# screening agent flagged every AI-assisted resume as fraudulent - the detector had no calibration for the role

## TL;DR

The AI-writing detector had no calibration, so it flagged every polished resume as fraudulent. Calibrate per role: measure the detector's false-positive rate on known-good resumes first, and treat the score as one signal, not a verdict. Uncalibrated detectors manufacture fraud.

```text
screening agent flagged every AI-assisted resume as fraudulent - detector uncalibrated
```

## Steps

1. Measure the false-positive rate: run the detector over hired employees' resumes.
   Expected: It flags the known-good set too.

2. Calibrate per role: set the flag threshold from the role's baseline, not a global default.
   Expected: The threshold reflects reality.

3. Downgrade the signal: an AI-assist flag triggers human review, never auto-reject.
   Expected: No one is rejected by a detector alone.

4. Re-screen the flagged candidates with human review.
   Expected: Wrongly flagged candidates recover.

5. Monitor the flag rate ongoing; detectors drift as writing tools evolve.
   Expected: Calibration stays current.

## When to use

- mass fraud flags on normal resumes
- the detector has no per-role calibration
- AI flags auto-reject

## When not to use

- the flags are accurate on verification (then the problem is real)
- the role legitimately bans AI assistance and states it (then enforce transparently)
- the detector is one of several signals already (check the weighting)

## Tool compatibility

- any AI-writing detector
- screening pipelines

## Variant phrasings

- "AI-assisted resume flagged fraudulent"
- "detector no calibration screening"
- "false positive AI writing resume"

## Why it happens

AI-writing detectors are probabilistic and uncalibrated out of the box; polished professional writing trips them constantly. Wired as an auto-reject, the detector becomes a randomness generator with consequences. The flag is a weak signal that needs a baseline before it means anything.

## Edge cases

- detector vendors overstate accuracy; measure it yourself on your own data
- disclosure policies ('did you use AI assistance') are a process alternative to detection
- the calibration set must reflect the actual applicant pool, not just employees

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_pRLoE3Y_IfzDiYIZVdByEg