Candidates use AI in your interviews whether you allow it or not, and CEOs now expect AI use to appear in performance reviews. This module redesigns the hiring loop for both realities and translates AI mandates into performance expectations that are fair — measured by observed work quality, never by tool telemetry.
Divya, a senior EM at a Pune fintech, was running the debrief for a backend candidate who had aced the coding round — cleanest solution the panel had seen in months, finished with twenty minutes to spare. Then her newest interviewer, fresh from an AI-tooling workshop, said something uncomfortable: "I pasted our question into a chatbot last night. It produced essentially his solution, variable names and all, in one prompt."
The debrief unraveled from there. Had the candidate used AI through an earpiece or a second screen? Nobody knew — the loop had no stated policy, so nobody had asked, and there was nothing to enforce. Worse, the panel realized the question itself was the problem: it was a solved problem on the internet, which meant it had been measuring memorization for years and was now measuring prompt access. And worst of all: if the candidate had used AI skillfully, wasn't that... the job now? Their own company mandated Copilot. They were about to reject a candidate for doing what they would require of him on day one — or hire him without knowing whether he could function without it. They had no way to tell which.
Divya's loop failed on three counts that this module fixes: no problem design that survives AI, no published stance on candidate AI use, and no way to evaluate the skill that actually predicts success now. And the same confusion was waiting in her next perf cycle, where the CEO wanted "AI adoption reflected in reviews" and nobody knew what that meant fairly.
The instinct after an experience like Divya's is detection — proctoring, lockdown browsers, suspicion. Before committing to that road, look at where each posture actually leads.
Lockdown environments, webcam proctoring, gotcha questions about the candidate's own screen. Costs: an adversarial candidate experience, an arms race you lose to a second phone, and — the quiet failure — a loop that still measures a job that no longer exists. The candidates who pass are the ones best at working unassisted, which is no longer the work.
AI-allowed rounds that observe prompting, validation, and debugging of AI output; AI-free rounds that protect fundamentals; problems engineered to survive one prompt. Costs: real design effort up front. Returns: signal on the skill you are actually hiring for, and a candidate experience that markets the team instead of insulting it.
The industry's leading indicators have chosen the second road. Meta rolled out AI-enabled coding interviews for mid-to-senior engineers in late 2025, explicitly scoring how candidates work with an assistant. Google piloted a similar format in 2026, pairing an approved assistant with a code-comprehension round. Canva requires AI use in its technical interviews and designs problems intended to be unsolvable by a single prompt. The bet these companies share: since the job now includes AI, the interview should measure the job.
The redesigned loop has two round types, each with a distinct purpose:
The attack log matters beyond the loop: problems decay as models improve, so re-run the attack quarterly. Lab 2 has you do exactly this against a provided loop — including proving which of three problems falls to a single prompt.
Run every interview problem through the one-prompt attack before it meets a candidate, and keep the transcript. It takes ten minutes, it tells you what the problem actually measures now, and the log is your evidence when a hiring committee asks why the old favorite got retired.
Whatever you decide, write it down and publish it to candidates before the loop. A written stance beats interviewer improvisation on every axis: candidates prepare for the actual exam, interviewers enforce consistently, and covert-use incidents become policy questions instead of ad-hoc moral judgments. The stance needs only four sections: which tools are allowed in which rounds; disclosure expectations; what happens if undisclosed use is discovered; and why the AI-free rounds exist. Divya's debrief had no answer to "did he cheat?" because the loop had never defined what cheating was.
Upstream of the loop, the JD deserves the same honesty pass. AI-fluency expectations now appear verbatim in postings across every engineering family — "AI Native Tech Lead" is a live title — so write yours realistically: name the fluency you actually need (works effectively with coding assistants, verifies output, contributes to team prompts and standards) rather than pasting buzzword salad that attracts keyword-matchers and repels the skeptical seniors you want. AI drafts the posting well; have it also run a bias and cliché scan, and generate the India/remote variants where your hiring actually happens.
Interview evaluation gets AI help too, with a bright legal line. Scorecard synthesis across a panel — evidence-vs-rating gaps, halo and recency flags, one interviewer's consistent harshness — is genuinely useful debrief input. But AI advises and the panel decides, always and in writing: automated employment decision tools attract legal exposure under NYC Local Law 144-class regulation, and beyond compliance, an AI-synthesized ranking laundered into a hiring decision is exactly the kind of process no one can later defend. Same rule at the JD stage — drafting with realistic AI-fluency expectations and a bias scan is a solid Foundation use; auto-screening candidates is a different, regulated act.
Now the harder half. The CEO reads the Shopify memo — AI use is a baseline expectation, reflected in reviews — and forwards it to you. Between that sentence and a fair rubric lies the difference between an adoption program and a morale disaster, and the difference turns on one distinction.
Measure observed work quality. Never measure tool telemetry.
Copilot acceptance rates, prompt counts, seat activation, AI-generated line percentages — none of it measures whether someone works well with AI, all of it is trivially gameable, and using it in reviews converts your measurement system into a surveillance system, poisoning both (Module 6 returns to this). What you can assess fairly is what you already assess: the quality of work products and working behaviors, now including AI-era ones — sound judgment about when to use AI, output that has plainly been verified, learnings shared with the team.
| Rubric dimension | Observable behavior — what fair looks like | The telemetry trap to refuse |
|---|---|---|
| Judgment | Chooses AI for fitting tasks; knows when hand-crafting or a no-AI zone applies | "Uses Copilot daily" (seat-activation data) |
| Verification | AI-assisted PRs arrive reviewed, tested, explainable line-by-line | Acceptance-rate percentages |
| Leverage | Uses AI to expand scope or quality — better tests, docs, migration coverage | Lines of AI-generated code |
| Multiplication | Shares prompts, patterns, and failure stories that make others productive | Prompt counts, message volume |
Three edge cases will find you within one cycle, so rule on them in advance. The accommodation case: an engineer whose disability or workflow makes a specific tool a poor fit gets an equivalent-outcomes path, not an exemption from quality. The client-blocked case: engineers on accounts whose contracts prohibit AI tools (Module 5) cannot be penalized for compliance — their rubric weights the non-AI dimensions. The conscientious objector: skepticism is legitimate and often well-calibrated, but "I don't use AI" cannot become "I deliver less than the role requires" — the expectation is outcomes, with AI as the default path and the burden shifting to the objector to meet the bar another way. Lab 3 makes you write all three rulings against a planted HRBP draft that scores telemetry — and rejecting that draft in writing is the graded act.
The moment engineers believe AI metrics feed their ratings, they will optimize the metrics — acceptance rates and activity counts will soar, telling you nothing, while trust in both the review system and the measurement program quietly dies. Grandfather existing goals, announce the rubric before the cycle, and put the no-telemetry promise in writing.
Performance season is where leadership-ops AI (Module 2) meets its strictest privacy rules — enterprise-tier, no-training tools only, and boundaries beyond even those.
Redesign the loop instead of policing it — AI-allowed rounds that measure fluency, AI-free rounds that protect fundamentals, problems proven against the one-prompt attack, and a published candidate stance. Then hold performance to the same standard you hold hiring: judge observed work quality, never tool telemetry, with the edge cases ruled on in advance.