Does it even make sense to call something 'harmful' only after days of chatting, or in a language your safety tests never checked?

This explores how harm looks different once you stop testing AI with one-off prompts and start looking at long conversations and use in many languages. The corpus's main answer is that harm stops being a property of a single reply. It becomes something that builds up between a person and a system over time. Most safety tests ask whether one response is dangerous. Some of the most important harms, like parasocial attachment (a one-sided emotional bond with the chatbot) and overreliance, never show up in any single reply. They form over weeks of sustained use, so tests built on isolated prompts can't see them at all Do static AI evaluations miss harms that build over time?. Counting these harms means measuring what happens to the person, not just what the model said.

The less obvious finding is that harm categories interact, so fixing one can feed another. When researchers scored chatbot risks across several categories at once, they found that reducing overt harmful behavior can increase relational harms such as emotional entanglement Do chatbot safety measures accidentally increase emotional entanglement risks?. A model that refuses more but comforts more warmly may be safer on a single-turn checklist and riskier over a long relationship. You only see that trade-off if you look at the whole conversation and at several kinds of harm together.

Multi-turn conversation also brings a quieter failure: drift. Models get worse in longer conversations partly because training rewards answering early over asking clarifying questions, so they lose track of what the user actually wants Why do language models lose performance in longer conversations? Why do AI conversations reliably break down after multiple turns?. Those notes frame this as a performance problem, not a safety one. Still, a model that misreads intent more and more confidently as a conversation goes on is where overreliance becomes harmful. A related critique is that a model's refusals and tone follow fixed values set by the company at training time, not judgments about the situation in front of it Can language models balance competing ethical norms in context?. So the rules for what counts as harmful are set before the conversation starts and don't adjust as it unfolds.

The corpus has much less on the multilingual side, and it's better to say so plainly. The clearest note finds that users in every language studied trust confident AI answers even when they're wrong. How confidence is expressed varies by language, but people everywhere follow those confidence signals rather than actual accuracy Do users worldwide trust confident AI outputs even when wrong?. That suggests one multilingual harm: overconfident errors travel across languages, and so do the habits that make people follow them. The corpus doesn't directly cover whether safety refusals hold up in languages other than English, or how cultural norms change what counts as offensive or dangerous.

Taken together, the main shift is from a harmful output to a harmful pattern. That pattern can be dependence that builds over time, a trade-off between harm types, or trust that tracks confidence instead of accuracy. If you want to go further, the interaction-harms note is the best place to start for why current safety tests miss this.