How Hard Is It to Cheat with ChatGPT in Technical Interviews? Results from an Interviewing.io Experiment

ChatGPT has become a popular coding assistant, but a recent CoderPad report indicated that 23% of recruiters consider AI use in technical tests to be cheating, prompting concerns among developers.

To measure how difficult it is to cheat with ChatGPT, former Google engineer Michael Mroczka organized an experiment on the Interviewing.io platform, recruiting professional interviewers and candidates who were instructed to rely solely on ChatGPT during the interview.

The participants included interviewers from Interviewing.io’s pool (who were unaware that the study was about AI‑assisted cheating) and candidates who were actively job‑searching, had over four years of experience, rated themselves as moderately to highly familiar with using ChatGPT for coding, and claimed they could cheat without being caught.

A total of 37 interview rounds were conducted, of which 32 were valid. Interviewers asked three types of questions: (1) original LeetCode problems, (2) slightly modified LeetCode problems, and (3) custom interview questions designed by the interviewers. The pass rates for candidates using ChatGPT were 73% for original LeetCode questions, 67% for modified LeetCode questions, and only 25% for custom questions.

Remarkably, none of the interviewers suspected any candidate of cheating. Post‑interview surveys revealed that 72% of interviewers were confident in their hiring decisions, while 81% of candidates were not worried about being discovered; only 6% thought interviewers suspected them.

For reference, the baseline success rate of candidates on Interviewing.io’s regular mock interviews is 53%, and the platform’s overall success rate is about 57%, making the ChatGPT‑assisted results notably higher for standard LeetCode‑style questions.

LOG FILE START
NullPointerNinja: "who's going to the event tomorrow night?" - 100%
LambdaLancer: "wat?" - 5%
NullPointerNinja: "the event which is on 123 avenue!" - 100%
SyntaxSorcerer: "I'm coming! I'll bring chips!" - 80%
SyntaxSorcerer: "and something to drink!" - 80%
LambdaLancer: "I can't make it" - 25%
LambdaLancer: "🙁" - 25%
LambdaLancer: "I really wanted to come too!" - 25%
BitwiseBard: "I'll be there!" - 25%
CodeMystic: "me too and I'll brink some dip" - 75%
LOG FILE END

The study concludes that relying on off‑the‑shelf LeetCode questions makes it easy for candidates to leverage ChatGPT, and recommends that companies adopt custom, domain‑specific interview problems to both reduce cheating risk and better assess genuine engineering ability.