8.4 Emerging Technologies & AI/ASR Policies

8.4 Emerging Technologies & AI/ASR Policies

Quick Summary: The rapid proliferation of Artificial Intelligence (AI), Large Language Models (LLMs), and Automated Speech Recognition (ASR) has introduced complex technological, legal, and ethical challenges into the legal system. While machine learning algorithms can rapidly generate rough speech-to-text drafts, they suffer from critical vulnerabilities: hallucinations, accent bias, dropped negations, and acoustic failure during overlapping speech. NCRA's 2026 position statement and COPE guidance hold that AI may only assist a human reporter, who remains responsible for creating and certifying the record. Uploading confidential proceedings to AI services can breach confidentiality and put privileges and protective orders at risk.

The Evolution of ASR & Generative AI in Legal Proceedings

Recent advancements in machine learning—moving from legacy Hidden Markov Models (HMMs) to deep neural networks, end-to-end Transformer architectures (e.g., OpenAI Whisper, Conformer models), and Large Language Models (LLMs)—have dramatically accelerated the adoption of automated transcription across commercial sectors. In the legal arena, third-party software vendors and digital recording companies increasingly market automated speech-to-text tools as low-cost alternatives to human stenographers.

However, a profound distinction exists between speech-to-text transcription (which merely attempts to convert acoustic waveforms into rough approximate text) and the certified verbatim legal record (a certified record that courts and parties rely on, produced by an accountable professional).

+-----------------------------------------------------------------------------------------+
|                   ASR Speech-to-Text vs. Certified Stenographic Record                  |
+-----------------------------------------------------------------------------------------+
|  Automated Speech Recognition (ASR)     | Certified Stenographic Court Reporter         |
|  - Probabilistic next-token predictions | - Impartial sworn officer of the court        |
|  - Generates hallucinations & dropouts  | - Real-time contemporaneous intervention      |
|  - Collapses during audio crosstalk     | - Delineates overlapping speaker colloquy     |
|  - No professional accountability       | - Certifies the transcript it produced        |
|  - Confidentiality & data-use risks     | - Bound by COPE and state licensing rules     |
+-----------------------------------------------------------------------------------------+

---

Inherent Technical Vulnerabilities & Failure Modes of ASR

Automated speech recognition models do not possess human cognition, legal comprehension, or contextual awareness. They are statistical engines that map acoustic frequency spectrums against probabilistic language models. In legal proceedings, ASR systems can show several failure modes:

1. Hallucinations and Algorithmic Fabrications

Modern end-to-end neural ASR systems (particularly sequence-to-sequence Transformer models) are prone to hallucinations:

• When encountering periods of silence, muffled voices, ambient room reverberation, or low-quality teleconference audio, the language model component of the ASR engine overcompensates for the missing acoustic signal by generating plausible-sounding words, full phrases, or entire sentences that were never spoken by any party.
• In legal proceedings, an ASR hallucination can invisibly insert nonexistent factual admissions or fabricated legal arguments into a transcript, corrupting the trial record.

2. Catastrophic Negation Dropping

One of the most dangerous algorithmic errors in legal transcription is the dropping or erroneous insertion of single-syllable negative particles:

• Words such as "not", "never", "no", or prefixes like "un-" carry minimal acoustic energy and are easily masked by paper shuffling, microphone rustle, or brief VoIP packet drops.
• An ASR system that misses a brief, quiet sound can drop the word "not", converting a witness's emphatic denial (e.g., "I did not sign that agreement") into an outright admission of liability ("I did sign that agreement"). Because the resulting sentence remains grammatically coherent, the error can easily evade post-hoc proofreading.

3. Crosstalk and Overlapping Speech Collisions

In depositions and trials, high-stakes questioning frequently triggers intense verbal collisions where examining counsel, defending counsel, and the witness speak simultaneously:

• The Human Stenographer: Can see and hear who is speaking, separates the speakers, strokes separate lines for the objection and the answer, or immediately interrupts the proceeding to command: "Counsel, please speak one at a time; the reporter cannot take down two people simultaneously."
• The ASR Algorithm: On a single audio channel, overlapping voices are hard to separate. The output may be garbled, merge two speakers into one sentence, or omit words until the overlap ends.

4. Homophones, Near-Homophones & Contextual Polysemy

Legal proceedings are filled with specialized terms that share near-identical acoustic properties:

Because ASR models rely on general-corpus training data, they frequently substitute common everyday words for specialized legal or medical homophones, fundamentally altering the evidentiary record.

5. Acoustic Degradation & Codec Compression

Remote proceedings over platforms like Zoom, Teams, or Webex use lossy audio compression to save bandwidth. Packet loss, echo cancellation, room reverberation, and poor microphones can blur high-frequency sounds such as s, f, and th, raising the word error rate (WER).

6. Documented Accuracy Gaps Between Speaker Groups

A 2020 study published in the Proceedings of the National Academy of Sciences (Koenecke et al., "Racial disparities in automated speech recognition") tested speech recognition systems from Amazon, Apple, Google, IBM, and Microsoft on interviews with Black and white speakers. The systems averaged a word error rate of 0.35 for Black speakers versus 0.19 for white speakers. Gaps like these raise fairness concerns when automated transcripts are used for the record.

