Interaction safety
Longitudinal evaluation
Longitudinal evaluation assesses an AI system by following the same people using it over an extended period, rather than scoring isolated replies. It can detect cumulative effects, including trust, dependency, attachment, and influence, that snapshot benchmarks cannot detect.
Also called long-term AI evaluation, in-the-wild evaluation
Benchmarks answer one question well: given this input, how good is the output? They are reproducible, comparable, and cheap. They cannot see what takes time to happen, including much of what matters about a system someone speaks with every day.
Longitudinal work asks what happens to a person and a relationship across months of contact. It can reveal trust rising without testing, agreement eroding someone’s appetite for challenge, or reliance displacing something else. No single exchange displays those effects.
The work is difficult. Longitudinal studies are slow, expensive, difficult to control, and involve highly sensitive material. Credible studies also need access to genuine ongoing relationships. That requirement helps explain the gap between their scarcity and these systems’ widespread use.
Why it matters
The relational-AI effects that warrant the most attention are often invisible to dominant evaluation methods. Without longitudinal work, the field measures convenience instead of the effects at issue.
Where TAICU stands
TAICU operates HUG partly to make this research possible: a real system in real use over real time. The lab treats it as a place to form better research questions, rather than evidence by itself.
Common questions
- What do AI benchmarks fail to observe?
- They score replies in isolation. Trust development, displaced reliance, and a system’s effect on someone across months do not exist at the scale of one reply, so they cannot be measured there.
- What resources and conditions does longitudinal AI research require?
- It requires access to genuine ongoing relationships over long periods, careful treatment of highly sensitive material, and patience for slow, expensive work that yields fewer clean comparisons than benchmarking.