In 2019 — three years before the release of popular U.S. chatbot ChatGPT — Purdue University was already thinking about harnessing “AI” for the justice system. While not nearly as advanced as modern AI, the $2 million, federally funded project would have upgraded the technology for monitoring probationers and parolees. This month, the U.S. government released an audit of the program, which collapsed in 2023 on deployment failures and insufficient oversight and rules. The case sheds light on novel questions around testing new technology on people involved with the justice system.

Attention on AI “harnesses” is growing as AI systems move from mere chat sessions on the web to work settings where they can read files and take actions. A harness is the operating layer around a model — safety checks that determine what the AI can do.

Early public-sector guidance, including a 2024 report from the U.S. Department of Commerce, put many safeguards at the “harness” layer of AI. Researchers at Princeton have also argued that some of the more publicized improvements to AI systems were the result of better harnesses, not models (Kapoor et al., 2024).

A bipartisan bill would kick off formal research on using AI speech recognition for court transcription. (FedScoop)

The data of almost 300 million students and teachers across the globe may be compromised, with the FBI issuing a warning. (AP)

The U.S. Treasury’s latest report on money laundering suggests AI is a fraud amplifier. (Treasury)

Nvidia is helping a major U.S. drug maker build its own private AI engine. (Reuters)

EY Canada’s hallucination blunder has the company reviewing how it publishes. (Financial Times)