The Pause That Gives Them Away

Hey everyone, Neil here. You're reading High-Signal Hiring. Hiring systems from 20+ years of global recruitment experience and 500+ technical hires. Zero noise and instantly actionable.

It's been a few weeks since the last issue. For the fourth time we’ve (me, my wife, 3 kids and our dog) decided to move countries, so the past month has been a little crazy to say the least! We’re now back in my homeland of the UK (Essex to be precise!) and looking forward to the next chapter (More on this to follow!). Thanks for sticking around; normal service resumes now.

For anyone who missed Issue #28, I argued the hiring bar has moved from writing code to owning it. Hire the engineer who can answer for the output, not the one who generates it fastest.

This week sits one level up from that. Before you can judge whether a candidate owns their work, you need an interview that still tells you the truth about them. For a lot of founders, it no longer does. Most of the signals you're leaning on can now be fed to a model in real time, and the candidate reading an answer off a second screen looks identical to the one who knows it cold.

You'll learn which parts of your process have stopped telling you anything, why the live follow-up is the one signal that survives, and how to rebuild the loop around it this week.

Not a subscriber yet? Sign up here

The formats founders lean on hardest are the ones that broke first.

Fabric's 2026 analysis of nearly 20,000 technical interviews found more than a third of candidates showing clear cheating behaviour, with adoption more than doubling from 15% in June 2025 to 35% by December. The tools doing this aren't fringe. Cluely, Interview Coder and Final Round AI sit as invisible overlays on the candidate's screen, listen to your voice through the system audio, and render a suggested answer the candidate reads while looking straight down the webcam.

Take-homes took the hardest hit. A take-home built to fill three hours now gets finished in eight minutes, so what you're measuring is whether the candidate pays twenty dollars a month for a subscription. One-way recorded video is worse, because the candidate gets infinite retries and a transcript feeding an answer with nobody watching. Even a scripted live round, where you ask the same set questions in the same order, is easy to game. The tool has heard the question before you finish saying it.

Anything a candidate can anticipate, they can now hand to a model. That covers most of a standard loop.

Every one of these tools has the same weakness, and it's latency.

When a candidate leans on an overlay, the flow isn't instant. Your question has to be transcribed, sent to a model, answered, and read back. Fabric clocks that round trip at one to four seconds, every time, and it shows up as a small, repeating pause before answers that should be reflexive. The prepared answer comes out smooth. The unscripted follow-up, the "why did you build it that way and not the other" that you invent from their last sentence, is where the lag appears.

That matches what twenty years of interviewing taught me long before any of this existed. The moment that ever told me the most about an engineer was never the rehearsed answer. It was the second and third question they didn't see coming. A model round trip just makes that gap easier to hear now.

The live follow-up holds up because it can't be prepared for and it can't be delegated fast enough. It's the last channel in your process running at human speed, which is why it's the only one still telling you the truth.

The instinct when you learn all this is to fight the tools. Proctoring software, eye-tracking, harder take-homes, a ban on AI in the room. The founders who go down that road lose. Detection sits one release behind the cheating tools, permanently, so you burn your energy policing instead of assessing. I wrote the small-team playbook on interview fraud back in Issue #20, and the update a few months on is simple. The gap between the cheating tools and the tools that catch them has only widened.

Stop trying to catch cheating. Build a process where it buys the candidate nothing, because the weight sits on the parts a tool can't help with. That needs no security budget. It needs you to stop grading the formats that are already compromised and put your judgement where the lag lives.

Most of the fix comes down to three moves, and none need new software.

First, move your evaluation weight off anything asynchronous. Cut the scored take-home and the one-way video. If you keep a take-home at all, treat it as a conversation starter rather than a gate, and assess it live by asking the candidate to extend it in front of you.

Second, stop running a fixed question list. Generate each question live from what the candidate just said. If they mention picking Postgres over Dynamo, your next question comes from that choice, not from row three of your sheet. A fixed script can be transcribed and pre-loaded before you finish asking it. An answer that branches off their own words can't be. This isn't about grilling them on trade-offs, I did that in Issue #28. It's about making the next question impossible to see coming, so the tool never gets its head start.

Third, go deep on something only they lived through. Pick a project from their history and branch into it live. What broke, what they'd redo, who disagreed with them and why. Nobody has a pre-generated answer for the specifics of their own project, and a model can't invent them convincingly on a one-second delay.

One caution, because this cuts both ways. A pause is not proof. Good engineers think before they speak, and the strongest candidates often go quiet because they're weighing a hard trade-off, not because they're reading a screen. The difference is in what comes after the pause. Honest thinking produces a messy, specific answer. The lag produces a clean, generic one that sounds like it was written down somewhere. Listen for which one you get, and don't punish someone for taking a breath.

The interview didn't get harder to run. It got easier to fake everywhere except the one place you improvise. Put your weight there, and the candidates who need a second screen go quiet on their ow

Cheers
Neil