Model drift

What it means

You buy an AI tool, test it, and it works. Months later its answers are worse, and nobody changed anything you can see. That is the problem the term names. IBM defines it this way: "Model drift refers to the degradation of machine learning model performance due to changes in data or in the relationships between input and output variables."¹ A model is the trained part of an AI system that makes its predictions or answers.

Here is an illustration. Picture a tool that learned to flag unusual expense reports from last year's records. This year the company opens offices in new countries, and travel patterns change. The tool works exactly as it did before, but the world it was built for has moved, so its flags get less reliable. IBM names one cause of this kind: "Data drift— also known as covariate shift—happens when the underlying data distribution of the input data has changed."¹

Where the word comes from

The idea is older than today's chatbots. In 1986 the researchers Jeffrey C. Schlimmer and Richard H. Granger published a paper called "Beyond Incremental Processing: Tracking Concept Drift."² It describes systems that must "tolerate noise (less than perfect feedback) and drift (concepts that change over time)," and that must tell whether a mistake is just noise or "an indication that the concept is beginning to drift."² A companion paper that year began from the same problem: learning "is further exacerbated when the concepts may change over time."³

The research field kept the term "concept drift."⁴ Business explainers such as IBM's now use "model drift" as the broader label, with data drift and concept drift as its causes.¹ With large language models, the word has also come to cover a different case: the service behind a familiar product name changes. Researchers studying ChatGPT found that "the behavior of the 'same' LLM service can change substantially in a relatively short amount of time."⁵ (A large language model, or LLM, is the kind of AI behind chatbots like ChatGPT.)

How others define it

• IBM: "Model drift refers to the degradation of machine learning model performance due to changes in data or in the relationships between input and output variables."¹
• Schlimmer and Granger (1986): drift means "concepts that change over time."²
• Lu and colleagues, in a review of the research: "Concept drift describes unforeseeable changes in the underlying distribution of streaming data over time."⁴

Where they agree and differ

All three definitions agree that something changes over time and that a system's past performance does not guarantee its future performance. They disagree about what changes. In the 1986 papers and the research review, the world changes: the concepts or the data.²⁴ IBM puts the stress on the result, a drop in performance, and lists changes in data or relationships as the causes.¹ The ChatGPT study points to a third case, in which the service itself changed.⁵

That difference matters, and the strongest view against using one word for all of it deserves a fair hearing: these are different problems with different fixes. If the world shifted, you retrain or adjust the model. If a vendor changed the model, you need the vendor to tell you, and you need your own tests. We think the single word is still useful for leaders, because the response starts the same way in both cases: you keep measuring. That is our view so far, and it is open to testing. Evidence that teams handle the two cases better under separate names would change our mind.

In the news

The best-documented example we found is a study by Lingjiao Chen, Matei Zaharia, and James Zou, which compared versions of GPT-3.5 and GPT-4 from March and June 2023.⁵ They report: "For example, GPT-4 (March 2023) was reasonable at identifying prime vs. composite numbers (84% accuracy) but GPT-4 (June 2023) was poor on these same questions (51% accuracy)."⁵ Their conclusion was practical, "highlighting the need for continuous monitoring of LLMs."⁵

The people who felt this were users: same product name, different answers, and no notice that anything had moved. Our News Editor Agent collected no recent stories using this term, so we chose this documented study over a newer but weaker report.

What it means for your organization

AI agents are built on models, so drift is a leadership problem, not only a technical one. An agent that passed your review last spring may not behave the same way today, either because your business changed or because the model under it did.¹⁵ Approval is not a one-time event. Someone in your organization has to own the job of checking, against the same tasks, that the tool still does what you approved it to do.

One question to ask your team: For each AI tool we rely on, who re-tests it, how often, and how would we find out if its vendor changed the model underneath?