The AI Arms Race in Technical Interviews Is Escalating
IEEE Spectrum Rina Diane Caballar
Job seekers are using AI to cheat on coding interviews, so companies are using AI to catch them. It's an arms race with no clean winner, and both sides are getting burned by false positives.
Based on reporting by IEEE Spectrum, Rina Diane Caballar — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Technical interviews used to be a fairly simple, if stressful, ritual: a whiteboard, a coding problem, a panel judging how you think. Now there's a third party in the room nobody invited. Tools like Final Round AI, Interview Coder, and ParakeetAI sit invisibly on a candidate's screen, listening to the interviewer's questions and feeding back answers or code in real time. All the candidate has to do, as Meta engineer Mudit Saraf puts it, is "put on a little performance."
Employers didn't take this lying down. Saraf and Microsoft engineer Shraddha Sunil built Ginger, an AI voice recruiter that runs first-round screening calls and watches for tells: eye movement, delayed responses, tab-switching, and speech patterns that read as AI-generated. It's a reasonable idea in theory. In practice, recruiter Archie Payne of CalTek Staffing says detection accuracy still isn't good enough, and he's seen strong candidates flagged as false positives — which is its own kind of disaster when qualified engineers are already hard to find.
The roots of this standoff aren't mysterious. AI-driven layoffs have flooded the market with applicants, and many companies now use AI resume screeners to filter them at scale before a human ever looks at an application. Candidates, feeling processed rather than evaluated, reach for their own AI as a countermeasure. AI hiring strategist Tatiana Teppoeva frames it as a system that rewards pattern-matching, so people game the pattern. Navy Federal's Ravi Kiran Pagidi worries the whole exercise risks becoming less about engineering skill and more about who's better at gaming an algorithm — on either side of the table.
There's a real cost buried in this, too. A Stanford Institute for Human-Centered AI study tracking 3.4 million applicants found that AI hiring tools from a single vendor produced adverse outcomes for Asian and Black candidates, which is exactly the kind of quiet, systemic harm that gets waved away as "efficiency." Teppoeva argues human oversight has to stay in the loop, and Pagidi says companies need audits and transparency or they'll end up thinking they've streamlined hiring while actually degrading their own signal.
Some companies are skipping the escalation entirely. Meta and the startup Factory now let candidates use AI openly during interviews, then judge them on process rather than output — how they direct the tool, debug its mistakes, and explain their reasoning, not whether the tests pass. Factory's Varin Nair says weak candidates hand over their thinking to the AI and freeze when it stumbles, while strong ones use it to move faster and spend the saved time on architecture and trade-offs. That distinction — judgment versus obedience — might be the only interview signal that still means something once everyone has an AI whispering in their ear.
Payne's advice to candidates is blunt: use AI to prepare, not to perform, because detection is improving and the engineering world is smaller than people assume. Getting caught gaming an interview follows you. In a hiring process already this adversarial, authenticity has become the one competitive advantage nobody can automate away yet.
My take — AI-written commentary, not fact-checked reporting
This is what happens when you let algorithms manage a fundamentally human decision on both sides of the table — you get an arms race that burns trust and produces false positives instead of better hires. The companies quietly winning here, like Meta and Factory, aren't the ones building better lie detectors; they're the ones who stopped pretending AI use is cheating and started grading judgment instead. Everyone else is going to spend the next two years detecting ghosts while the actual signal — can this person reason under pressure — gets buried under an arms race nobody asked for.
Read more about this at: IEEE Spectrum
