Yes, you can require candidates to use AI in technical interviews, and three of the largest engineering employers now do. Canva, Meta and Google have each opened one specific round to AI assistants, told candidates in advance, and rewritten the problem so a single prompt cannot solve it. None of them flipped a company-wide switch. The decision that matters is not whether to allow AI. It is which stage allows it, and what you score when it does.

## Canva stopped policing AI in interviews and started requiring it

On 11 June 2025, Canva engineering published a post by Simon Newton with a title that left no room for interpretation: "Yes, You Can Use AI in Our Interviews. In fact, we insist you do."

For backend, machine learning and frontend candidates, Canva deleted its Computer Science Fundamentals interview, the algorithms and data structures round, and replaced it with a new competency called AI-Assisted Coding. Candidates bring their own tools. The post names Copilot, Cursor and Claude. The rationale was blunt: "Rather than fighting this reality and trying to police AI usage, we made the decision to embrace transparency and work with this new reality."

The most instructive part is not the policy. It is what happened to the problem. Canva retired self-contained exercises like Conway's Game of Life and replaced them with ambiguous product scenarios. Their published example: build a control system for managing aircraft takeoffs and landings at a busy airport. You cannot one-shot that. There is no answer to memorize and no way through it without deciding what to build first and defending the decision out loud.

Canva also told candidates what was coming. In the company's own words, "we now inform candidates ahead of time that they'll be expected to use AI tools, and we highly recommend they practice."

The internal justification was mundane and, for most engineering teams, familiar: almost half of Canva's frontend and backend engineers were already daily active users of an AI coding tool. The interview had drifted away from the job.

## The defensive playbook is not producing signal

Most teams have gone the other way, and the data suggests the effort is wasted.

Karat's 2025-2026 AI Workforce Transformation Report, published 7 January 2026 and based on 400 engineering leaders across the US, India and China, found three things that only make sense together. **62% of organizations still prohibit AI use in technical interviews.** Those same leaders estimate **over half of candidates use AI anyway**, despite being told not to. And **71% say AI has made technical skills harder to assess**.

Read that as one sentence: most teams ban it, most teams believe the ban fails, and assessment got harder regardless. The prohibition is buying compliance theatre, not measurement.

The enforcement side reads worse. Internal Amazon documents seen by Business Insider in February 2025 tell candidates not to use GenAI tools during interviews "unless explicitly permitted," warning that failure to comply "may result in disqualification." Amazon also circulated guidance helping interviewers spot AI use, with tells like candidates typing while a question is being asked, reading rather than responding, and eyes wandering. That is amateur polygraphy in a 45-minute window, and it is the best available argument that detection does not work. We covered that arms race in [hiring engineers when everyone has the same AI](/blog/hire-engineers-when-everyone-has-same-ai).

The cost lands on your calendar too. Robert Half surveyed more than 2,000 US hiring managers in November 2025: **67% say reviewing AI-generated applications has slowed hiring**, and **65% say verifying candidate skills has gotten harder**.

## What Canva, Meta and Google actually changed

Three companies, three different loops, one identical structural move. Each scoped AI to a single stage, kept the rest unchanged, and told candidates before the stage rather than after.

| Company | What changed | Scope | Tooling |
|---|---|---|---|
| **Canva** (Jun 2025) | Replaced the CS Fundamentals interview with an AI-Assisted Coding competency | Backend, ML, frontend roles; one competency | Candidate's own (Copilot, Cursor, Claude) |
| **Meta** (pilot from Oct 2025) | Added an AI-enabled coding round, 60 minutes, in CoderPad with a built-in assistant | Replaces **one of two** onsite coding rounds; the other stays no-AI | Supplied in the environment |
| **Google** (pilot from H2 2026) | AI-assisted code comprehension round: analyze an existing codebase, find bugs, improve performance | Junior and mid-level US roles only | Gemini, supplied |

Meta's spokesperson framed it as unremarkable: "We're obviously focused on using AI to help engineers with their day-to-day work, so it should be no surprise that we're testing how to provide these tools to applicants during interviews." Internal framing added a second motive: an AI-enabled round is more representative of the real developer environment and makes LLM-based cheating less effective. When the tool is permitted, smuggling it in stops being an advantage.

