Model drift is the gradual decay of an AI model's accuracy as the world it was trained on changes — audiences shift, industries consolidate, behaviour moves. A matchmaking or churn model tuned on past editions slowly mismatches the current one unless it's monitored and retrained on fresh data.
Events are unusually drift-prone because they change in steps, not gradually: each edition brings a partly new audience, renamed categories, a shifted industry mix, sometimes a new venue or format. A model trained on 2023's behaviour meets 2026's reality — the hot product category didn't exist, half the job titles changed, the pandemic-era engagement patterns it learned are ancient history. The result is quiet degradation: matchmaking that was sharp two editions ago now feels generic, churn scores that flag the wrong exhibitors, and nobody notices because the system still produces confident numbers on schedule. Commercially, drift is a stealth tax — you keep paying for AI features whose value erodes yearly while blame lands on "the algorithm" in general. In practice, the defence is measurement and refresh: track outcome metrics per edition (match acceptance, meeting completion, prediction hit rate), compare against earlier editions, and ask your vendor when models were last retrained and on whose data. The common mistake is treating AI as installed rather than maintained — set up at purchase, never revisited, quietly aging. One honest nuance: retraining isn't free of risk either. A model refreshed on one unusual edition — a strike year, a venue change — can learn the anomaly as the new normal. Drift management is judgement plus process, not a checkbox that says "auto-retrain: on."
Direct answer
Model drift is the gradual decay of an AI model's accuracy as the world it was trained on changes — audiences shift, industries consolidate, behaviour moves. A matchmaking or churn model tuned on past editions slowly mismatches the current one unless it's monitored and retrained on fresh data.
More terms
No related terms yet.