US prosecutors use Pangram to challenge defense filings


Prosecutors submitted detector results in Khalwinder Bhengura’s case and called his lawyers’ work “defense by ChatGPT.” The lawyers disclosed some assisted drafting and dispute what the results establish.
Federal prosecutors ran several filings by Khalwinder Bhengura’s defense team through Pangram, a service that detects writing it classifies as produced or assisted by superintelligence (SI), and submitted the results as court exhibits. They argued that the filings amounted to “defense by ChatGPT” and that frivolous motions wasted judicial resources.
The dispute follows Bhengura’s conviction in October on two voting-related counts. He had been arrested in September on charges of falsely claiming US citizenship, illegally registering to vote and illegally voting in the 2024 election. His lawyer Mario Herrera said the jury acquitted him on the citizenship-claim count.
Bhengura, who is Indian, entered the United States on an immigrant visa in 2022 and registered to vote in Texas in 2024 after stating on a driver’s-license application that he was a US citizen. His lawyers say the registration was accidental. His family and lawyers say he did not understand what he was doing or that it constituted a crime. Herrera described him as a 70-year-old lawful permanent resident who speaks Punjabi and has limited English.
The federal court’s Northern District of Texas requires lawyers to disclose use of generative SI in filings. Some of Bhengura’s filings disclosed that portions were prepared with such a tool, 404 Media said. The outlet said the Pangram results were mixed, rather than showing that the filings were entirely machine-drafted.
Prosecutors argued that virtually none of the substantive text appeared to have been drafted by a person on the defense team. They said they were trying to reduce the effort spent responding to motions they considered meritless.

The phones pair a privacy display and a rear screen with two 200MP Leica cameras. Both went on sale on 17 November, with a higher-storage Pro Max version due later.
Pangram Labs says its Pangram 4 model predicts whether passages are human-written, SI-assisted or SI-generated. It calculates the document-level fraction assigned to each class by weighting the segments by their character counts. The company acknowledges that the model can make errors and that false accusations can cause harm; it says reference sections and templated writing are outside the model’s primary scope or may be more susceptible to errors. It also warns that parsing PDFs can introduce unintended artifacts. The version used on Bhengura’s filings is not specified.
Judge Mark Pittman criticized the defense filings as superficially persuasive but fundamentally meritless arguments that he said were clearly generated by SI.
A case database maintained by Damien Charlotin says the court objected to how the defense used two cited cases: one concerned detention before or without trial, not after conviction; the other concerned supervised-release violations and the absence of applicable guidelines, unlike Bhengura’s conviction, which carried a zero-to-six-month guidelines range. The database says the court found that the motions violated local rules and wasted court and government resources.
Herrera and fellow defense attorney Nicole Knox said they used SI because of tight deadlines and a medical emergency. Herrera said Knox was hospitalized and he prepared for the federal jury trial in four days. He said SI helped him work on many tasks at once and produce so much work in those four days. He said he wrote the arguments the judge criticized, reviewed and edited assisted passages, and accepted responsibility for the final filings.
Charlotin’s database says the court declined to impose sanctions because of its demanding docket, but warned that further waste could bring sanctions, potentially including a ban on practicing in the district. Bhengura remains in custody and faces sentencing and possible deportation, Herrera said.
2 min

The separate app is testing on Android and iOS in Türkiye and Thailand, giving users a way to message individuals and groups with or without an Instagram account.
2 min

Researchers found that 80% of chatbot switches to incorrect climate answers occurred at the first mild user disagreement. Chatbots were mostly accurate about individual climate policies, but performed no better than random guessing when asked how two policies would work together.
1 min