Pangram Labs says its text detector flags human writing as AI-generated about 41 times in 1,000,000 samples — a false-positive rate of 0.0041%, roughly one error per 24,000 documents, by the company’s own benchmark. That is the tool federal prosecutors in the Northern District of Texas ran this month against a defense lawyer’s filings, then cited in a motion accusing him of “defense by ChatGPT” and “AI slop,” according to court filings reviewed by 404 Media.
The mechanism that made this possible is a local rule. Northern District of Texas Local Civil Rule 7.2(f), and its criminal counterpart Local Criminal Rule 47.2(e), require any brief “prepared using generative artificial intelligence” to say so on its first page. Silence counts as an affirmative certification that no AI was used at all — so once a lawyer discloses any AI assistance, under a rule effective since September 2025, the filing invites scrutiny of exactly how much AI did the work. The government’s attorneys supplied that scrutiny themselves, running the disclosed filings through Pangram and attaching the results as exhibits.
The case is a federal prosecution of Khalwinder Bhengura, a 70-year-old Indian national and lawful permanent resident who was arrested in September alongside six other foreign nationals in a North Texas sweep of illegal-voting cases. Bhengura became a permanent resident on June 29, 2022. On Aug. 1, 2024, he applied for a Texas driver’s license in Tarrant County and, through the state’s “motor voter” process, was registered to vote; on Nov. 5, 2024, he voted in person in Arlington in the general election that included the presidential race. A jury convicted him on Oct. 27 of voting as a noncitizen and falsely claiming U.S. citizenship, according to the U.S. Attorney’s Office for the Northern District of Texas. That verdict covered two of the three counts in his original arrest; his lawyers say the jury acquitted him on the one tied to whether he knowingly claimed citizenship, and that he does not speak English well. Sentencing is set for Feb. 11, 2027, after which he faces deportation.
Government attorneys told the court they were tired of responding to what they called meritless filings. “This case has turned into defense by ChatGPT,” they wrote, adding that “virtually nothing of substance in these filings seems to have been drafted by a human on the Defense team” apart from quoted case law and record evidence. The Pangram results attached to that motion were mixed — none of the filings came back as 100% AI-written — but prosecutors argued that didn’t matter: the United States had been trying to “economize the amount of effort it puts into responding to Defendant’s meritless motions given that Defendant appears to be almost entirely using AI to generate them.”
This Court is disturbed by counsel’s repeated filings which subtly bend the law by proffering superficially persuasive but fundamentally meritless arguments so clearly generated by artificial intelligence. Defendant is not a pro se litigant; he has three counsel of record with decades of practice between them. It is distressing to see how quickly the advent of artificial intelligence has undermined counsels’ ability to properly advocate for their client.
That’s Judge Mark Pittman, who is also the trial judge who presided over Bhengura’s conviction. His order doesn’t cite a specific fabricated case or a false quotation — the failure mode that has sunk dozens of other AI-assisted filings since 2023 — it cites the arguments themselves as artifacts of a machine. Bhengura’s lawyers, Nicole Knox and Mario Herrera, dispute that framing. Herrera told 404 Media that Knox “experienced a medical emergency while we were at the U.S. Attorney’s Office preparing the case with the prosecutors,” leaving him to prepare a federal jury trial in four days with lead counsel hospitalized. “I utilized all available legal tools and did not violate any rules or laws in doing so,” he said. Knox said the same: “AI could have never come up with our argument,” and that under the new disclosure rule, “if you obtain one sentence or phrase through the use of A.I., then you must disclose it. We have nothing to hide.”
Herrera’s sharper objection is to what Pangram actually measures. “An AI-detection software identifies patterns in writing styles based upon a document’s content. It cannot determine who wrote a document, how it was written, or if the law applied within a document is correct,” he said, noting the government never identified a false citation in his filings — the thing that has actually gotten other AI-assisted lawyers sanctioned. That’s a fair description of the tool. Pangram, founded by Stanford graduates Max Spero and Bradley Emi in 2023 and renamed from Checkfor.ai in 2024, trains on paired human and AI text matched for topic, length and tone, then scores a document token by token for statistical fingerprints of machine generation. It says nothing about legal reasoning, factual accuracy or who gets credit for an argument — the same way a spell-checker can’t tell you if the thesis is wrong.
Pangram’s growth is itself a data point in how fast this kind of tool has moved into institutions that depend on trusting written text. The company told the New York Times its monthly users rose from 2,700 in June 2025 to 120,000 in June 2026, a roughly 44-fold increase in a year, alongside a 35-fold rise in annualized revenue, according to the company’s account summarized on its Wikipedia page. It raised $9 million from Menlo Ventures in July 2026 on top of an earlier $4 million seed round, bringing total funding to roughly $13 million. Its current model, Pangram 4, claims a false-negative rate — AI text it misses — of 0.3396%, down from about 1.99% in the prior version, per the company’s own benchmark reporting. Independent academic tests, including a National Bureau of Economic Research working paper and a University of Chicago Booth study, have found Pangram outperforms rivals like GPTZero and Originality.ai at holding false positives near zero while catching AI text — but Epoch AI’s July 2026 evaluation found it still missed 10.1% of AI passages written to imitate a specific human author’s style, a failure mode plausibly relevant to lawyers drafting in a formulaic house voice. Pangram’s CEO, Max Spero, told 404 Media that most law-firm use he’s heard of is internal — firms checking whether their own junior associates are quietly outsourcing drafts to a chatbot, not prosecutors checking opposing counsel.
Pangram's user base, one year Monthly active users
| | Monthly active users (users) |
|---|---|
| June 2025 | 2,700 | | June 2026 | 120,000 |
Source: Pangram Labs, via The New York Times. Two data points as reported by the company; no intermediate months disclosed The prosecution’s underlying case belongs to a category that is, itself, measured in single digits against denominators in the tens of millions. Illegal voting by noncitizens shows up at vanishingly low rates whenever anyone counts it with real data rather than anecdote. A review of 84 million votes cast across 22 states turned up 14 referrals for prosecution — 0.000017% of ballots. A Justice Department unit examining the 2002 and 2004 federal elections found provable fraud in 0.00000013% of ballots cast. A Heritage Foundation review of Pennsylvania elections from 1982 to 2022 found 39 cases, or 0.000039%, among more than 100 million ballots. None of that makes Bhengura’s conviction wrong on its own facts — the jury found he voted as a noncitizen, whatever his understanding of the paperwork — but it is the backdrop against which the Trump administration has made public examples of the dozen-odd cases like his.
What’s new here isn’t that a lawyer used AI badly — that’s been litigated, publicly and often, since 2023, almost always over a hallucinated citation a judge could check against a reporter. It’s that an AI-detection score, with an error rate its own maker measures in parts per 24,000 under laboratory conditions and parts per ten in an independent test of style mimicry, is now sitting in the record as evidence that a human lawyer’s reasoning doesn’t deserve a response. No federal rule of evidence or procedure sets a reliability threshold for that kind of score, the way Daubert does for expert testimony. Judge Pittman’s order suggests he didn’t need Pangram’s numbers to reach his own judgment about the filings in front of him. The open question is what happens in the next case, with a less skeptical judge, where a detector’s score is the argument rather than an exhibit to one.