Texas Supreme Court and the Admissibility of AI Deposition Transcripts
In response to nationwide shortages of licensed stenographic certified shorthand reporters (CSRs), state judicial systems—headlined by the Texas Supreme Court and its specialized advisory task forces—have confronted an urgent operational question: To what extent can artificial intelligence-driven automated speech recognition (ASR) generate official deposition records for use in civil litigation?
While preliminary drafts produced by multi-channel neural ASR models (such as Whisper, Conformer, and specialized legal LLM fine-tunes) offer drastic turn-around speed and lower upfront delivery fees, trial attorneys and judges face immense evidentiary vulnerabilities. In adversarial proceedings, deposition transcripts are not mere informal meeting summaries—they constitute sworn testimonial instruments governed by strict statutory rules of evidence, subject to cross-examination, impeachment, and dispositive summary judgment motions.
A deposition on oral examination must be taken before an officer authorized to administer oaths. The officer must record the testimony of the witness by stenographic, audio, or audiovisual means. Under Rule 203.2, the deposition officer must certify on the transcript that the witness was duly sworn and that the transcript is a true record of the testimony given by the witness. AI algorithms are not licensed officers capable of administering oaths under Texas Government Code Chapter 52.
The Anatomy of ASR Failure in Adversarial Depositions
Comprehensive litigation testing across multi-party commercial, medical malpractice, and patent disputes reveals that automated speech models exhibit failure modes that differ fundamentally from human stenographic errors:
| Discrepancy Category | Automated AI Speech Model (ASR) | Certified Shorthand Reporter (CSR) | Evidentiary Consequence |
|---|---|---|---|
| Legal & Technical Jargon | Substitutes common phonemes (e.g., "res ipsa loquitur" → "race it so low quitter") | Stenomask / dictionary expansion with immediate clarification requests | Substantive misrepresentation of medical standards of care and proximate cause. |
| Simultaneous Cross-Talk & Objections | Suppresses the lower-amplitude voice or blends statements into a single speaker line | Halts proceedings or transcribes overlapping lines with parenthetical markers | Waiver of preserved objections under TRCP 199.5; failure to preserve error for appellate review. |
| Numeric & Financial Values | Misinterprets digit phrasing (e.g., "four hundred fifty thousand" parsed as "$45,000") | Cross-verifies numbers with marked deposition exhibits on the record | Catastrophic errors in damages valuations, settlement negotiations, and financial disclosure audits. |
| Witness Affirmations vs. Hesitations | Normalizes "uh-huh" (affirmative) and "uh-uh" (negative) inconsistently | Explicitly notes verbal confirmation or requests a spoken "yes" or "no" | Reversal of deponent admissions during trial impeachment. |
Admissibility Challenges Under Texas Rules of Evidence 901 and 1002
Opposing counsel seeking to exclude or strike an AI-drafted deposition transcript typically rely on two pillars of evidentiary doctrine:
- TRE 901 (Authenticating or Identifying Evidence): To satisfy the requirement of authenticating an item of evidence, the proponent must produce evidence sufficient to support a finding that the item is what the proponent claims it is. If an AI system generated the transcript without continuous human certification of the underlying audio recording, establishing who verified each disputed word becomes an impossible hurdle.
- TRE 1002 (Best Evidence Rule / Requirement of the Original): When the accuracy of a written transcript is contested, the original synchronized audio recording remains the best evidence of the witness’s statements. An uncertified AI text output lacks independent legal weight without the primary audio stream.
The Texas "Human-in-the-Loop" Verification Mandate
The prevailing trajectory endorsed by court reporting boards requires a certified human transcriber or digital court reporter to act as the ultimate reviewing editor. In this workflow, automated speech recognition serves strictly as an initial scoping tool. The human reporter listens to the multi-channel audio master, audits each acoustic collision, manually corrects speaker attribution during heated colloquy, and signs the final certification certificate.