48% of Technical Candidates Cheat. Your Interview Loop Is the Wrong Fix.
Fabric's 19,368-interview study shows 48% of technical candidates cheat and 61% pass anyway. Sourcing verified engineers is the only fix that scales.
Fabric's analysis of 19,368 AI interviews landed with a number nobody in engineering hiring wants to look at directly: 48% of technical candidates are cheating, and 61% of them clear the pass threshold. If your loop still ends with a live coding round graded on a rubric, you are running a filter that half the pipeline routes around. The Blind thread "SWE interviews in 2026 are so stupid" hit 18k this week because everyone already knows.
The fix isn't better proctoring. Detection is a depreciating asset. The fix is upstream: source engineers whose work is externally verifiable before they ever hit your calendar.
The Fabric numbers are worse than the headline
Fabric's 19,368-interview study, run between July 2025 and January 2026 with a cheating threshold of probability > 40% and a pass threshold of score ≥ 7.0, found that 38.5% of all candidates cheated and 48% of technical-role candidates did. Sales roles came in at 12%. The 4x gap between coding and sales tells you exactly what tool the cheaters are using.
The trend line matters more than the average. In July 2025, the cheating rate sat around 9%. By September 2025, it had tripled to 45% and stayed there through January 2026. The step change lines up with the mainstream release of graphics-layer overlay tools. Sunday interviews now peak at 47.1%, because the people cheating are also the people willing to interview on a weekend.
Then the kill shot: 61% of flagged cheaters scored above the 7.0 pass threshold. Not "some got through." A clear majority. Your score is no longer a signal. It is a coin flip weighted against you, because the candidates who cheat also study which questions to expect.
Cluely made cheating a $20/month SaaS
Cluely (Roy Lee's April 2025 rebrand of Interview Coder, launched after his Columbia suspension for the viral Amazon-interview video) uses DirectX and Metal graphics-layer overlays that are invisible to Zoom, Teams, and Google Meet screen share. It reads the interviewer's question, pipes it to an LLM, and renders the answer on a layer the video capture pipeline literally cannot see.
The economics are the point:
- Cluely and its clones cost $20 to $50 per month.
- The target job pays roughly $150,000 per year.
- Payback on a single successful interview is under 4 hours of the first day.
- No install on a monitored machine is needed. A second phone works.
The graphics-hook trick is the load-bearing piece. Everything above the network layer (WebSocket relays, humanizer prompts, voice-clone overlays for phone screens) can be re-engineered in a weekend. Proctoring vendors ship updates in quarters. That is not a temporary gap. That is the paradigm losing.
Only about 11% of FAANG interviewers report that their company uses any detection software at all. 62% of hiring managers already admit that candidates are better at faking than recruiters are at catching. The arms race is over at most companies because most companies never showed up.
"Just let them use AI" is the trap that already sprung
Meta and Canva went the other direction and officially allow AI in interviews. The theory sounds clean: if the tool is legal, cheating disappears by definition. Brian Jenney (Parsity, 8.3k Medium followers) actually ran the experiment. His May 24, 2026 post, which is the piece re-circulating on LinkedIn and Blind this month, describes an "AI-allowed" engineering interview where Claude and Cursor were explicitly encouraged. Roughly half the candidates still cheated.
They weren't using the sanctioned tools. They were running overlay tools that generated whole solutions in one shot, and they could not explain the code they had just "written." Jenney caught them the moment he asked follow-ups.
The deception isn't about tool access. It's about faking the underlying skill, which no interview format catches after the fact. </pull> Allowing AI relabels the problem. It does not remove the incentive to fake competence you don't have. The candidates who cheat when it's banned are the same candidates who cheat when it's allowed, because the fraud was never about the tool. It was about the skill gap. ## Google and Amazon are going back to in-person. Most of you can't afford to. Sundar Pichai confirmed in a Google town hall that the company is reintroducing in-person interview rounds. Amazon is doing the same. Jane Street and Hudson River Trading never dropped in-person finals in the first place. Karat reports a 5x increase in cheating detection over the last two years and is leaning on live human proctoring. In-person works. It is also structurally unavailable to almost everyone reading this: - You do not have office space in the candidate's city. - You cannot fly a Series A pipeline of 40 engineers to HQ per month. - Your best candidates are remote by preference and will decline the flight. - Junior candidates skew hardest toward remote and cheat at roughly 2x the senior rate. In-person finals filter juniors out of the funnel entirely. If Google's answer is "spend $2,000 per onsite," and you have neither the office nor the budget, you need a different answer. The different answer is that you stop treating the interview as your primary filter. ## The score is now anti-signal for juniors For junior candidates (0 to 5 years of experience), a high interview score is now weakly correlated, possibly negatively correlated, with real ability. Juniors cheat at roughly double the senior rate, and 61% of cheaters pass. The candidate who aces your LeetCode-medium in eight minutes is more likely to be running Cluely than to be genuinely fast. The mechanism is simple. Real strong juniors show their thinking, ask clarifying questions, and take honest wrong turns before recovering. Overlay-driven candidates produce clean solutions with suspicious latency (roughly 4 seconds of overlay lag) and cannot answer "why did you pick this approach over a hash map?" The upstream fix is to filter on artifacts that predate the interview and cannot be generated in a 4-second window: 1. Public commits to named repositories over a period of 12+ months. 2. Merged PRs into projects the candidate did not create. 3. Conference talks with recorded video (RustConf, PyCon, KubeCon). 