Some people are using AI models to hone biological pathogens and build autonomous drone swarms, the newsletter reports.
MIT Technology Review’s newsletter argues that the industry places too much faith in AI models refusing harmful requests. Chatbots turn down prompts such as how to poison a colleague, but those refusals fail in some cases.
Arthur Holland Michel wrote the piece for the next print issue. Researchers and governments are debating where to draw the line between prompts a model should obey and prompts it must disobey. Some people are attempting to use AI to hone biological pathogens and build autonomous drone swarms.
Refusal training cannot serve as a reliable safety layer on its own. The same genetics expertise that could help cure cancer could also produce bioweapons, so teams building dual-use systems carry real risk. Failing refusals on high-risk prompts matter more than on benign ones.
Governments will draw their own lines about which prompts models must refuse, and those rules may restrict free speech. The newsletter also covers new research on GLP-1 weight-loss drugs, including possible links to hair loss, nail disorders, and nerve damage behind the eye. Envision Energy appears in the 2026 Climate Tech Companies to Watch list for powering data centers with wind and batteries.
What matters
- MIT Technology Review reports that AI models refuse many prompts, but those refusals fail in some cases.
- Builders cannot treat refusal training as a reliable safety layer for high-risk model use.
- Governments will draw their own lines on what models must refuse, which may restrict speech.
Why it matters
Governments will draw their own lines on what models must refuse, which may restrict speech.
This GenAI News article was prepared in original wording using reporting and materials published by MIT Technology Review AI. Source reference: https://www.technologyreview.com/2026/10/09/1146250/the-download-ai-refusal-problem-weight-loss-drug-side-effects/.
Drafted by the GenAI News review pipeline.