Interaction safety
Interaction safety
Interaction safety studies harms arising from an ongoing relationship between a person and an AI system, rather than from one output. It includes influence, dependency, attachment, misplaced trust, and emotional failure: risks that appear across time and context.
Also called relational AI safety, safety in human-AI interaction
Most AI safety work evaluates outputs: dangerous content, jailbreaks, and accuracy. That work remains necessary. Systems people live alongside create further harms that do not sit inside a single reply.
A system may pass every content evaluation and still leave someone less able to tolerate disagreement, more isolated, or more certain of beliefs nobody tested after six months. Each reply along that path can appear safe. The trajectory is unsafe, and reply-by-reply evaluation cannot see it.
This work needs different methods. The relationship becomes the unit of analysis, months replace turns as the timescale, and evidence often lies in what people do outside the conversation. That evidence is correspondingly difficult to collect responsibly.
Why it matters
Relational systems reach millions of people while the evaluation methods for their characteristic risks barely exist. The problem is methodological, rather than a lack of willingness to measure.
Where TAICU stands
Interaction safety is one of TAICU’s three research pillars. The lab holds that benchmarks alone cannot evaluate these risks. Real relationships must be studied over time, which is part of why TAICU operates a system of its own.
Common questions
- How does interaction safety differ from AI safety?
- Conventional AI safety evaluates outputs, including harmful content. Interaction safety examines harms emerging from an ongoing relationship, including dependency, misplaced trust, and influence. A single reply cannot reveal those harms.
- Why do benchmarks fail to measure interaction safety?
- Benchmarks score isolated replies. These harms belong to a trajectory across months, where each individual reply may be defensible while the pattern is harmful.