4. Named prior teams at companies with real technical bars. 5. Package downloads, GitHub stars on original work, or maintainer status. None of these can be faked in an interview window. All of them are queryable if you know where to look. This is the exact gap [Refolk](/) closes: you describe the engineer in plain English ("Rust systems engineers with 200+ commits to tokio or actix in the last 18 months, US-based") and get a ranked shortlist whose skill is already verified before the first message goes out. ## The 1,800:1 ratio nobody is pricing in In Refolk's index of professional profiles, roughly 19,500 US-based Software, Senior, and Staff Engineer profiles surface with an "Open Source" or "GitHub" skill signal. Only about 12 self-identify as "Open Source Maintainer" in their headline. The verifiable-shipper pool is roughly 0.06% of the engineering-skill pool. Meanwhile, Refolk's index shows roughly 21,600 US technical recruiters and sourcers currently in-role, concentrated at staffing shops (Experis, K2 Partnering) and hyperscalers (SpaceX, Blue Origin, Snowflake). Divide those numbers and you get a recruiter-to-verified-maintainer ratio of roughly 1,800 to 1. | Segment | Count / Rate | Source | |---|---|---| | Technical-role cheating rate | 48% | Fabric 19,368-interview study | | Sales-role cheating rate | 12% | Fabric 19,368-interview study | | Cheaters passing the 7.0 threshold | 61% | Fabric 19,368-interview study | | US SWE profiles with OSS/GitHub skill signal | ~19,500 | Refolk's index | | US SWE profiles claiming "open source maintainer" | ~12 | Refolk's index | | US technical recruiters currently in-role | ~21,600 | Refolk's index | | Recruiters per verified maintainer | ~1,800:1 | Refolk's index | | Technical vs sales cheating multiple | 4.0x | 48% ÷ 12% | | Score inflation, cheater pass-through | 1.59x | 61% ÷ 38.5% | The competitive read: most of those 21,600 recruiters still cannot read a GitHub contribution graph. They search LinkedIn for "Senior Software Engineer" and let the interview loop sort it out. The interview loop no longer sorts it out. The recruiters who can query the graph directly are pricing in a signal the market is systematically ignoring.
refolk prompt: Find US-based backend engineers with 300+ commits to Postgres, ClickHouse, or DuckDB in the last two years, currently at a company with fewer than 500 people. note: You get a ranked shortlist of engineers whose work is already externally verified, no take-home required. slug: 3m6wv4kq0q
## The Prisoner's Dilemma is why the rate keeps climbing
Fabric's own framing explains why 9% became 45% in ten weeks: honest candidates start cheating because they assume everyone else already is. This is a self-accelerating equilibrium and it does not stabilize on its own. It stabilizes only when the interview stops being the primary filter, which happens two ways:
- **The Google path.** In-person finals return industry-wide. Expensive. Slow. Filters out remote and international candidates. Available to companies with real estate and travel budgets.
- **The sourcing path.** The bar moves upstream. You source engineers whose skill is verifiable from public artifacts before the interview. The interview becomes a culture and comms check, not a skill assessment. Cheap. Scales. Available to any team that can query a contributor graph.
Option two is what most of you can actually execute this quarter. It is also what tools like [Refolk](/) exist to make routine: ask in plain English, get engineers whose GitHub graph is their resume, and stop paying the Cluely tax.
What to actually change in your loop this quarter
Rebuild the funnel around verification-before-conversation, in this order:
- Kill the initial LeetCode screen for anyone with 200+ commits to a named public repo in the last 18 months. The commits are the screen.
- Move the take-home to a 30-minute walk-through of code the candidate has already shipped in public. Ask them to explain a PR they merged six months ago. Overlay tools cannot fake this.
- Rank inbound applications by external artifacts, not by resume claims. GitHub graph, package downloads, conference talks, named prior team.
- Use in-person only for finals, and only for the 15% of the pipeline that already cleared verified-artifact screening.
- Redirect the proctoring budget to outbound sourcing. A Cluely-resistant hire starts with a search query, not a Chrome extension.
The 21,600 recruiters chasing a pool of 12 headline maintainers is the market inefficiency. Refolk is where you exploit it: describe the engineer, get the shortlist, verify the artifacts, and skip the theater.
FAQ
Does allowing AI in interviews actually solve the cheating problem?
No. Brian Jenney's May 2026 experiment with an "AI-allowed" interview loop, where Claude and Cursor were explicitly encouraged, still caught roughly half of candidates cheating with overlay tools they could not explain. Allowing sanctioned AI relabels the problem without removing the incentive to fake underlying skill. The candidates who cheat when it is banned are the same candidates who cheat when it is allowed.
How is Cluely different from a regular ChatGPT tab?
Cluely uses DirectX on Windows and Metal on Mac to render the AI's answer on a graphics layer above the video capture pipeline, so Zoom, Teams, and Meet screen share cannot see it. A regular ChatGPT tab is visible to any screen share. Cluely costs $20 to $50 per month against a target salary near $150,000, so the payback is a few hours of the first day on the job. That economic asymmetry is why the tool tripled cheating rates between July and September 2025.
If proctoring is losing, why are Karat and CodeSignal still growing?
Karat reports a 5x increase in cheating detection over two years, which measures the arms race, not who is winning it. Aiseptor's synthesis is explicit that detection-after-the-fact is a structurally losing paradigm, because every layer above the network hop can be re-engineered faster than proctoring vendors ship updates. The vendors will keep growing because most companies still buy them, but the leverage has moved upstream to sourcing.
What does "verified GitHub sourcing" mean in practice?
It means filtering candidates on public artifacts before the first interview: 200+ commits to a named repository over 12+ months, merged PRs into projects the candidate did not create, conference talks with recorded video, and named prior teams at companies with real technical bars. These signals cannot be generated in a 4-second overlay lag. Refolk indexes those signals across GitHub, LinkedIn, and the open web, so you can describe the engineer in plain English and get a ranked, pre-verified shortlist instead of a cheating-adjusted score.