Google's Brian Ong, VP of Recruiting, gave the same reasoning on the record: the pilot exists "to be more reflective of how our teams are operating in the AI era." Per the internal document reported by Business Insider, interviewers assess three things: prompt engineering, output validation, and debugging skills.

The fourth data point is a reversal, and it is the most persuasive one. Anthropic prohibited applicants from using AI in May 2025 and abandoned the ban by July 2025. What it landed on is not a yes or a no: AI is permitted for application materials and interview prep, permitted in take-homes when the instructions say so, and not permitted in live interviews unless stated otherwise. An AI lab could not make a blanket ban stick on its own funnel, and it replaced the ban with a stage-by-stage rule.

## AI-open or AI-closed is a stage decision, not a company decision

"Do we allow AI?" is the wrong question, because the loop is not one thing. A take-home and a system design conversation measure different faculties, run under different conditions, and have completely different exposure to AI. A single company-level toggle guarantees one of those stages is misconfigured.

Two definitions, because the vocabulary is doing real work here:

- **An AI-open stage** expects the candidate to use an assistant, and the rubric scores how they direct, verify and correct it.
- **An AI-closed stage** expects no assistant, and stays short and narrow enough that unaided performance is a fair thing to ask for.

A default loop looks like this:

| Stage | Policy | Why |
|---|---|---|
| Application and written materials | **AI-open**, no disclosure theatre | Roughly 40% of candidates already use AI here, per Gartner's 2Q25 survey of 3,000 candidates. Policing it costs more than it returns. |
| Take-home or code assignment | **AI-open**, problem redesigned | This is where the job actually lives. Unsupervised and unenforceable anyway, so make it honest and raise the difficulty. |
| Live coding | **Pick one.** AI-open with a genuinely hard, ambiguous brief, or AI-closed and deliberately short | Meta runs both, one of each. That split is the single most copyable structure in this article. |
| System design | **AI-closed by default** | You are testing whether they can reason about constraints in conversation, not whether a model can produce a reference architecture. |
| Architecture and debugging on your own codebase | **AI-open** | Matches the day job exactly. Comprehension and correction, not generation. |
| Reference and values conversations | Not applicable | No policy needed. Do not write one. |

The rule of thumb: keep a stage AI-closed only when you can articulate what unaided performance tells you that the AI-open version would hide. "It feels more rigorous" is not an answer. "This round tests whether they can hold a distributed systems tradeoff in their head while I push back on it" is.

## Redesign the assignment so one prompt cannot solve it

Opening a stage to AI without changing the problem is the failure mode. If your take-home was solvable by a competent engineer in three hours in 2023, an assistant finishes it in twelve minutes now, and you have converted an assessment into a formality. Four changes do most of the work:

1. **Make the brief ambiguous.** Canva's aircraft control system has no stated schema, no defined scope, and no obvious first move. The candidate has to choose, and that choice is the signal.
2. **Start from imperfect code, not a blank repo.** Ship a starter repository with real problems in it: a subtle race condition, an N+1 query, a test that passes for the wrong reason. Comprehension and correction is a harder target than generation, which is why Google's pilot round is built on it.
3. **Require an artifact about the process.** Ask for a pull request description naming one thing the assistant got wrong and how they caught it. Candidates who never verified anything cannot fake that paragraph convincingly.
4. **Score the tradeoffs they raised, not the tests they passed.** If your grading is a test suite, AI wins the assessment and teaches you nothing.

## The scorecard: grade the collaboration, not the output

In an AI-allowed technical interview, score six things: how the candidate decomposes an ambiguous brief, when they choose not to use AI, how they validate generated output, how they diagnose and correct it, whether they raise production concerns unprompted, and whether they can explain code they did not type.

| Criterion | Strong looks like | Weak looks like |
|---|---|---|
| **Decomposition** | Splits an ambiguous brief into ordered, testable pieces before prompting | Pastes the whole brief in and takes the first plausible output |
| **Tool selection and restraint** | Uses AI where it pays; writes the tricky 20 lines by hand | Prompts for everything, including things faster to type |
| **Output validation** | Reads generated code adversarially, checks edge cases, runs it | Trusts it because it compiles |
| **Correction** | Diagnoses why the generated code is wrong and fixes the cause | Re-prompts until something passes |
| **Production standards** | Raises error handling, tests, security, cost and observability unprompted | Ships the happy path |
| **Explanation** | Defends every decision in code they did not type | "The AI did that" |

This is a composite of Canva's five published competencies and Google's three stated criteria, and it builds on the judgment rubric in [how to screen engineers for real skill in the AI era](/blog/screening-engineers-ai-dependency). The weighting mechanics are in our guide to [structured scorecards](/blog/skills-based-hiring-structured-scorecards), and weighting matters more here than in a traditional loop, because Output validation and Correction are where the variance is.

