Fabric Flagged 48% of Technical Candidates. Beat the 4-Second Tell.

Fabric's October 5, 2026 Buyer's Guide turned "AI interview cheating" from vibes into measurement, and the numbers are brutal for anyone sitting a live technical screen this quarter. If you are an engineer with a strong accent, a fast typist who writes clean code, or someone who just happens to think in even three-second beats, you are in the dragnet whether you cheated or not.

This is a field guide to what the detectors actually watch, where they misfire, and how an honest candidate makes it through.

What Fabric's 19,368-interview study actually found

Fabric flagged 38.5% of candidates for cheating behavior across 19,368 AI interviews, with technical roles hit four times harder than sales. The headline that matters to engineers: 48% of technical candidates got flagged, versus 12% in sales roles. Junior candidates (0 to 5 years of experience) were flagged at nearly double the rate of senior engineers, and 61% of the flagged "cheaters" scored above pass thresholds, meaning they would have advanced if detection had been off.

Here is the breakdown of what Fabric caught, by method:

The important structural fact in that table: 82% of the signal weight is behavioral, not tab-switch monitoring. Fabric is not watching your browser. It is watching your eyes, your typing cadence, and your response latency. That is where honest candidates get caught in the net.

The 4-second tell, and why uniform latency is the real giveaway

The single most reliable AI interview cheating detection signal is not that candidates take too long to answer. It is that they take the same length of time to answer every question regardless of difficulty. The Cluely pipeline (audio capture, send to the LLM, process, render) needs three to five seconds on an easy question and three to five seconds on a hard one. A human does not do that.

An honest junior engineer thinks for 15 seconds on a hard graph problem and 2 seconds on "walk me through your resume." A Cluely user looks identical on both. That flat response curve is what the Fabric model calls out, and it is also what gets a nervous but honest candidate flagged if they default to a counting-to-three pause before every answer because an interview coach told them to.

What to actually do about it:

• Answer easy questions fast. If someone asks what language you prefer, say it in half a second.
• Think out loud on hard ones, visibly, for 10 to 30 seconds. Narrated thinking is both legitimate signaling and a latency alibi.
• Do not rehearse a uniform "let me think about that" preamble. That phrase, repeated at the same cadence on every question, is a flag by itself.
• If you need a long pause, say why: "I want to sketch the data structure before I answer." That is a human thing. A pipeline does not announce itself.

Why non-native English speakers get flagged for Cluely they never installed

The ugliest finding in the research is that detectors trained on written text misclassify non-native English writers as AI 61% of the time. Liang et al. at Stanford (2023, Patterns/Cell Press) ran seven commercial detectors against 91 TOEFL essays. The average false-positive rate was 61.22%. Eighty-nine of the 91 essays (97.8%) got flagged by at least one detector, and 18 (19.78%) were unanimously flagged by all seven. These essays were written by humans, by hand, with no AI involved.

The mechanism is simple and worth naming. Non-native English writers tend to produce text with:

• Lower lexical diversity (smaller working vocabulary deployed per paragraph)
• Simpler syntactic patterns (fewer nested clauses, less inversion)
• More predictable word choices (the common collocation over the unusual one)

Language models produce exactly those features, because they optimize toward statistically likely outputs. A detector trained to flag "predictable text" cannot distinguish a careful second-language writer from GPT-4o. Pangram, an academic detector, reports a 0% false-positive rate on the TOEFL benchmark, so the problem is solvable, but most enterprise detectors have not solved it yet.

What this means for the written screens that come before the live interview

Short, polished, templated answers (exactly what interview coaches teach) look like LLM output. If English is your second language and you are writing a take-home or a screening response:

• Include specific proper nouns the model would not invent: team names, internal tool names, your manager's first name, a city.
• Add numbers with awkward precision ($847K ARR, not "nearly a million").
• Leave one aside or opinion in. "I did not love this decision, but" is a very human phrase.
• Mix sentence lengths deliberately. One short sentence next to a long one is harder for a statistical detector to classify than three medium ones.

This is also the exact friction that writing your resume from scratch should take off your shoulders. Paste a posting into Refolk and the resume you get back is built from your actual history (named projects, real metrics, your own phrasing), not the templated prose that trips written-text detectors.

The 64% India problem hiding inside the 48% flag rate

Part of the 48% technical cheating rate is a measurement artifact caused by where engineers actually live. In Refolk's index of professional profiles, a combined query across US and India software and backend engineers returns 937,482 people. 599,731 of them, or 64%, are in India.

If any text-analysis component in Fabric's 20-plus signal stack behaves like the detectors Stanford tested, roughly 366,000 Indian engineers are structurally exposed to being flagged for language patterns alone. Fabric's own stated false-positive rate is 3 to 5%, which on 19,368 interviews is 580 to 970 wrongly-flagged people. That is almost certainly an undercount for the ESL subset.

