Your model did not get lazy. AI performance degradation over time is at least six different problems wearing one name.
AI performance degradation over time broke out as a search term this week, and the searches are ahead of the coverage. People are not asking something abstract. They want to know why the assistant that solved a task last month fumbles it today. The honest answer is that several unrelated failures share that one label, and each needs a different fix.
What AI Performance Degradation Over Time Actually Means
Half the confusion online comes from people arguing past each other. One person means a fraud model that has quietly gone stale since 2024. Another means a chatbot that started repeating itself yesterday afternoon. Those are not the same event, and they do not share a cause.
Six things people call AI degradation
Sorting them out first saves a lot of wasted debugging. Here is what each one really is.
| What people call it | What actually causes it | Where you notice it |
|---|---|---|
| Model drift | The world moves on, training data does not | Accuracy slides month after month |
| Temporal quality degradation | The same aging, measured across many models at once | Published error curves in research |
| Context rot | More input tokens dilute the model’s attention | Long chats and big documents |
| Model collapse | Training runs feeding on AI-generated text | Rare cases and variety disappear |
| Infrastructure bugs | Routing, caching or compiler errors in production | Sudden, provider-wide, temporary |
| Silent config changes | Prompt or reasoning settings shipped without notice | A good model feels dumber overnight |
Above all, notice how different the timescales are. Two of these unfold across years. Two can arrive and vanish inside a single week.
AI Model Performance Degradation Over Time Is Measurable
The oldest version of this problem has a name in the literature. A 2022 paper in Scientific Reports, titled Temporal quality degradation in AI models, put numbers on it for the first time.
The team tested 128 model and dataset pairs drawn from four industries: healthcare operations, transportation, finance and weather. In 91% of them, quality fell as the models aged. Furthermore, the decline was not one shape. Some models drifted up in error slowly, some fell off a cliff, and some simply became unpredictable.
In fact, none of this involves a language model. It happens because a model is frozen at the moment of training while the data around it keeps moving. Consequently, anything trained once and left alone is already on the clock.
Claude AI Performance Degradation: What the Postmortems Found
The loudest complaints of the past two years came from developers convinced that Claude had been quietly weakened. Anthropic eventually published the receipts, and the answer was less sinister and more embarrassing than the theories.
In September 2025, it detailed three overlapping infrastructure bugs. A context window routing error introduced on August 5 peaked at 16% of Sonnet 4 requests on August 31, touching roughly 30% of Claude Code users. A second bug corrupted output, producing stray Thai characters and broken syntax. A third was a compiler fault on TPU hardware. You can read Anthropic’s full postmortem for the timeline. Its flat denial is worth quoting: “We never reduce model quality due to demand, time of day, or server load.”
Then it happened again. An April 23, 2026 postmortem traced a fresh round of Claude Code complaints to three of its own changes.
| Change | Shipped | Undone |
|---|---|---|
| Default reasoning effort dropped from high to medium | March 4 | April 7 |
| Caching bug discarding earlier reasoning each turn | March 26 | April 10 |
| Prompt instruction capping response length | April 16 | April 20 |
The middle one made Claude read as forgetful and repetitive. The third cost about 3% on benchmarks. Users of AI coding tools were right that something had changed. They were wrong about what.
AI Context Degradation and the Long Chat Problem
Meanwhile a subtler effect has its own name now: context rot. Research published by Chroma in 2025 tested 18 frontier models, including GPT-4.1, Claude Opus 4 and Gemini 2.5.
Every single one got worse as input grew. Crucially, the drop did not wait until the context window was nearly full. Performance slid at each increment the researchers tested, so a million-token window still rots long before you reach a million tokens.
In addition, distractors make it worse. Text that looks relevant but is not will actively pull a model off course, which is why a bloated prompt often beats a short one to the wrong answer.
Model Collapse: When AI Trains on AI
By contrast, the slowest failure on the list is also the most quoted. In 2024, Ilia Shumailov and colleagues published work in Nature showing that models trained repeatedly on their own synthetic output degrade in a specific way.
In practice, they lose the tails first. Unusual phrasings, rare cases and minority patterns thin out with each generation, until the model knows only the middle of its own distribution. The paper calls the resulting defects irreversible.
Still, it is worth being precise about scope. Model collapse is a training-time risk for the next generation of models. It is almost certainly not why your chatbot annoyed you this morning.
How to Detect AI Performance Degradation Over Time
Start from an uncomfortable fact: providers update models behind stable API names. A version string you pinned six months ago does not guarantee identical behavior today.
Independent trackers now watch for exactly that. AI Stupid Level scores major models on coding, reasoning, tool calling and speed, then applies change-point detection to separate a genuine sustained decline from ordinary noise. Similarly, IBM uses the term agentic drift for the way small errors compound across a long multi-step run.
Your own defense is duller and more effective. Keep a fixed set of real tasks, run them on a schedule, and log the failures rather than trusting your memory of last month. Teams building on agentic AI frameworks need this most, because an agent hides its own mistakes inside a chain of steps.
Ultimately, the stakes explain the panic. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, blaming costs, unclear value and weak risk controls. It also reckons only about 130 of the thousands of self-described agent vendors are the real thing.
Graceful Degradation Is the Opposite Problem
One phrase in the same search cluster means something entirely different, and it trips people up. Graceful degradation is not a fault. Rather, it is a design goal.
A system degrades gracefully when it loses a dependency and keeps behaving predictably instead of pretending nothing broke. Voice assistants show it clearly. When the model stalls, the agent plays a filler phrase instead of dead air. Unsure speech recognition triggers a polite request to repeat. If the good voice fails, a plainer one takes over.
So the two ideas sit at opposite ends. One is quality leaking away while nobody watches. The other is an engineer deciding in advance what the system gives up first.
Want More on AI Performance Degradation Over Time?
If you mostly notice this in chat, compare the best AI writing tools and keep your own scorecard on them. For the harder question of whether any of it pays off, read what the numbers say about AI coding tools and productivity.
Frequently Asked Questions
What is AI performance degradation over time?
It is an umbrella term for several unrelated failures: model drift, context rot, model collapse, infrastructure bugs and quiet configuration changes. Each has a different cause, timescale and fix, so the label alone diagnoses nothing.
Do AI models really get worse over time?
Yes, and it has been measured. A 2022 Scientific Reports study of 128 model and dataset pairs found quality fell in 91% of them as the models aged, through gradual error growth, sharp drops or rising unpredictability.
Why does Claude sometimes feel worse than last month?
Anthropic has published two postmortems. Three infrastructure bugs degraded responses in August 2025, and a reasoning default, a caching bug and a length-capping prompt did it again in spring 2026. All were reverted.
What is AI context degradation?
Known as context rot, it means output quality falls as input grows. Chroma tested 18 frontier models in 2025 and every one degraded at each length increment, well before the context window was anywhere near full.
What is model collapse?
Nature research from 2024 showed that models trained repeatedly on AI-generated data lose the rare and unusual parts of their distribution first. Variety thins out with each generation, and the paper describes those defects as irreversible.
What is graceful degradation in AI systems?
A deliberate design choice, not a fault. When a dependency fails, the system keeps working predictably at reduced capability: a filler phrase during a stall, a request to repeat, or a simpler backup voice.
