Can AI Forget What It Learned? Why AI Forgets Old Things When It Learns New Ones
SUMMARY OF THIS ARTICLE
- Yes, AI can forget what it learned, and researchers call the most dramatic version of it “catastrophic forgetting”.
- AI does not remember like a person. It learns patterns, stored as millions of tiny numbers, and new training can change those numbers.
- When a model is taught one thing after another, the new lesson can overwrite the old lesson, sometimes very fast.
- This is different from a chatbot seeming to forget in the middle of a chat, which is usually about limited short-term context, not lost training.
- Researchers reduce forgetting by replaying old examples, protecting important parts of the model, giving new tasks their own space, or retraining from scratch.
Introduction: A Strange Kind of Forgetting
Imagine you teach a child to ride a bicycle, and then you teach the same child to swim. You would not expect the swimming lessons to make the child forget how to ride a bicycle. A person can add a new skill without losing an old one. That is one of the most ordinary, and most remarkable, things about human learning.
Artificial intelligence does not always work that way. Ask a simple question and the answer can sound surprising: can AI forget what it learned? The short answer is yes. In some cases, an AI system can learn something new and, in the process, lose much of what it learned before. Researchers have a dramatic name for this problem: catastrophic forgetting. The word “catastrophic” does not mean the computer breaks or explodes. It means the forgetting can be sudden, large, and much worse than the gentle forgetting we see in people.
This article explains, in plain language, what AI learning really means, why forgetting happens, and how it is different from the everyday feeling that a chatbot has “forgotten” what you just told it. It also looks at real examples, from shopping recommendations to fraud detection and self-driving cars, and at the clever ways researchers try to stop the forgetting. At the end, it explains what this means for ordinary users, and what is still not fully solved.
You do not need to know maths, coding, or computer science to follow it. If you have ever used a voice assistant, a photo app, a music app, or a chatbot, you already know the kinds of systems this article is about.
What Does It Mean for AI to “Learn”?
When people say an AI has “learned” something, they do not mean it has stored a memory the way you remember your first day at school. A typical AI system, especially a neural network, learns by finding patterns in a large amount of data.
Think of a child learning to tell cats from dogs. The child sees many pictures, notices that cats often have pointed ears and whiskers, and dogs often have a different shape and bark, and slowly builds a sense of the difference. An AI system does something similar, but in numbers. It is shown many examples, makes guesses, gets told when it is wrong, and adjusts itself a little each time. After many rounds, its guesses become much better.
Inside the system, that learning is stored as a huge collection of numbers, often called weights. Each weight is like a tiny dial that controls how strongly one part of the system influences another part. When the system is trained, those dials are turned, bit by bit, until the system gives good answers for the examples it has seen.
This is important for understanding forgetting. An AI does not keep a separate, neat file for each thing it learned, the way you might keep separate folders for school, work, and hobbies. Instead, the same dials are used for many different tasks. A dial that helps the system recognise a cat might also help it recognise a dog, a car, or a face. Because the knowledge is shared and spread out across the whole system, changing the dials to learn something new can accidentally change what the system already knew.
It also helps to remember what AI learning is not. A trained AI does not understand the world the way a person does. It does not know that a cat is a living animal with feelings, or that water is wet in the way you feel it. It has learned statistical patterns: which shapes, words, sounds, or numbers tend to go together. That is powerful, and it is why AI can translate languages, recognise faces, and suggest songs you might like. But it is also fragile, because a pattern that works well for one set of examples can be pushed out of shape by a very different set of examples.
The Famous Experiment: Learning One Thing, Then Losing It
The problem of AI forgetting was shown clearly in the late 1980s, in work by researchers Michael McCloskey and Neal Cohen, published in 1989. Their experiments were simple, and that simplicity made the result striking.
In one version, a small neural network was trained to learn pairs of words, a bit like learning that one word goes with another. First it learned one list of pairs until it could recall them well. Then it was trained on a second list, using some of the same starting words but with different partners. When researchers tested it again on the first list, the network had forgotten almost all of it. People given a similar task did much better. They could learn the second list without losing the first list almost completely.
In another version, the network learned simple addition problems, such as 1 + 1 and 1 + 2, until it answered them correctly. Then it was taught new addition problems, such as 2 + 1 and 2 + 2. After learning the new problems, it could no longer solve many of the old ones reliably. Even a small amount of new training could cause a large amount of forgetting.
At the time, some researchers had hoped the problem would be mild. They thought the many small changes made while learning the new task might mostly cancel each other out, leaving the old knowledge largely intact. McCloskey and Cohen showed that this hope was wrong, at least for the simple networks they tested. The forgetting was not gentle. It was drastic, and that is why they described it as catastrophic.
Later research, including work by Roger Ratcliff in 1990, confirmed the same basic pattern. The exact size of the effect depends on the system, the tasks, and how the training is done, but the core lesson has held up for decades: if a neural network is trained on one task, and then trained only on a different task, it will tend to forget the first task, sometimes very quickly.
