The remote technical interview had a good run. A candidate shares a screen, solves an algorithm problem, and the interviewer observes their reasoning. That format assumed the candidate had nothing but their own knowledge. Real-time AI assistance — running on a second monitor, a phone, or an overlay the interviewer cannot see — removed that assumption entirely.
The response from most hiring teams has been to reach for detection. That is the wrong tool, and the reasons are worth understanding before you buy one.
What changed in 2026
- Assistance became invisible and instant. Tools that read a screen and produce a solution in seconds made the gap between assisted and unassisted candidates enormous on classic puzzle questions.
- Detection claims outpaced detection accuracy. Vendors offering to identify AI-assisted candidates faced the same false-positive problems that plagued AI text detection in education.
- Onsite interviews returned selectively. Some employers reintroduced in-person final rounds, accepting the cost and the reduced candidate pool for the integrity.
- The legitimacy question sharpened. As AI tools became standard in daily engineering work, banning them in an interview started to look like testing a skill the job does not require.
Approaches compared
| Approach |
Effectiveness |
Cost |
| AI detection software |
Low; false positives are the main output |
High reputational and fairness risk |
| Invasive proctoring |
Moderate; defeated by a second device |
Poor candidate experience, drop-off |
| In-person final round |
High |
Travel cost, narrower candidate pool |
| Questions about candidate own work |
High |
Requires skilled interviewers |
| Collaborative debugging on unfamiliar code |
High |
Preparation effort per role |
| Explicitly AI-allowed exercises |
High, if scored correctly |
Requires rethinking the rubric |
The bottom three are where the durable answers are. They share a property: they measure something that assistance does not substitute for. An AI can write a binary search. It cannot tell an interviewer why the candidate chose a particular tradeoff in a system they actually built, or answer three levels of follow-up about a decision that was never made.
Interviews that hold up
Ask about their work. A conversation about a project the candidate genuinely did — the constraints, the alternatives rejected, what went wrong, what they would do differently — is very hard to fake in real time, because the follow-up questions go somewhere that generated text cannot follow.
Give them unfamiliar code to debug. Reading and reasoning about an existing system is closer to the job than writing a function from scratch, and it produces a conversation rather than a solution artifact.
Or allow AI explicitly and grade the judgment. If the role involves using these tools daily, watch how the candidate uses one: whether they verify output, notice when it is wrong, and can explain what it produced. That measures something the job actually needs. The skill set involved is much closer to what AI coding agents require in practice than a whiteboard puzzle ever was.
For candidates, the practical implication is that the interview is shifting toward depth on real experience. Being able to discuss your own work in detail, including the parts that went badly, is now more valuable than pattern-matching a hundred practice problems.
Common mistakes
- Accusing a candidate based on detector output. These tools produce false positives, and an accusation you cannot substantiate is both unfair and a liability.
- Adding proctoring without redesigning questions. It degrades experience for everyone and does not solve the second-device problem.
- Banning AI for roles where AI is part of the job. You are filtering for a skill the role does not use.
- Keeping the same rubric while allowing tools. If assistance is permitted, what you score has to change.
- Treating this as a candidate-integrity problem only. The format was fragile; the tools exposed it.
FAQ
Do AI detection tools for interviews work?
Not reliably. They share the underlying difficulty of all AI-output detection, and the consequence of a false positive in hiring is more severe than in most other contexts.
Should we go back to in-person interviews?
It solves the problem and costs you candidate reach and scheduling flexibility. Many teams use it only for a final round, which is a reasonable compromise.
Is using AI in an interview always cheating?
Only if the employer said not to. If tool use is permitted and the interview is designed around it, it is just working.
What should candidates do if unsure of the rules?
Ask before the interview starts. An explicit answer protects both sides, and the question itself signals good judgment.
Where to go next
For the tools reshaping the work itself, read AI coding agents compared. For workplace norms around disclosing tool use, AI disclosure at work.