AI Errors Infiltrate Federal Court Transcripts

On July 24, 2026, the Fifth Circuit Court of Appeals issued a per curiam opinion in Scott v. Collier that contained an unusual addendum. Judges James Ho, Andrew Oldham, and Kyle Duncan did not just rule on the employment law question before them. They flagged something far more alarming: the official court transcript from the district court's bench trial appeared to contain errors generated by artificial intelligence. The court's language was blunt. "[W]e remind the Court Reporter that this court relies on transcripts being true and accurate representations of the transcribed proceedings." That single sentence carries enormous weight. The integrity of the American legal system depends on court records being exact reproductions of what happened in a courtroom. When they are not, justice itself becomes unreliable.

The Anatomy of AI-Generated Transcript Errors

The Fifth Circuit identified several distinct categories of AI contamination in the trial transcript. The first involved homophones, words that sound alike but carry different meanings. The transcript recorded that "Collier's testimony was 'incredible,' meaning 'not credible.'" In context, that substitution reversed the meaning entirely. An AI speech-to-text system, trained to predict likely next words, selected "incredible" when the speaker clearly said "not credible." A human court reporter would catch that distinction instantly.

The second category was tense errors. The transcript quoted Collier as saying he would "sign off" on overtime when the grammatically and contextually correct form was "signed off," the past tense. AI transcription systems often default to base verb forms because they optimize for probability rather than grammatical accuracy within specific conversational contexts.

The third and most disturbing category was outright hallucination. In one passage, an attorney's question was recorded as: "And did you ever sign off on Mr. Scott's overtime? A No." But no witness response appears in the record. The "A No" was fabricated, likely inserted by an AI system that predicted a response where none existed. This is the same hallucination problem that plagues large language models in every domain, from legal research to medical diagnosis. The AI filled a gap with plausible-sounding text that had no basis in reality.

A fourth pattern showed the AI "completing" answers based on the content of questions rather than transcribing actual testimony. When asked whether he "ever denied Mr. Scott overtime," Collier's recorded response, "I never denied him overtime," mirrors the question almost verbatim. Human witnesses rarely parrot questions back in such clean, syntactically parallel forms. This is a signature behavior of predictive text generation.

The Scale of the Problem

Court reporting in the United States has been under strain for years. The National Court Reporters Association has warned of a growing shortage in certified stenographers. A 2023 survey estimated roughly 5,200 official court reporters working in federal and state courts, down from over 8,000 two decades earlier. The average age of working court reporters continues to climb. Fewer students enter stenography programs each year, deterred by the two to four years of training required to achieve the 225-words-per-minute certification standard.

This shortage has created market pressure to adopt AI-powered alternatives. Companies offering automated transcription services have pitched their products to courts, law firms, and government agencies. The technology works well enough for casual use: meeting notes, podcast transcriptions, rough drafts. But "well enough" is a dangerous standard when applied to legal proceedings where a single misheard word can alter the outcome of a case.

The Fifth Circuit's opinion in Scott v. Collier did not identify which AI transcription tool was used or whether the court reporter employed it as a primary tool or a supplement. That ambiguity is itself part of the problem. There are no federal rules requiring court reporters to disclose their use of AI assistance. The Federal Rules of Appellate Procedure and the Judicial Conference guidelines address the duties of court reporters but were drafted in an era when the primary tools were stenotype machines and audio recordings reviewed by human ears.

Institutional Trust and Procedural Integrity

The legal system runs on trust in its own records. When an appellate court reviews a trial court's decision, it works almost entirely from the written transcript. Judges on appeal were not in the room. They did not see the witnesses, hear the tone of voice, or observe the body language. The transcript is their window into what happened. If that window is distorted by AI artifacts, the appellate process itself becomes unreliable.

Consider the practical consequences. A hallucinated "A No" inserted into an attorney's question could be mistaken for actual testimony. A homophone error that flips "not credible" to "incredible" could change how a reviewing court evaluates a witness's reliability. A fabricated answer that mirrors a question could create the false impression that a witness admitted something they never said.

These are not hypothetical risks. They materialized in an actual federal case. The Fifth Circuit caught them, but only because the errors were conspicuous enough to trigger suspicion. Subtler AI errors, small word substitutions that shift meaning without obvious absurdity, might pass undetected through the entire appellate process. No one knows how many already have.

