AI-rewritten: This is a summary of an article from MIT Technology Review, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

People often assume artificial intelligence can naturally refuse harmful requests, but early models lacked this ability. Steven Adler of OpenAI noted that initial versions would answer any question, including how to kill oneself with a gun. Ryan McBain of Harvard added that these early systems could easily generate violent instructions because they were trained on vast amounts of web content containing violence and vitriol.

Today, companies train models to refuse many prompts by rewarding refusals for harmful questions and punishing over-refusals for harmless ones. They use other AI systems to teach the core models how to say no. However, these refusal mechanisms are probabilistic and often fail. Some users have successfully used advanced AI to attempt creating biological pathogens or building autonomous drone swarms.

The line between what a model should obey and what it must disobey is unclear. Zico Kolter of OpenAI’s board stated that determining where to draw this line is a huge question. Currently, private companies decide these boundaries secretly, but governments may soon intervene. There is a risk that strict refusal rules could stifle legitimate speech or allow oppressive regimes to block critical ideas.

The technical process behind AI refusal involves specific patterns in the model’s internal calculations. A recent Google-funded study describes these patterns as "high-dimensional polyhedral cones," which Jannes Elstner of Apollo Research explains are simply lines pointing in roughly the same direction. If researchers eliminate these specific activation patterns, a model previously shown to refuse will no longer refuse the same prompts.