that delay adopting automated phone screening because of a perceived candidate experience risk are often comparing AI to an idealized version of human screening that does not exist. A fairer comparison: AI screening with occasional device issues versus human screening with inconsistent availability, scheduling delays, and interviewer bias.

The "wanted redo" signal nobody talks about

About 10% of the negative cohort with written feedback asked to redo their interview after a technical issue. That is a small share, but the behavioral signal is strong.

A candidate who hates the format walks away. A candidate who asks for another shot is telling you they were engaged enough to want a fair attempt. They were rejecting the dropped call, not the format.

This pattern showed up in our analysis of 70,000 voice AI interviews too. Candidates who experience a technical issue and get reconnected tend to complete the screen at rates comparable to candidates who had a clean first attempt. The willingness to re-engage shows up consistently.

Pacing and environment: the feedback we already acted on

A small percentage of candidates said Joy moved to the next question too quickly. That was the most actionable feedback in the dataset because it was the only negative theme that pointed to something the product could directly address.

So we addressed it. Joy now includes pacing controls that employers can configure per role. Customers can adjust how long Joy waits after a candidate finishes speaking before moving on, so the conversation matches the pace of the person on the other end.

Organizations also have full control over the SMS message that goes out when a candidate first applies. That means employers can guide candidates to take the call when they have a good signal and a quiet space. You cannot fix a candidate's broken AirPods, but you can set expectations upfront so candidates know to find a good environment before the interview starts. That alone targets the single biggest source of negative feedback in the entire dataset.

And the underlying voice models keep getting better. We upgrade Joy's conversation engine on a quarterly basis, moving to the latest voice AI models as they become available. Each generation improves natural turn-taking, handles interruptions more gracefully, and adapts better to different speech patterns. The pacing feedback from this study was a useful signal, but the product has already moved past it.

For hiring leaders evaluating AI recruiting tools, the question to ask is not "do any candidates have a bad experience?" Some always will, with any format. The question is whether the negative feedback points to fixable problems or fundamental flaws. Pacing and environment issues are fixable, and we have already fixed them. A systematic rejection of the format would be a different story, but that is not what this data shows.

What this means if you are evaluating AI phone screening

Separate device problems from product problems. When you review candidate feedback on any phone-based screening tool, ask whether the complaint is about the interaction or the infrastructure. A negative rating from a candidate with a broken microphone tells you nothing about whether candidates accept AI interviews.

The adjusted number is the one that matters for product decisions. 87% positive is the rate that reflects how candidates experience the actual AI interaction, once you remove the cases where the interaction could not happen at all. Use that number when comparing to your current phone screen satisfaction rates, if you measure them at all.

Negative feedback that asks for a redo is positive signal. Candidates who want to try again are candidates who bought into the format. Build your process to accommodate reconnection and you convert a dropped call from a failed screen into a completed one.

Pacing feedback is a feature roadmap. A small percentage of candidates wanted more time to answer. That is product feedback worth acting on, and Joy already has, with configurable pacing controls, pre-interview environment prompts, and quarterly voice model upgrades.

Stop comparing AI screening to perfect human screening. The right comparison is AI screening at 87% positive versus your current process, including the candidates who never get a call back, the ones who wait five days for a screen, and the ones who drop out because the recruiter was double-booked. Our research on how candidates feel about AI interviews consistently shows that speed and consistency matter more to candidates than whether the voice on the other end is human.

AI interview negative reviews: what candidates actually report

This post is Part 2 of a two-part series on AI interview candidate experience. Part 1 covers the 83% positive rate, first-timer reactions, stress reduction, and multilingual reach.