I want to write something useful about AI in interviews. The current discourse has two camps — "ban it, integrity matters" and "embrace it, it's the future" — and they're both wrong in the same way: they're not engaging with what an interview is actually trying to measure. Once you do engage with that question, the answer is harder than either camp wants it to be.

Here's the honest version, with the numbers and the moves for both sides.

What's actually happening (the numbers)

78% of technical candidates report using AI assistance during interviews. That's per BrightHire's 2026 hiring survey of 2,400 hiring managers and 8,000 candidates across mid-to-late-stage tech companies. Six points higher than 2025.

Of those, 71% don't disclose the AI use to the interviewer. Either because they weren't asked, or because the rules were ambiguous, or because they assumed the interviewer expected it.

47% of recruiters report that they've stopped using take-home assignments as a meaningful hiring signal. They still send them — but they've internally downgraded what take-home performance means. The signal got too noisy to trust.

31% of companies now explicitly ban AI assistance in technical interviews. Another 24% explicitly allow it. The remaining 45% have no policy, which functionally means the interview is unevenly run depending on the interviewer's individual stance.

The interview process, as a system, is broken in a way it wasn't in 2023. The systems we used to measure programming ability no longer measure programming ability — they measure a mix of programming ability, AI fluency, AI evasion, and willingness to use unsanctioned tools. Those are four different things, and the interview can't separate them.

Who's getting hurt, on both sides

This is the part the discourse usually misses. There are four cohorts and three of them are getting hurt:

• Strong engineers who don't use AI in interviews (because they'd rather show their unaided thinking). These candidates are getting beaten on speed and apparent confidence by candidates who do. They're losing offers to weaker engineers whose AI-augmented interview answers looked better in real time.

• Strong engineers who do use AI in interviews. They're often getting filtered out anyway, because experienced interviewers can usually tell. Or worse — they're getting the job, succeeding in the role, and the company is concluding that "AI use is fine" when really they hired a strong engineer who would have succeeded either way.

• Weak engineers who use AI to clear the bar. This is the big problem for companies. AI assistance lets a candidate produce convincing-looking code for problems they don't really understand. The hire happens, the on-the-job performance is bad, and the company concludes "we got fooled" and tightens interviews — which hurts the strong-honest cohort.

• Hiring teams. Every signal they had is now noisier. Take-homes don't work. Whiteboard interviews are theater. System design is the last reliable signal and even that's getting AI-augmented in coaching prep.

The only cohort that's not getting hurt is people who use AI cleverly and would have been hired anyway. That cohort is small.

What the "just ban it" companies are getting wrong

Banning AI in interviews has three problems and they're all real.

Problem one: it doesn't work. You can ask candidates to share their screen. You can require an in-person final round. You cannot detect a candidate who's running a second device with whisper-AI suggestions in their ear. The technology to assist invisibly is now consumer-grade, costs under $30/month, and a meaningful fraction of candidates are using it whether you ban it or not.

Problem two: it selects against the wrong things. A "ban AI" rule punishes candidates who would otherwise have disclosed AI use and worked with it transparently — i.e., the honest ones. The deceptive cohort uses AI either way.

Problem three: it doesn't measure what you actually want. You're hiring someone who's going to use AI heavily on the job. Whether they're good at solving problems without AI is, at best, a proxy for whether they're good at solving problems with AI — and proxies degrade fast in changing environments. Increasingly, "good at unaided coding" and "good at AI-augmented engineering" are different skills with mediocre correlation.

What the "just embrace it" companies are getting wrong

The "let candidates use AI freely" camp has the opposite problem. It optimizes for AI fluency, which isn't the same as the job skill.

When you let a candidate freely use AI in a 60-minute interview:

• The strong-engineer signal gets muddled — strong engineers and AI-fluent weaker engineers both produce good-looking output.
• The candidate's pure AI prompting skill matters more than their underlying engineering judgment.
• You can no longer compare a 2026 candidate to a 2023 candidate on the same scale, because the rubric changed under your feet.
• You select for candidates who happen to have practiced AI-augmented interview prep, which is a different population than candidates who happen to be good engineers.

"Embrace it" sounds modern. It produces hires whose on-the-job performance is more variable than the unaided-interview era.

What good hiring is doing now (the third path)

The companies hiring well in 2026 are doing something neither camp is: they're separating the interview into layers and matching the AI policy to what each layer is trying to measure.

I've talked to hiring leads at four companies that are getting better hire quality in 2026 than they were in 2024. The pattern they're converging on:

Layer one: pure-judgment interview, no AI allowed

A short (45-min) conversation about the candidate's actual recent work. Specific projects. Specific decisions. Specific tradeoffs. The interviewer's job is to find the place where the candidate's stated experience can't survive five follow-up questions. This part is unaided — no AI helps you remember why you actually chose Postgres over Mongo three years ago.

