Last week we drew a line between rules that steer and architecture that enforces. The word used to describe that enforcement layer was “deterministic”. “Deterministic” is likely a word you know, because determinism is one of the fundamental assumptions of science that every behavior analyst was taught. A challenge to productive communication, however, is that the software use of “deterministic” and the scientific use of “deterministic” are not the same. And, conflating them produces a specific, avoidable error in how people reason about AI. This week we unpack these differences.
Here is an experience you may have had. You ran the same prompt through an AI system twice and got two different answers. Reasonable enough. So you found the temperature setting, set it to zero, and ran it twice more. You still got two different answers, though the difference was smaller.
Most people attribute that to “AI is just like that” and move on. That attribution is an error. Something specific is happening, and knowing what is happening tells you which parts of your workflow you can make reproducible and which you can’t. It is also a useful lens on a word behavior scientists use in a different way.
