In a nutshell
- You're already making this decision, even if you think you aren't. Any stage you don't supervise live is AI-allowed in practice. The only question is whether you design for it.
- Big companies have split. Canva, Shopify and a Meta pilot let candidates use AI in coding interviews. Anthropic and Amazon keep live interviews AI-free. Google is adding an in-person round.
- Both positions are defensible. The one indefensible position is "don't ask, don't tell": no stated rule, and a quiet penalty for anyone caught.
- The practical answer for most teams is two lanes. Some stages allow AI and score judgment. Others are independent and score unaided reasoning. You tell candidates which is which.
- Allowing AI doesn't lower the bar. It moves it from "can produce code" to "can own code", which is harder to fake.
Why is this suddenly a decision I have to make?
Because the old default stopped working quietly.
Karat, which sells human-plus-AI technical interviews, surveyed 400 senior engineering leaders and found 71% say AI makes technical skills harder to assess, while 62% of organisations still ban AI in interviews. The same leaders estimate more than half of candidates use AI anyway.
That gap is the problem. A ban you can't see being broken doesn't produce a clean test. It produces a test where honest candidates compete against hidden help, and you can't tell which is which.
Candidates aren't even sure what the rules are. Greenhouse's 2025 survey of 2,200 job seekers found only 19% of US respondents consider using AI in a live interview to be cheating, and 27% had never seen an employer AI policy. Greenhouse sells hiring software, but the direction matches what you'd expect.
So whatever you decide, the first step is the same: decide, and write it down.
What are well-known companies actually doing?
Here's what has been publicly confirmed. I've left out format details that only appear on interview-prep sites.
| Company | Position | What they've said |
|---|---|---|
| Canva | AI expected | Replaced its CS-fundamentals screen with "AI-Assisted Coding". Expects backend, ML and frontend candidates to use tools like Copilot, Cursor and Claude |
| Shopify | AI allowed | Head of Engineering Farhan Thawar: "You let them use whatever they want" |
| Meta | Piloting | Developing a coding interview where candidates have an AI assistant |
| Anthropic | Mostly AI-free live | AI fine for prep; live interviews are "all you–no AI assistance unless we indicate otherwise" |
| Amazon | Not allowed | Candidates acknowledge they won't use unauthorised tools such as GenAI |
| Adding in-person | At least one in-person round "to make sure the fundamentals are there" |
A few of these are worth reading in full.
Canva's engineering blog explains that its own tests showed AI "can trivially solve traditional coding interview questions". So instead of fighting that, it changed what the interview asks for.
Shopify's Thawar put the reasoning bluntly on the Pragmatic Engineer podcast: "If they don't use a copilot, they usually get creamed by someone who does." He also said he wants candidates to fix trivial bugs by hand rather than keep prompting. That detail matters. AI-allowed doesn't mean judgment-free.
Meta's internal post, as reported by Wired, said the new format "also makes LLM-based cheating less effective." That's the underrated argument for allowing AI: you can't cheat on a rule that doesn't exist.
And Anthropic publishes a stage-by-stage candidate AI policy that's worth copying as a structure even if you disagree with the content. It says exactly where AI is fine and where it isn't.
What's the case for banning AI in interviews?
It's stronger than the "it's 2026, get with it" crowd admits.
- You need to see baseline reasoning. AI tools are wrong in confident ways. The engineer you hire has to notice, debug and reason when that happens. You can only see that with the tool out of the picture.
- Comparison is cleaner. If everyone uses different tools at different skill levels, scores vary for reasons unrelated to the person.
- Candidates think it's fair. HackerRank's 2025 developer survey found 73% of developers think losing to AI-assisted candidates is unfair. That's vendor data, but it suggests honest candidates want a level field.
What's the case for allowing it?
- It's the job. If your engineers use AI daily, a no-AI interview tests a job that doesn't exist at your company.
- It removes the incentive to hide. When the tool is allowed, the overlay that hides it is pointless.
- It shows what you actually hire for. Do they read what the AI wrote? Do they run it? Do they notice the subtle bug? Do they know when to stop prompting and fix the line themselves?
So the honest conclusion: both sides are right about different things. Baseline reasoning and AI judgment are both real skills, and you need evidence of each.
Why is "don't ask, don't tell" the worst option?
Because it rewards the wrong people.