That weighting has evidence behind it. In Stack Overflow's 2025 Developer Survey, the top AI frustration, at **66%**, is "AI solutions that are almost right, but not quite," and **45.2%** say debugging AI-generated code takes more time. Active distrust of AI accuracy rose to **46%** from 31% a year earlier, while only 3.1% report high trust.

Now the harder number. A [METR randomized controlled trial](https://arxiv.org/abs/2507.09089) published in July 2025 gave 16 experienced open-source developers 246 real issues from their own repositories. With AI tools available, they were **19% slower**. They estimated afterwards that AI had made them **20% faster**. The sample is small, but treat the result as the argument for the AI-open interview rather than against it. If experienced engineers cannot tell whether their assistant is helping, then knowing when to stop prompting is a scarce, hard-to-fake skill. It is also invisible in an AI-banned interview, because the candidate who accepts an almost-right answer and the candidate who catches it look identical on a whiteboard.

## Tell candidates before the stage, not after

The disclosure half is not a courtesy. It is the fraud control.

Gartner's 2Q25 survey of 3,000 job candidates found that **6% admitted participating in interview fraud**, meaning posing as someone else or having someone pose as them, and forecasts that one in four candidate profiles worldwide will be fake by 2028. Jamie Kohn, Senior Research Director in Gartner's HR practice, put it plainly: "It's getting harder for employers to evaluate candidates' true abilities, and in some cases, their identities."

Gartner's recommended mitigation is not a detection vendor. It is setting clear expectations and communicating hiring standards, explicitly including how you define acceptable AI use. Even a research house focused on candidate fraud lands on the constructive answer.

There is a second reason to publish. The same survey found only **26% of applicants trust AI to evaluate them fairly**. Candidates walking into an unstated AI policy are guessing at your standard and assuming the worst, and that ambiguity costs you good people at the top of the funnel.

Copy this, adjust the specifics, and put it on the stage itself:

> **Take-home assignment: AI tools expected.** Use whatever you use at work, including Copilot, Cursor or Claude. We are not testing whether you can write code unaided; we are assessing how you scope an ambiguous problem, verify what your assistant produces, and correct it. In your pull request description, tell us one thing the assistant got wrong and how you found it. Your next round is a live conversation about this code with no AI, so submit only work you can defend line by line.

That last sentence does more for assessment integrity than any proctoring product. If the AI-open stage is followed by an AI-closed conversation about the same code, faked submissions collapse on contact.

<div class="blog-inline-cta">
  <p><strong>Building this loop this week?</strong> Kit lets you set the policy, the candidate-facing wording, and the scorecard independently on every stage, so an AI-open assignment and an AI-closed live round sit in the same pipeline.</p>
  <p><a href="/users/sign_up">Start your free trial</a></p>
</div>

## The fairness problems nobody has solved

Three of them, and none has a published answer.

**Tool access.** Useful tiers of Cursor, Claude and Copilot cost money, so requiring AI can impose a subscription on an unemployed candidate. Google supplies Gemini and Meta supplies an assistant inside CoderPad. Canva's post does not address tool ownership at all. Supply the tool or supply the budget, and never let tool choice be a scored variable.

**Practice asymmetry.** Canva "highly recommends" candidates practice. Anyone currently employed somewhere with an AI budget has been practicing for a year on the company's dime. Publish the format and a sample problem, not just the policy.

**Uneven gains.** Karat found AI raised average productivity by 34%, self-reported, but unevenly, widening the gap between strong and weak engineers. For junior hiring that cuts both ways, and nobody has resolved it.

And if a take-home now runs several hours with paid tools attached, pay for it. [Candidate payouts](/docs/managing-jobs/03-candidate-payouts) are configurable on any stage in Kit.

## Does this produce better hires? Nobody knows yet

Be honest about the evidence, because the vendors selling into this space will not be.

