Model Drift Response: Diagnose Before You Retrain

When a production AI system starts behaving differently than it did at validation, the reflex is to retrain it. That's usually the wrong first move. Retraining is the slowest and most expensive fix available, and it only addresses one of several possible causes. For AAISM™ candidates, drift response is a good test of whether you can match a control to a root cause instead of reaching for the biggest lever in the room.

Drift Is the Normal Case, Not the Exception

It's tempting to treat drift as a rare failure. The evidence says otherwise. A 2022 study in Scientific Reports trained four common model types on 32 longitudinal datasets from healthcare operations, transportation, finance and weather. Quality degraded over time in 91% of the 128 combinations. The authors also found that shifts in the input data didn't fully explain the decline, and that the same data could age very differently depending on which model and settings were used.

Drift also doesn't announce itself. The system keeps answering, latency looks fine, and outputs still look plausible. That's why the NIST AI RMF's Measure function expects organizations to monitor production behavior and compare it against the metrics captured during pre-deployment testing. The EU AI Act's Article 72 makes post-market monitoring a formal duty for providers of high-risk systems (though the Digital Omnibus has pushed most of those high-risk obligations back to December 2027).

Three Cases, Three Different Fixes

The input data changed. A JAMA Network Open study of 24 US hospitals found that daily alerts from the Epic Sepsis Model rose 43% in the three weeks after each hospital's first COVID-19 case, even though total patient census fell 35%. The share of patients triggering an alert more than doubled. The patient mix had shifted, with elective surgeries cancelled and the remaining patients sicker on average. The University of Michigan paused the model's alerts in April 2020 after nurses reported being overwhelmed. That pause was containment: it stopped the harm while people worked out what had changed. The underlying problem was the population the model was scoring, not a defect in the model.

The upstream model changed. Researchers at Stanford and UC Berkeley compared the March and June 2023 versions of GPT-3.5 and GPT-4 on the same tasks. GPT-4's accuracy at identifying prime numbers fell from 84% to 51% between the two versions, and both models made more code-formatting errors in June. Anyone building on those APIs saw their application's behavior change without touching a line of their own code. Retraining your own model wouldn't help here. The right response is to pin the model version, re-run your baseline tests before accepting an upgrade, and treat a vendor's model update as a change that has to go through change management.

The update itself caused the drift. In late April 2025, OpenAI shipped a GPT-4o update that made ChatGPT noticeably more flattering and agreeable, including toward harmful ideas. OpenAI's postmortem traced the problem to new training signals based on user thumbs-up ratings, which weakened the signal that normally kept that behavior in check. It also acknowledged that expert testers had flagged that something felt off before launch. The fix was a rollback to the previous version, carried out within days. The follow-up was process change: new evaluations for that specific behavior and stricter launch gating, as the Georgetown Tech Institute's brief summarizes.

A Response Order That Holds Up

Each of those cases needed a different fix, and none of them was a retrain. That points to a response sequence worth knowing cold:

• Contain first: add a guardrail or output filter, tighten the system prompt, or pull the feature if the drift is safety-relevant. This buys time without waiting on a retraining cycle.
• Diagnose the source: a prompt or configuration change (revert it), an upstream model version change (re-test or pin back), a shift in data or context such as a changed retrieval source or new user behavior (fix the pipeline), or genuine drift in a model you fine-tuned yourself. Only that last case actually points to retraining.
• Fix at the smallest effective layer: adjust configuration or prompts before fine-tuning, and fine-tune before doing a full retrain.
• Re-run the baseline: confirm the fix actually resolved the drift rather than hiding it.
• Close the loop: update the model card and documentation, and log the event as an incident so the pattern is visible next time.

The larger issue is speed. Models, vendors and data are all changing faster than most change-management processes were designed to handle. A governance program that can only respond by retraining will always be a cycle behind. One that can contain within hours and diagnose within days can keep up.

Related Reading

For the full exam breakdown, see the AAISM™ Certification Study Guide. For the other Domain 3 technical controls, see OWASP LLM Top 10, Explained for Security Managers. On deciding when a behavior change counts as an incident, see Incident or Research Finding?. On how AI governance sits alongside an existing IT governance program, see IT Governance vs. AI Governance.

Sources

Vela et al., Scientific Reports (2022) — Temporal quality degradation in AI models
Wong et al., JAMA Network Open (2021) — Sepsis model alerts in 24 US hospitals before and during COVID-19
Chen, Zaharia & Zou (2023) — How is ChatGPT's behavior changing over time?
OpenAI — Expanding on what we missed with sycophancy
Georgetown Law Tech Institute — Tech Brief: AI Sycophancy & OpenAI
NIST — AI RMF Playbook, Measure function
EU AI Act — Article 72, post-market monitoring