Blog article Artificial intelligence
Observing AI also changes the question
Evaluation context is part of the behavior we measure. What does that mean when reading an experiment with generative agents?
Vinicius Covas1 min read
The situation matters
A model’s response does not appear in a vacuum. There are instructions, a task, interlocutors and expectations about what counts as an adequate answer. In research on AI communication, those conditions are part of the object of study. A useful comparison needs to explain both which system responded and the situation in which it did so.
Observable behavior, careful interpretation
The linked preprint studies register changes under different observation conditions. The question is how to describe this adaptation without automatically attributing human mental life to the system. “Knowing”, “wanting” or “feeling observed” may be intuitive expressions, but they need operational definitions. We should distinguish what is recorded in messages from the explanations offered to interpret it.
Three questions for an audit
What information does the system receive about the evaluation? Does the comparison keep other conditions constant? Does the result persist when the framing changes? Documenting these choices makes the scope of the evidence discussable. A convincing response in one test is not sufficient on its own to predict behavior in every setting. The task is to build comparisons that make that distance visible.
Communication researcher and professor. I study how humans and machines build meaning and relationships.
About the author
