Here is the uncomfortable thing nobody on a TA leadership call wants to admit out loud: the candidate you just passed to onsite probably had ChatGPT open the entire time. Your scorecard is not measuring them. It's measuring GPT-5 with a webcam pointed at a human face. And the longer you pretend that's not happening, the worse your hiring manager satisfaction is going to get.
The Number Nobody Wants To Print
In informal pulse checks across the recruiters we work with, the consensus estimate is that the majority of remote technical screens in 2026 involve a candidate using a second device, a second window, or an in-browser assistant. The honest number is probably north of 70%. We're not going to print a fake study to make you feel better. You already know.
And the worst part is it's not the bad candidates doing it. It's everyone. Strong engineers do it because "everyone else does." Mediocre ones do it because it's the only way they get past you. The signal is gone.
Why Proctoring Doesn't Work
The first instinct is to bolt on more surveillance: lockdown browsers, eye-tracking, "no second monitor" attestations. None of it works in 2026. A candidate with an AirPod in one ear and a phone face-down on their lap will pass any proctoring stack you can buy. You're playing whack-a-mole with a $20/month consumer product that updates faster than your vendor.
Worse, proctoring is hostile to the exact candidates you most want to hire. A senior engineer with twelve years of experience will not consent to having their pupils tracked. They'll just take the other offer. They're already refusing the easy version of this.
“You can't proctor your way out of an arms race against a free chatbot. You have to change what the interview is.”
How AI-On-AI Detection Actually Works
The only defense that works in 2026 is putting a probing AI on your side of the table. A static question — "implement a binary search" — is trivial for an LLM. A live conversation that asks "okay, now add a constraint that the array can be rotated, and explain why your invariant still holds" is not. The probing breaks the script.
This is what Talia AI is built to do. Talia, by UpStack, runs a structured 15-minute technical screen with adaptive follow-ups. When the candidate's answer doesn't match their stated reasoning, Talia asks again, differently, in a way no pre-baked GPT prompt is ready for. The scorecard surfaces the gap explicitly: "Candidate's spoken reasoning did not match implementation choice on follow-up."
That's not anti-cheating. That's just real interviewing — at machine scale and machine consistency. We covered the full landscape here.
The Fix You Can Implement This Week
Stop running first rounds as static coding tasks on CoderPad or HackerRank. Run them as live AI conversations. Pick one open role, route the next 25 candidates through Talia, and compare the scorecards to your last 25 human-screened candidates. The pass rate will drop. That's the point. You weren't filtering before — you were waving people through to engineers who then had to do the actual screening for you.
The companies who fix this in Q2 2026 will quietly out-hire everyone else by Q4. The ones who keep blaming "candidate quality" for the dip in onsite-to-offer conversion are about to learn that the candidate quality was always fine — the screen was the problem.
The Screening Room is an UpStack publication. Visit talia.ai.