This is the highest-signal interview component in 2026. It's hard to fake, AI can't help you, and it measures the thing that actually predicts on-job performance: do they have real judgment in real situations.

Layer two: AI-augmented work session, fully disclosed

A 90-minute work session on a realistic task with full AI access disclosed. The candidate uses Claude / Cursor / their tool of choice, narrates what they're doing, and the interviewer watches the process, not just the output.

The signal here isn't "can they produce good code with AI" — almost everyone can produce passably good code with AI. The signal is how they work with the AI:

• Do they read the AI's output critically, or do they accept it uncritically?
• Do they ask the AI good questions, or vibey ones?
• When the AI produces something wrong, do they notice?
• When the AI produces something right but suboptimal for the situation, do they notice that?
• How do they communicate about their work while doing it?

This is hard to fake the way the take-home was easy to fake. The interviewer is in the room.

Layer three: design and operational judgment, conversational

A senior interviewer walking through a real ambiguous problem with the candidate. AI isn't banned per se, but the format — discussion, whiteboarding, back-and-forth — naturally limits how much AI helps. The signal is whether the candidate can hold ambiguity, push back appropriately, and think structurally.

This is the "is this person going to do well at staff-eng-shaped work" interview. It was always the hardest one to fake and it remains so.

What candidates should actually do

If you're a candidate interviewing in 2026, the honest moves are different than the discourse suggests.

Move one: get genuinely fluent at AI-augmented work before you interview. Companies are increasingly testing it directly. Showing up with rusty AI skills now is what showing up with rusty Git skills was in 2018 — disqualifying for a lot of roles.

Move two: practice the "talk about your real work" interview. Companies are weighting this more and more. The candidates who do well aren't the ones who memorize talking points — they're the ones who have actual specific stories with specific decisions and tradeoffs. If you can't remember why you made a particular call on a real project, that gap shows in three follow-up questions.

Move three: disclose your AI usage proactively. If the interview format allows AI, use it well and openly. If the interview format prohibits AI, prohibit it for yourself fully, even if you think you could get away with it. The honest game has higher long-term EV because the dishonest game catches up with you in references, in trial weeks, and in your reputation. The AI-fluent honest cohort is the cohort that's going to dominate the next decade of senior roles. The deceptive cohort gets shorter half-life careers.

What companies should actually do

If you're hiring, three things that will improve hire quality in 2026:

Stop relying on take-homes. They were always a weak signal. They're now actively noise. The 47% of recruiters who downgraded them are correct.

Build the three-layer process above. It costs more interviewer time per candidate. Hire quality is what matters; volume isn't the metric. The companies winning the hiring war in 2026 are running fewer, deeper loops with more senior-engineer interviewer involvement, and they're hiring slower and better.

Be explicit about AI policy on each layer. Ambiguity benefits deceptive candidates and hurts honest ones. Spell it out per stage. "This interview, no AI. This interview, full AI, please share your screen. This interview, your call but tell us what you're doing." Removing the ambiguity is the single highest-leverage thing most hiring orgs can do.

What I think the next two years actually look like

The interview, as an institution, is mid-correction. Three things converge by 2028:

• Most companies converge on some version of the three-layer pattern. The "ban AI / allow AI" debate fades because it was always the wrong frame.
• Take-homes mostly die. Even the companies that kept them downgrade their weight to ~5% of the hire decision.
• A new interview signal emerges: "can you actually work with an agent fleet under time pressure." This is going to become the canonical mid-stage interview for senior engineering roles, and the candidates who practice for it specifically will dominate the comp distribution.

If I had to pick the single number that captures the situation: 78% of candidates are using AI in interviews, but only 24% of companies have a coherent policy on it. That gap is the dysfunction. Closing the gap — on either side — is where the better outcomes live.

What I'm not saying

I'm not saying companies that ban AI in interviews are stupid. The intent is right — measure the thing the candidate brings, not the thing the tool brings. The implementation just doesn't work in 2026.

I'm not saying candidates who use AI without disclosure are dishonest people. Most are operating in environments where the rules are ambiguous and the consequences for asking are unclear. The system is set up to incentivize quiet AI use, and the right response is to fix the system, not to moralize at the candidates.

The hiring process broke in a specific way over the last two years. The fix is in reach. It just requires both sides to stop pretending the old process was fine and start designing a new one. The companies that do that hire better. The candidates who play that game honestly get hired into better seats. That's the whole game.

Further reading

• BrightHire 2026 — State of Technical Hiring report — the underlying survey
• Latent Space — interviewing in the agent era — pattern-matching from senior engineering hiring
• Hacker News — discussion on the BrightHire numbers — particularly the thread on take-home death
• Erik Bernhardsson — why interview signal is decaying — the senior IC view
• Lenny's Newsletter — how product hiring is adapting — adjacent dynamics in PM hiring