The legal profession has begun grappling with AI in other contexts. In 2023, a New York attorney made national headlines when he submitted a brief containing fabricated case citations generated by ChatGPT. Judge P. Kevin Castel sanctioned the lawyer and his firm. Multiple federal courts have since adopted local rules requiring attorneys to disclose their use of AI in filings. But those rules target lawyers, not court reporters. The Scott v. Collier opinion exposes a gap: the system assumed the transcript itself was trustworthy while focusing its AI skepticism on the briefs that cite it.

The Bitcoin Parallel: Verifiable Records Matter

This case illustrates a principle that sits at the core of Bitcoin's design philosophy. Bitcoin exists because trusted intermediaries fail. The entire point of a decentralized, cryptographically verifiable ledger is to eliminate the need to trust a third party's record of what happened. Satoshi Nakamoto's whitepaper opens with the observation that "commerce on the Internet has come to rely almost exclusively on financial institutions serving as trusted third parties." The same critique applies to legal records.

A court transcript is a centralized, human-produced record with no built-in mechanism for verification. There is no hash function, no consensus algorithm, no way for the parties to independently confirm that the words on the page match the words spoken in the courtroom. The system relies entirely on the trustworthiness and competence of a single individual, the court reporter, and whatever tools that person chooses to use. When that individual quietly substitutes an AI system for their own ears and judgment, the trust assumption breaks down without anyone knowing it broke.

Sound-money advocates have long argued that systems built on trust in central authorities contain hidden fragility. The AI transcript problem is a vivid, non-financial example of exactly that fragility. A permissionless, timestamped, cryptographically secured record of events is not just a feature of better money. It is a feature of better record-keeping in every domain where accuracy matters and incentives to cut corners exist.

Regulatory and Professional Response

The Fifth Circuit's opinion stopped short of imposing sanctions on the court reporter. It issued a reminder, not a penalty. That restraint may reflect the novelty of the issue. But it also creates a precedent problem. If appellate courts merely remind court reporters to be accurate without attaching consequences, the incentive structure does not change. Court reporters facing heavy caseloads and tight deadlines will continue reaching for AI tools that reduce their workload, even when those tools introduce errors.

Several state bar associations and judicial conferences have begun studying the issue. The Judicial Conference of the United States, which sets policy for federal courts, has formed working groups on AI use in the judiciary. But their focus has been primarily on AI in judicial decision-making and attorney filings, not on the transcription pipeline.

The National Court Reporters Association has taken a firm public position against fully automated AI transcription, arguing that certified human reporters produce accuracy rates above 99 percent while current AI systems fall below 95 percent in noisy, multi-speaker courtroom environments. That 4-plus percentage point gap translates to dozens of errors per hour of testimony, any one of which could carry legal significance.

Some jurisdictions have moved toward digital audio recording as a backup or primary record, reducing reliance on real-time human transcription. But audio recordings have their own problems: poor microphone placement, overlapping speakers, mumbled testimony. They also shift the accuracy burden to whoever later transcribes the recording, which increasingly means an AI system.

What to Watch

Three developments will determine how this issue evolves. First, whether the Judicial Conference of the United States issues formal guidance on AI use by court reporters before the end of 2026. The Fifth Circuit's opinion has created pressure, but institutional change in the federal judiciary moves slowly.

Second, whether litigants begin challenging convictions and judgments on the ground that AI-contaminated transcripts denied them a fair proceeding. The Due Process Clause requires accurate records for meaningful appellate review. A defendant who can show that AI hallucinations corrupted the transcript of their trial has a plausible constitutional claim. If such challenges succeed, the cost of AI transcript errors could escalate from professional embarrassment to case reversals and retrials.

Third, whether the market for court transcription technology moves toward verifiable, auditable systems that pair AI assistance with cryptographic proof of accuracy, or whether it continues to treat transcription as a commodity where speed and cost matter more than integrity. The technology to create timestamped, hash-verified transcript segments already exists. Whether courts demand it is a question of institutional will, not technical capability. The same cryptographic principles that make Bitcoin's ledger trustworthy could make court records trustworthy. The question is whether the legal system recognizes the need before the next AI hallucination changes the outcome of a case.

---

Source: Reason