With no stated policy, the candidate who quietly uses AI gets a better score than the one who doesn't. The honest candidate who mentions it might get marked down. And your scores mix both groups, so you can't compare anyone.
A written policy turns "did they cheat?" into "did they follow the rules we gave them?" That question has an answer. You can start from our free AI interview policy template and adapt it to whichever position you take.
What is the two-lane model?
Label every stage as one of two lanes, tell candidates which is which, and score each lane differently.
| AI-allowed lane | Independent lane | |
|---|---|---|
| Measures | Judgment, verification, speed with tools, product sense | Baseline reasoning, debugging, explanation |
| Typical format | Work sample or paired build with their own tools | Live laddered problem, code review, defense of their work |
| What you score | What they asked for, what they kept, what they caught | Whether they can adapt, justify and own their answers |
| How it's enforced | It isn't. AI is allowed | Format: live, conversational, laddered, tied to their history |
That last row matters. Independent stages are not enforced by cameras. They're enforced by questions that a pre-generated answer can't survive: change a constraint on screen, ask why they chose this over the obvious alternative, ask what they shipped and what broke. I walk through that method in technical interview follow-up questions.
How do I decide the mix for my team?
Answer these four questions.
| Question | If yes | If no |
|---|---|---|
| Do your engineers use AI tools daily? | At least one AI-allowed stage, weighted heavily | Independent lane can carry more weight |
| Is any stage unsupervised (take-home, online test)? | It's AI-allowed. Label it and design for it | — |
| Will this hire make decisions others depend on (senior, staff, founding)? | Give the independent lane more weight | Balance roughly evenly |
| Is the role junior, high-volume? | Async AI-allowed exercise for triage only, then independent stages decide | — |
My default for a mid-level product engineer: one AI-allowed build or work sample, one laddered independent technical round, and one independent defense or review stage. That's three technical stages, roughly a third each.
How do I score an AI-allowed round so it doesn't just measure the tool?
Don't score the output. Everyone has the same tools, so output converges. Score four things instead:
- Verification. Did they read and run what the AI produced? Did they reject or edit anything, and say why?
- Detection. Did they catch a subtle bug, a misleading requirement, or a security trap you planted in the code?
- Tradeoffs. Did they say what they chose not to do?
- Tests. Did they test the risky part, or only the happy path?
And follow it with a short independent defense where they change their own code while you change the requirements. That's how you know the work is theirs. More on this in take-home assignments and AI.
Things not to score: how much code the AI wrote, which tool they used, or how elegant their prompts were. A clumsy prompt followed by a careful review beats a beautiful prompt followed by blind acceptance.
How do I tell candidates?
Send the policy with the first invite, link it from the job post, and repeat it at the start of each stage. A short version for a two-lane loop:
"Some of our stages are AI-allowed and some are independent. In the work sample and the paired build, use whatever AI tools you normally use. We'd like you to. In the live technical round and the review of your work, please work without AI assistance. We'll remind you at the start of each stage. If you need an accommodation for any stage, just ask."
And at the start of an independent stage:
"This one is independent, so no AI tools for the next 45 minutes. I care more about how you think than whether you finish. Talk me through it, including ideas you reject."
Talogy, which sells assessments, reported in 2025 that an upfront honesty agreement cut unauthorised help on its assessments from 28% to 13%. Stating the rule clearly isn't a formality. It changes behaviour.
Isn't allowing AI just lowering the bar?
No. It moves the bar.
"Write a function that reverses a linked list" was never the job. "Here's 400 lines an assistant wrote, three things in it are wrong, find them and extend it" is much closer to the job, and much harder for a weak engineer to fake. The tool produces the code. Only the candidate can explain why it's wrong.
Checklist: deciding your AI policy
- [ ] Every stage labelled AI-allowed or independent
- [ ] Every unsupervised stage labelled AI-allowed
- [ ] At least one independent stage that uses follow-up ladders
- [ ] AI-allowed stages scored on verification, detection, tradeoffs and tests
- [ ] Any offline work followed by a live defense
- [ ] Policy sent with the first invite and repeated at each stage
- [ ] Accommodations offered in the policy itself
What should I do next?
Take the free cheat-risk audit to see which of your current stages are exposed, then adapt the free policy template. If you want the two-lane loops already built, with question ladders by stack and AI-allowed work samples with rubrics, that's what Unscripted, the paid kit is for. If you're also wondering whether the classic algorithm round still has a place, read is LeetCode dead.