AI Dictionary › AI Fundamentals
Deriva (drift)
Drift is the gradual departure of an AI system's behavior from what was set up, without anyone having decided it. A model with fixed weights does not change from one day to the next; what changes is what goes into it: the sources, the accumulated conversations, the growing instructions, the environment it acts in. The system follows the most abundant direction in its input, and within a few weeks it writes, decides or answers differently from how it did at the start.
Inside a single long conversation drift has a precise cause: the model imitates its own earlier text. A mistake left uncorrected in the morning becomes, in the context, a successful example, and by the afternoon it is house style. In agents that run for weeks drift comes through the state they carry and the environment that rewards them: an agent on a social network learns what the network applauds.
Two kinds of drift call for opposite treatment. Drift in form (tone, format, rhythm adapting to the environment) is adaptation, and often intended. Drift in fact (an invented number, a wrong attribution, a forgotten constraint) is corruption, and must be stopped every time. The operating rule: drift is not eliminated, it is measured. You need an external, dated check that compares today's output with last month's.
Grace keeps a sentinel on the model that powers it: every week it compares the evaluations against a recorded baseline and warns if the judgment shifts. That is drift measured from outside, applied to the one doing the measuring.
From our network
HSE Genius: AI for Safety Data Sheets
Extract SDS data, H phrases and ECHA compliance checks in seconds, powered by AI.
Visit hsegenius.com →From the Agora Intelligence blog
Say what you mean. Get what you need.
Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.