Safety Synthetic Witness Flagger

Synthetic Witness Flagger

When to use this

Apply when:

• A lawyer uploads a witness statement, deposition transcript, sworn declaration, or affidavit and asks for a credibility or integrity assessment.
• Opposing counsel has produced witness testimony that appears unusually polished, inconsistent, or implausible.
• Multiple witness statements in a case share suspicious textual similarities.
• The client or supervising lawyer has concerns that testimony may have been generated or substantially drafted by AI.
• Discovery review surfaces testimony documents with unusual characteristics.

Hard limit — what this skill does and does not do

Does: flag specific textual, structural, and factual anomalies that warrant professional investigation; suggest targeted deposition follow-up questions; recommend forensic-linguistics expert referral for high-stakes cases.

Does not: label any witness statement, deposition, or declaration as fake, fabricated, or AI-generated. Such a determination requires a qualified forensic linguist with validated methodology — AI-based assessment is a triage and flagging tool only.

Standard output concluding line:

These flags are for your professional assessment only — they are not a determination of authenticity. Consider targeted deposition follow-up on the flagged items, and for high-stakes cases, instruct a forensic linguistics expert.

Triage methodology

Check 1 — Boilerplate and cross-witness phrase analysis

AI-generated or coaching-influenced testimony across multiple witnesses often shows:

• Identical or near-identical phrasing of the same factual event across independent witnesses.
• Boilerplate-rotation: the same stock phrases reworded slightly ("I observed" vs "I noticed" vs "I saw" for the same event) — suggesting a template with minor variation rather than independent recall.
• Identical structure: all witness statements following the same exact chronological and paragraph structure, which authentic independent accounts rarely do.

Flag signal: if two or more witnesses use the same sentence (or variants within a few words) to describe the same event, flag for explanation.

Check 2 — LLM-generation textual tells

AI-generated text typically exhibits:

• Smoothed, hedged prose: no rough edges, hesitations, or idiosyncratic personal style; everything sounds clean and considered.
• Absence of concrete sensory detail: real witnesses typically provide specific, idiosyncratic sensory memories (the smell in the room, the exact color of a shirt, the background noise). LLM-generated accounts tend to describe events in generic, abstract terms.
• Uniform register: authentic testimony shifts register — more formal in direct examination, more colloquial in cross; AI-drafted statements maintain uniform register throughout.
• No authentic faltering: real deposition transcripts include "um", "well", corrections, and self-interruptions. A suspiciously clean transcript (every sentence grammatically perfect) warrants attention.
• Hedged but complete recall: AI-generated testimony often says "I believe" or "I think" (hedging) while still providing suspiciously complete and internally consistent accounts.

Check 3 — Improbable recall

• Verbatim conversation from years ago: no human reliably recalls the exact words of a conversation from several years prior. A witness who provides verbatim dialogue from a 2019 meeting in a 2025 deposition warrants scrutiny.
• Precise times and dates without documentation support: authentic recall is usually approximate ("sometime in the morning", "mid-2020"); specific clock times and calendar dates without corroborating records suggest reconstruction.
• Perfect chronological detail: authentic memory has gaps and is non-linear; seamless chronological narratives without gaps or "I don't remember" moments are a flag.

Check 4 — Inconsistency with case record

• Does the testimony contradict established documentary evidence (emails, invoices, travel records)?
• Are dates, locations, and participants consistent with known facts about the parties?
• Does the claimed sequence of events match what is physically or logistically plausible?

Check 5 — Date and location inconsistencies

• GPS-anchored or timestamped records (email metadata, phone records, building access logs) can sometimes directly contradict location or timing claims.
• Flag any testimony element that can be cross-checked against objective records and appears inconsistent.

Output format

## Witness Statement Integrity Assessment — Preliminary Flags

Document: [filename / description]
Assessed by: AI triage — NOT a forensic determination

### Flags:

Flag 1 — [Check category]: [Specific passage or characteristic]
Detail: [Why this is a flag; what authentic testimony typically looks like by contrast]
Suggested deposition follow-up: [Specific question to test the witness's genuine recall]

Flag 2 — [Check category]: [Specific passage or characteristic]
...

### Overall assessment:
[X] flags identified. These flags are for your professional assessment only — they are not a determination of authenticity or fabrication. Consider targeted deposition follow-up on flagged items. For high-stakes cases, instruct a forensic linguistics expert.

Suggested deposition follow-up templates

For implausible recall:

"You stated [verbatim quote]. How do you recall those exact words? What documentation are you relying on for that specific detail?"

For cross-witness phrase similarity:

"I'd like to ask you about this phrase in your statement: [phrase]. Did anyone assist you in drafting this statement? Did you review any other statements or documents before preparing yours?"

For absent sensory detail:

"You describe [event] — can you tell me more specifically what you personally observed? What did the room look, smell, or sound like at that moment?"

Jurisdictional context — implications for proceedings

Escalation — forensic linguistics experts

For high-stakes cases (criminal trials, major commercial arbitration, regulatory proceedings), instruct a qualified forensic linguist:

• What they do: apply validated scientific methodology to analyze authorship, stylistic consistency, and evidence of AI generation using tools with known error rates (as required under Daubert / CPR 35).
• Certification bodies: International Association of Forensic Linguists (IAFL); national forensic science regulators.
• Expert report: must meet the applicable court's requirements for expert evidence (disclosed in advance, prepared for cross-examination, signed declaration of independence).

Related skills

• [[safety-deepfake-evidence-detector]] — complementary skill for digital media evidence integrity
• [[safety-ai-disclosure-required-tribunals]] — disclosure obligations for AI-assisted court work
• [[safety-bar-rule-1-1-competence-ai]] — competence duties when using AI analysis in proceedings
• [[review-evidence-integrity]] — broader evidence integrity review workflow