Why clean code is now a liability

Keystroke dynamics catch AI-generated code because the LLM produces clean bursts with minimal backspaces, which is not how a human engineer types under pressure. Fabric's detector, Codility's June 2026 "Cheating Apps Detection" update (in preview for every account as of August 2026), and CoderPad's analog all weight this. The counterintuitive implication: do not clean up your scratch work during a live coding interview.

Signatures that read as human to a detector:

1. Write a function, delete half of it, rewrite from a different angle.
2. Misspell a variable name, backspace, correct it.
3. Add a // TODO comment then delete it when you solve the subproblem.
4. Change for to while and back.
5. Leave a console.log or print statement in while debugging, then remove it.

Signatures that look like Cluely output:

• A 40-line function typed top-to-bottom with no backspaces.
• Perfect variable naming on the first try.
• Comments appearing in order, before the code they describe, in grammatically perfect English.

If you type fast and clean because you are good, you are at a disadvantage against this signal. Deliberately verbalize your false starts. "I was going to use a hashmap here but let me try a set first" is both honest problem-solving and a legitimate way to make your editor history look like a person's.

The overlay arms race, in one paragraph

Cluely and Interview Coder work by exploiting the graphics layer: on Windows they use DirectX overlays, on macOS they use Metal framework layers, both rendering at a depth below what Zoom, Teams, and Meet capture during screen share. Cluely raised $5.3M seed and $15M Series A from Andreessen Horowitz after founders Roy Lee and Neel Shanmugam were suspended from Columbia. A fellow Columbia student, Patrick Shen, launched Truely as an "anti-Cluely." Codility named InterviewCoder and Cluely directly in its June 2026 release notes. Polygraf claims 98%+ detection against Cluely, Yoodli, InterviewCoder, and Pickle. The pattern: cheaters iterate weekly, detectors iterate monthly, and honest candidates are stuck being measured by whatever signals the detectors fall back on when the overlay fingerprints go stale. Those fallback signals (gaze, latency, keystroke bursts) are things nervous humans do too.

How to pass an AI interview without getting falsely flagged

The short answer to how to pass an AI interview is: look like a specific person, not a polished one. Detectors flag statistical averages. The further your behavior is from the average AI output, the safer you are, even if your behavior is a little weird.

A practical checklist:

• Vary your response latency by difficulty. Fast on easy questions, visibly slow on hard ones, with narrated thinking in between.
• Keep your eyes moving naturally. Look at your editor, your notes, the camera, in the pattern you normally would. Do not stare at a second monitor off to the side.
• Type like a human. Backspace, misspell, reorder. Do not optimize your editor history.
• Name specific things. Companies, teams, cities, tool versions, dates. LLMs hedge. Humans overshare.
• Do not read from a script. Reading cadence is uniform. Speaking cadence is not.

The honest version of the AI interviewer Fabric detects is a human with a weird cadence, specific memories, and messy typing. That is defensible. A clean, uniform, Cluely-paced answer is not, even if you typed it yourself.

FAQ

Does Fabric actually detect Cluely and Interview Coder by name?

Yes, and so does Codility as of its June 2026 update. Fabric's 20-plus signal stack targets dedicated overlay tools directly, and those account for 45% of the flagged cheating cases in the 19,368-interview study. Cluely's DirectX and Metal overlays are specifically named in Codility's release notes. Polygraf claims 98%+ detection on this class of tool. The arms race is real, but as of late 2026 the detectors are ahead on the overlays specifically.

I am a non-native English speaker. Am I going to get flagged no matter what?

Not on live video, but on written screens you are at structural risk. Stanford's 2023 study found a 61.22% average false-positive rate across seven commercial detectors on TOEFL essays. The fix on your end is to make your writing sound specifically like you: name real teams, real tools, real numbers, and leave in one or two opinions or asides an LLM would not invent. Pangram's 0% false-positive rate on the TOEFL benchmark suggests the next generation of detectors will be fairer, but you cannot wait for employers to upgrade.

Will using a tool like Refolk to write my resume get it flagged as AI?

Refolk writes your resume from your own history (your roles, your projects, your actual metrics), tailored to the specific posting, which is the opposite of the generic polished prose text detectors flag. The 61% false-positive problem comes from text that sounds generic and predictable. A resume full of named employers, dated projects, and specific numbers reads as human because it is specific to one person. The risk shape is "templated and bland," not "AI-assisted."

What is the single thing that gets honest candidates flagged most often?

Uniform response latency. If every answer takes you the same 4 or 5 seconds to start, you look like a Cluely pipeline whether you are one or not. The fix is deliberate variance: snap-fast on easy questions, visibly slow and narrated on hard ones. Pair that with messy typing (backspaces, false starts, reordered lines) and you have defeated the two biggest behavioral signals that AI-interview-flagged candidates fail on.