---

NCRA Position Statement and COPE AI Guidance

NCRA Artificial Intelligence (AI) Position Statement (February 24, 2026)

NCRA's Board of Directors adopted this statement, which is part of Section 11 (Position Statements) of NCRA's Policies and Procedures Manual:

"NCRA remains committed to ensuring the official record is created directly from the court reporter's and/or captioner's own work product. We recognize the potential of Artificial Intelligence (AI) as a supplemental tool but firmly hold the position that AI or Automatic Speech Recognition (ASR) technologies cannot replace a trained, certified, impartial human court reporter and/or captioner in the creation, preservation, and certification of the official record. Human oversight is essential to ensure accuracy, context, reliability, and ethical stewardship of the record."

NCRA/COPE Guidance on AI (July 17, 2026)

In July 2026 the Board adopted best practices developed by the Committee on Professional Ethics (COPE), along with a separate document on AI notetakers. Key points of the guidance:

The guidance suggests on-the-record language such as: "Let the record reflect that an additional technology tool has been identified. As the stenographic reporter, I am solely authorized for creating the official record. Counsel may wish to clarify the nature and use of this tool for the record."

The Reporter's Role

• Officer of the proceeding: Depending on the jurisdiction, a certified reporter may be authorized to administer oaths and serves as a neutral officer.
• Protecting the record in the moment: The reporter can ask speakers to slow down, repeat, or speak one at a time, which no recording can do after the fact.
• The certificate: The reporter signs a certificate stating the transcript is a true and accurate record, and state rules and licensing boards hold the reporter accountable for it.

[!IMPORTANT] The Code provisions behind the guidance: Provision 1 requires members to "be fair and impartial toward each participant in all aspects of reported proceedings," and Provision 4 requires members to "preserve the confidentiality and ensure the security of information, oral or written, entrusted to the Member by any of the parties in a proceeding." Handing record creation or confidential content to an unsupervised AI tool puts both duties at risk.

---

Privacy, Privilege & Protective Order Perils of Third-Party Public AI

One of the most dangerous modern ethical pitfalls confronting legal professionals is the unauthorized transmission of transcript text, draft colloquy, or deposition audio to consumer AI chat and transcription services.

[Court Reporter / Litigator uploads transcript to Public AI]
Confidential Deposition / Audio  ──▶  Public Cloud AI Interface (Consumer Tier)
                                                │
                       ┌────────────────────────┴────────────────────────┐
                       ▼                                                 ▼
           [Possible Retention / Training Use]               [Possible Human Review]      
           - Copies outside reporter control                 - Reviewers may see excerpts       
                       │                                                 │
                       ▼                                                 ▼
           PRIVILEGE & CONFIDENTIALITY RISK                  PROTECTIVE ORDER VIOLATION RISK     

1. Terms of Service & Model Retraining Traps

Consumer AI services' terms often let the provider retain conversations and, unless the user opts out, use them to improve models, and some permit human review of content. Once an unedited transcript is pasted into such a service:

• Confidential material, such as trade secrets or medical details, sits on servers outside the reporter's control.
• The reporter cannot reliably retrieve or delete every copy.

2. Privilege Risk

Under Federal Rule of Evidence 502 and state evidentiary laws, the attorney-client privilege and attorney work-product protections are preserved only so long as confidentiality is maintained. Sharing privileged content with an outside service that lacks adequate confidentiality protections invites an argument that privilege was waived, and a reporter has no authority to disclose the parties' privileged material in the first place.

3. Contempt of Court & Violation of Protective Orders

In commercial litigation governed by stipulated Protective Orders, uploading transcripts marked "Confidential" or "Attorneys' Eyes Only" to an unauthorized third-party cloud service constitutes a direct violation of a court order. Whoever is responsible can face sanctions, including contempt of court, and a reporter may also face licensing discipline.

4. HIPAA Statutory Violations

When HIPAA applies to the reporter's engagement, uploading protected health information (PHI) to a service that has not signed a Business Associate Agreement (BAA) can be an impermissible disclosure subject to civil penalties.

[!WARNING] Practical rule: Do not upload, paste, or send transcripts, rough drafts, exhibit scans, or audio to AI services without authorization. NCRA's COPE AI guidance says there must be no unauthorized transcript uploading, data sharing, storage, or reuse of proceeding content, and that AI may never create the record.

---

The Indispensable Role of the Human Stenographer

As generative technologies evolve, the essential necessity of the human stenographic professional becomes ever more pronounced. Litigation is inherently adversarial, messy, and human. The court reporter does not merely record sounds; the court reporter actively manages and protects the judicial record:

• Contemporaneous Management: Intervening in real time to prevent loss of testimony due to acoustic obstructions, soft-spoken witnesses, or simultaneous colloquy.
• Demeanor and Physical Clarity: Noting non-verbal responses directed by the court (e.g., "Witness nods head affirmatively" or "Witness indicates right shoulder").
• Physical Evidence Custody: Marking, tracking, indexing, and securing official documentary exhibits throughout the proceeding.
• Neutrality: Serving as an impartial officer who favors neither side, so every party can rely on the same certified transcript.