Canva has published no pass rates, no candidate volumes, and no quality-of-hire comparison. What it reported at ATC2025 in November 2025, via Global TA Lead for Technology Joel Knudsen, is qualitative: richer discussions of tradeoffs and edge cases, the ability to pose more complex realistic scenarios, and the observation that candidates need both prompting skill and critical code-review ability. Useful, but not a controlled result. Meta and Google are running pilots. METR's trial had 16 developers in it.

So instrument your own change. Track stage-level pass rates for six months before and after, candidate drop-off at the stage you changed, and the correlation between that stage's score and 90-day manager ratings. If the AI-open stage does not separate candidates better than the round it replaced, you have learned something concrete instead of adopting a trend. That discipline is worth building given [how confidently wrong AI hiring advice can be](/blog/ai-hiring-advice-overconfidence-accuracy-gap), and how carefully teams need to draw the line on [AI autonomy in the interview itself](/blog/ai-agents-interview-candidates-trust-gap).

## Configuring an AI-open loop in Kit

AI policy is a property of a stage, not a property of your company, so it belongs in stage configuration. Kit models a pipeline as independent stages with independent config, which means Meta's split loop needs no new concept: an AI-open code assignment and an AI-closed live interview sit side by side in one pipeline.

Three things to set up when you [create the job posting](/docs/managing-jobs/01-creating-a-job-posting) and [build the pipeline](/docs/managing-jobs/02-managing-your-pipeline):

1. **Write the policy into the stage's candidate description.** Every stage type carries candidate-facing text, so the rule is read in the portal at the moment it applies, not buried in an email sent three weeks earlier. That is Canva's advance notice, implemented. The top-of-funnel summary belongs in your application disclosure.
2. **Put the AI contract in the assignment instructions.** For a code assignment stage, the instructions are committed into the candidate's private repository alongside a starter template. That is where you name the permitted tools and require the pull request paragraph about what the assistant got wrong. A deliberately imperfect starter repo turns the task into comprehension and correction.
3. **Give the AI-open stage its own weighted scorecard.** Drop in Decomposition, Tool restraint, Output validation, Correction, Production standards and Explanation as weighted criteria, and score them during [application review](/docs/working-with-candidates/02-reviewing-applications). An AI-open stage should not share a rubric with an AI-closed one.

Kit does not sell AI detection and has no AI-policy checkbox, because neither is needed. The policy is stage config, candidate-facing copy, and a rubric. Once the loop works, save it as a [process template](/templates) so the next role starts configured instead of rebuilt.

## Frequently asked questions

### Is it cheating to use AI in a coding interview?

Only if the stage said not to. Cheating is a violation of a stated rule, so the fix is stating the rule per stage rather than assuming a shared norm. Karat found 62% of organizations prohibit AI while leaders estimate more than half of candidates use it anyway, which is what happens when an unenforceable rule meets a tool everyone already has open.

### What companies allow AI in technical interviews?

Canva expects it in its AI-Assisted Coding competency for backend, ML and frontend roles. Meta pilots an AI-enabled round replacing one of two onsite coding interviews. Google is piloting an AI-assisted code comprehension round for junior and mid-level US roles using Gemini. Anthropic permits AI for application materials and for take-homes when instructed, but not in live interviews. Amazon prohibits it unless explicitly permitted.

### How do you test engineers when everyone has AI?

Change what the problem demands. Give an ambiguous brief with no single correct answer, start from an imperfect codebase instead of a blank one, require an explanation artifact about the process, and follow the AI-open stage with a short AI-closed conversation about the same code.

### Should take-home assignments allow AI?

Yes, in almost every case. Take-homes are unsupervised and the prohibition is unenforceable, so a ban converts the stage into a test of candidate honesty rather than skill. Allow it, raise the difficulty, and score verification and correction rather than output correctness.

### How do you tell candidates your AI policy?

On the stage it governs, in the candidate's view of that stage, before they start. Summarize the overall policy at application time, then restate the specific rule at each stage. Gartner recommends communicating how you define acceptable AI use as a fraud mitigation, not just as a courtesy.

Canva, Meta and Google did not decide whether AI belongs in hiring. They decided which round it belongs in, rewrote the problem accordingly, and told candidates first. That is a config change and an afternoon of rubric writing, not a philosophy. Pick the one stage where AI use most resembles the job, open it, make the problem too ambiguous for a single prompt, and score the collaboration. Then measure whether it separated candidates better than what it replaced.