Prep-site data shows 438 users accepted AI-written interview code 93% of the time

A prep-platform operator has published usage data suggesting that candidates rehearsing for AI-assisted coding interviews rarely push back on what the assistant produces. In a post on dev.to, the founder of nyrion.ai, a site built to simulate this interview format, reports that across 438 people practicing on the platform, decisions about AI-proposed code went in the assistant's favor 93% of the time, and most users could not fully explain the code they had just submitted.

Ten companies, one new round

According to the post, ten employers now include an AI assistant as part of their coding interview: Canva, which the author dates to June 2025, followed by Shopify, Meta, DoorDash, Amazon, Atlassian, Stripe, Coinbase, Google and Instacart. Because nyrion.ai replicates that round, its session logs offer a rare look at how candidates actually behave when an assistant sits in the editor alongside them.

Nearly every proposal gets through

Users made 1,659 decisions to accept or reject code the assistant suggested. The remaining 1,546 of those, or 93%, were acceptances. The pattern held for heavy users as well: of the 109 people who judged five or more proposals, 62 turned down none at all. The author writes that they had expected more resistance and found essentially none.

Tests pass, explanations falter

Submissions also fared better when the assistant had contributed. When the AI proposed code before a candidate's first submission, 87% of those first attempts cleared the complete test suite for the stage, against 53% when no proposal had been made. The gap appeared on first, second and later stages alike, but the author is careful to label it a correlation that establishes nothing about why it exists.

The explanation step is where sessions fell apart. After tests passed, users were asked to describe the code they had submitted. Of 497 written responses, only 9% were complete, 79% were partial and 12% missed the mark. Three quarters of the partial answers conveyed the general idea while leaving out a point the author treats as essential. Users also skipped 29% of open-ended questions, compared with 5% of multiple-choice ones, which fits the author's broader conclusion that people rehearse the code generation and dodge the verification and articulation.

Six assistants you do not get to choose

The post also maps how the ten companies configure the round. Amazon, Stripe, Instacart, Meta, Atlassian and Google supply their own assistant inside their internal tooling, while Canva, DoorDash and Shopify let candidates use one they bring themselves. For Coinbase, the available reports do not say. The author argues the distinction matters for preparation, since an assistant you have never used behaves differently from the one already living in your editor.

Caveats worth noting

These were practice sessions, not real interviews, and nyrion.ai's assistant follows rules of its own, so the figures describe how people rehearse rather than how any specific company's round unfolds. Every number is an aggregate covering at least 20 people, the author's own and test accounts were excluded, and a full methodology page with sample sizes and time windows is published separately on the platform. There is also a clear commercial interest at play: the site selling preparation for this format is the one reporting how underprepared people are.

Why it matters

If this interview style becomes standard, and ten major employers suggests it is heading that way, the skill under test is no longer writing code but reviewing and defending it. The data indicates most candidates practice the opposite: they let the assistant write, watch the tests pass and move on. Interviewers will almost certainly probe exactly the step candidates skip, because verification and explanation are the only parts of the loop the human still owns. The numbers also echo a wider workplace question about AI code generation, since a 93% acceptance rate in practice looks much like the reflexive way many developers already accept autocomplete suggestions in daily work. Until candidates treat an assistant's output as a claim to be checked rather than an answer to be submitted, this round will keep failing the very people it was designed to evaluate.