US prisons are exploring the use of AI to analyze inmates' phone calls, a practice privacy advocates say could amplify racial bias in the justice system
Reuters : Tweets: @aaschapiro , @lanceulanoff , @onekade , and @foxcahn Tweets: Avi Asher-Schapiro / @aaschapiro : Prisons & jails are quietly deploying tech to transcribe, & analyze the phone calls of incarcerated people - & even “learn” slang. But this speech-to-text tech is riddled by racial bias - often making twice as many errors with Black speakers. https://news.trust.org/... Lance Ulanoff / @lanceulanoff : “Technology that transcribes voice conversations is flawed and has a particularly high error rate when applied to the voices of Black people, according to a 2020 paper on the 5 leading systems by researchers at Stanford University + Georgetown University” https://www.reuters.com/... @onekade : The Democrats in the House want to explore expanding the use of racially biased AI to listen to and analyze prison calls, looking for who knows what. This will disproportionately harm Black people. https://www.reuters.com/... Albert Fox Cahn / @foxcahn : Why are Congressional Democrats trying to fund AI experiments on people in prison? No, AI CAN'T create a crystal ball and predict pre-crime. This biased tech is going to get BIPOC individuals falsely accused of crimes and normalize racist tech. https://www.reuters.com/...
Context & Ripple Effects
This story extends a decade-long arc of algorithmic tools moving into the US justice system — from [[a:869915|crime-prediction software shown by ProPublica to carry strong bias against Black defendants]] to the proprietary recidivism algorithms still feeding sentencing and bail decisions. What is new here is the target: prisons and jails quietly deploying speech-to-text systems that transcribe and 'learn' slang from inmates' phone calls.
The stakes are sharpened by two facts in the coverage: a 2020 paper found the leading transcription tools make roughly twice as many errors with Black speakers, and House Democrats are seeking to fund and expand these AI call-analysis experiments rather than constrain them.
First-order effects
- Incarcerated people's phone calls become machine-analyzed data, and Black speakers bear disproportionate transcription error rates — meaning the record fed to investigators and analysts is systematically less accurate for them.
- House Democrats' push to fund the experiments gives vendors of call-analysis tech a growing, government-backed customer base inside prisons and jails.
Second-order effects
- Defense counsel have already begun attacking AI tooling on reliability grounds in court — as seen in challenges to Cybercheck's accuracy in thousands of US cases — and transcripts generated by biased speech-to-text offer the same attack surface for evidence derived from inmate calls.
- Vendors selling race- or ethnicity-sensitive AI face the same researcher scrutiny raised when companies deployed such software for market research, now applied to a criminal-justice context where errors carry liberty consequences.
Third-order effects
- If the funding pattern holds, algorithmic analysis extends from patrol routing and sentencing support into the daily communications of incarcerated people — deepening the shift toward state-mediated AI across the justice system without a matching governance framework.
- A documented racial-error gap in widely used commercial speech-to-text could become the next flashpoint for regulating accuracy standards in government AI procurement, much as biased risk scores did for sentencing algorithms.
The trend: AI tools are migrating deeper into every stage of the US justice system — patrols, sentencing, probation, and now inmate communications — with each deployment inheriting the bias and reliability problems documented in the previous ones.