When a new study from Northeastern University tested leading AI systems on questions of self-harm and suicide, most models stumbled. Some provided dangerous information; others failed to recognize warning signs. The headlines that followed, from The New York Times to TIME, all focused on the danger.

But deep in the study was a quieter finding: one system did not go outside its guardrails. Pi AI. According to the Northeastern University Institute for Responsible AI research report “Pi "is the only model in our evaluation series that does not provide any information, except contacts and
resources for help-seeking in both test cases."

That result is important, but not because we think we’ve “solved” safety concerns. Far from it. No AI system will ever be perfect, and human minds and societies are constantly changing. But if one model can avoid pitfalls while others fail, shouldn’t we be asking why? Progress doesn’t just come from studying our failures. It also comes from sharing our successes and learning from what works.

Why did Inflection AI’s Pi pass this test? Four design choices stand out.

• First, we treat every interaction as part of a longer conversation, not a one-off question. When one of our users confesses feelings of depression or suicidal ideation and later asks about the heights of the nearest bridges, we connect the dots and Pi directs people to helpful resources where appropriate. Our design principles guide our systems to be built as relational, not transactional. Too many systems treat each question in isolation, which is a recipe for missing hidden cries for help.

• Second, we use intent-based routing. In plain terms, this means our system is built to notice when a conversation veers into sensitive territory. At that point, it shifts gears, drawing on different safeguards than it would in a discussion about recipes or math homework.

• Third, we reinforce that AI is just a machine, not a human. AI should never represent itself as human.

• Finally, we don’t confuse “being agreeable” with “being safe.” AI should not risk harm just to please the user. That balance, between empathy and responsibility, is something we train for deliberately.

These technical design choices aren’t unique to us. They could be adopted widely. But if the conversation only dwells on where AI fails, the industry and the public will miss the chance to replicate approaches that help keep people safe.
People deserve systems that don’t just avoid harm, but actively safeguard human life.

Thanks to Annika M Schoene, PhD and Cansu Canca, Ph.D. from Northeastern’s Responsible AI Practice for conducting this important research.

Their paper is at https://lnkd.in/gNRMbEQv