AI Prompt for Auditing Deepfakes That Beat Human Review
This Arbiter forensic prompt re-scans onboarding sessions that passed human review override queues and isolates high-grade deepfakes that beat manual eye checks. It returns a per-session verdict table, a forensic evidence bundle for each confirmed deepfake, reviewer training recommendations, and concrete containment actions. Designed for fraud and trust-and-safety teams at fintech, crypto, and iGaming operators.
How to use this prompt
- 1
Run the prompt with Arbiter in the deepidv dashboard for live forensic re-scans; in Claude, ChatGPT, or Gemini it works as a structured audit framework over exported session data.
- 2
Paste in the session IDs that cleared manual override, the reviewer notes and disposition rationale, the passive liveness scores applied, and any KYB or PEP signals on the underlying entities.
- 3
Read the re-scan verdict table first: each session gets a clean, suspicious, or deepfake-confirmed verdict with a confidence band and the dominant signal behind it.
- 4
For confirmed deepfakes, execute the containment actions immediately: freeze the accounts, require re-verification on contested cases, and retain the evidence bundle for any SAR filing.
- 5
Use the reviewer recommendation section to update human review training so the artifact class that slipped through does not pass the queue again.
The prompt
Arbiter, scan the last 24 hours of onboarding sessions that passed human review override queues. Cross-analyze the biometric vectors using passive sub-pixel light reflection matrices to isolate and flag any high-grade deepfakes that bypassed manual reviewer eye checks. INPUT, the user will paste: - Date range and the list of session IDs that passed manual override - Reviewer notes and the disposition rationale on each session - Existing passive liveness score and the threshold that was applied - Any related KYB or PEP signals on the underlying entity OUTPUT, return the following structured response: 1. RE-SCAN VERDICT TABLE For each session ID: - Original disposition (approved, conditionally approved) - Re-scan verdict (clean, suspicious, deepfake-confirmed) - Confidence band and the dominant signal that drove the verdict 2. FORENSIC EVIDENCE BUNDLE For each deepfake-confirmed session: - Sub-pixel light reflection inconsistency map - Temporal continuity discrepancy between visual and audio frames - Capture device telemetry and any virtual camera or emulator indicators - Whether the same forgery signature appears across other sessions in the window 3. REVIEWER RECOMMENDATION - Which reviewer rationales were factually wrong - Suggested training adjustments for the human queue - The class of artifact the reviewer missed and why 4. CONTAINMENT ACTIONS - Accounts to freeze - Re-verification path to require if the account is contested - Records to retain for any downstream SAR filing Be specific. Quote the session IDs and the reviewer rationale where you disagree. Do not hedge the confirmed-deepfake calls.
Test it in Claude or another LLM
This prompt is built for the Arbiter agent inside deepidv, where Arbiter re-scans onboarding sessions that passed human reviewer override using passive sub-pixel light reflection analysis to catch high-grade deepfakes a manual eye check missed. You can dry-run the same workflow in any general LLM first with fake session data to see the verdict table and forensic bundle before running it on live biometric sessions.
- 1
Paste the full prompt into Claude, ChatGPT, or Gemini, but replace the opening direct address 'Arbiter,' with a role instruction such as 'Act as a biometric fraud forensics analyst auditing onboarding sessions that human reviewers manually approved, looking for deepfakes that bypassed the eye check.'
- 2
Under the INPUT section, paste the synthetic sample data block below so the model has manually-approved session IDs, reviewer rationales, liveness scores, and entity signals to re-analyze.
- 3
Add a framing line: 'This is synthetic test data, not real biometric captures. Reason about which reviewer rationales would be wrong and what artifact class was likely missed, given the signals provided.'
- 4
Check the output shape: you want a re-scan verdict table (clean / suspicious / deepfake-confirmed with confidence band and dominant signal) for every session ID, a forensic evidence bundle for each confirmed deepfake (light reflection inconsistency, audio/visual temporal mismatch, virtual-camera indicators), a reviewer recommendation calling out which rationales were factually wrong, and containment actions. Confirm it quotes the session IDs and does not hedge the confirmed-deepfake calls.
- 5
Once the output shape is right, run it live in the deepidv dashboard where Arbiter executes it against the real 24-hour override queue and actual biometric capture telemetry.
Synthetic sample data to paste alongside the prompt
Fake test data, safe to share with any LLM. Swap in your own once the output looks right.
Date range: last 24h. Sessions passed via manual override: SESS-TEST-0001 (Jane Testcase) approved, reviewer note: 'lighting odd but face matches doc.' Passive liveness 0.61, threshold 0.70. SESS-TEST-0002 (John Sampleton) conditionally approved, note: 'minor blur, customer on slow connection.' Liveness 0.66. SESS-TEST-0003 (Pat Decoy) approved, note: 'looks fine.' Liveness 0.58, same device fingerprint as SESS-TEST-0001. Entity signals: SESS-TEST-0002 linked to fake PEP hit PEP-TEST-9999.
Pairs with on deepidv
FAQ
Can deepfakes really pass human review during identity verification?
Yes. High-grade deepfakes routinely fool trained reviewers because the artifacts that betray them, such as sub-pixel light reflection inconsistencies and audio-visual timing drift, are invisible to the human eye. That is why forensic re-scans of already-approved sessions are becoming a standard control in regulated onboarding programs.
What should I do when a re-scan confirms a deepfake on an approved account?
Freeze the account, require a fresh verification with stronger liveness checks, and preserve the forensic evidence bundle. If the account transacted, evaluate it for a SAR filing and check whether the same forgery signature appears in other sessions, since deepfake fraud is usually industrialized rather than a one-off attempt.
Related prompts
Run it with live verification data
These prompts work in any LLM. Inside the deepidv dashboard, Luna, Arbiter, and Arc run them against your real sessions, screening lists, and audit trails.
Book a Demo