The Stability-Plasticity Dilemma: A Tug of War
Researchers describe the core problem as the stability-plasticity dilemma. Those two long words hide a simple idea, and it is a tug of war that every learning system faces.
Stability means keeping what you already know. Plasticity means being able to change and learn something new. A system needs both, but they pull in opposite directions.
If a system is too stable, it refuses to change. It remembers the old lessons perfectly, but it cannot learn anything new. Imagine a student who will not change a single answer, even when shown clear evidence that the answer is wrong. That student will never improve.
If a system is too plastic, it changes too easily. It learns the new lesson quickly, but it overwrites the old lessons in the process. Imagine a student who erases the whole notebook every time a new lesson starts. That student can learn the latest lesson, but cannot keep a growing body of knowledge.
AI systems, especially when trained one task at a time, struggle with this balance. The dials that store old knowledge are the same dials that must move to learn new knowledge. If the system moves them a lot, it learns fast but forgets fast. If it moves them only a tiny amount, it forgets slowly but also learns slowly. Finding the right balance, for each part of the system and each new task, is one of the central challenges of modern AI research.
Two Kinds of Forgetting: In a Chat, and After Retraining
Many people first notice AI forgetting in a conversation with a chatbot. You tell the chatbot your name, your plan for the weekend, or the details of a story you are writing. A few messages later, it seems to have forgotten what you said. It asks you to repeat yourself, or it gives an answer that ignores the earlier detail. This can feel like the AI has a bad memory, and it is natural to wonder whether the model has “forgotten” its training.
In most cases, that kind of forgetting is different from catastrophic forgetting, and understanding the difference helps a lot.
A chatbot in a conversation works with two kinds of information. The first is its training, the patterns it learned from a huge amount of text, pictures, or other data before you ever met it. The second is the context of your current chat, the words you have just exchanged in this conversation. That context is like a short-term notepad. The chatbot can only look at a limited amount of recent conversation at once. When the conversation gets long, older messages may fall off that notepad, or become less clear to the system. The chatbot has not lost its training. It has simply run out of room to keep every detail of this one chat in view.
The second kind of forgetting happens after retraining, and this is the kind researchers worry about most. Suppose a company trains a model to recognise road signs, and later trains the same model, on its own, to recognise animals. If the model is only shown animal pictures during the second training, it may become worse at road signs. The dials that helped it recognise a stop sign have been turned to help it recognise a cat instead. This is not about a chat getting too long. It is about the stored patterns themselves changing.
Everyday Examples: Where Forgetting Shows Up
Catastrophic forgetting is not only a laboratory curiosity. It can show up in ordinary products, sometimes in ways users notice, and sometimes in ways they do not.
Consider a music or shopping recommendation system. It learns what you like from your past choices. If your tastes change, or if the system is retrained only on recent behaviour, it may start to forget your longer-term preferences. You might notice that it suddenly suggests only the newest trend, and stops suggesting the older songs, books, or products you loved for years. The system has not deleted your account. It has shifted its patterns toward the newest data, and the older patterns have faded.
Fraud detection is another example. A bank’s AI system learns what normal spending looks like, and what suspicious spending looks like. Fraud patterns change over time, so the system must keep learning. If it is retrained only on the newest fraud examples, it might become worse at spotting older fraud tricks that are still in use. That is why banks and other companies usually retrain on a mix of old and new examples, not only on the newest ones. They want the system to add new knowledge without losing old knowledge.
Self-driving cars face a similar challenge, with much higher stakes. A car system must recognise road signs, pedestrians, cyclists, and many other things, in rain, sunshine, snow, and at night. If it is trained in one city, and then trained only in a very different city, it might become worse at handling the first city. If it is trained only on sunny days for a long time, it might become less reliable on rainy days. In safety-critical systems, forgetting is not just an annoyance. It can be a risk, which is why these systems are tested very carefully, and why they are usually trained on a wide mix of conditions at once, rather than one condition at a time.
How Researchers Try to Stop the Forgetting
If forgetting is such a basic problem, how do researchers deal with it? There is no single perfect fix, but there are several clever ideas, and most modern approaches combine more than one of them. Each idea can be explained with a simple picture.
The first idea is replay, also called rehearsal. Instead of training only on new examples, the system is also shown old examples again, mixed in with the new ones. This is like a student who, when learning a new chapter, also reviews earlier chapters a little each week. By revisiting old examples, the system keeps the old dials close to their useful positions, even while it learns new things. Replay is simple and often works well, but it has limits. Keeping and reusing old data can raise privacy questions, especially if the old data includes personal information. It can also be expensive, because training takes longer when old examples are included every time.
The second idea is to protect the most important parts. Researchers try to work out which dials matter most for old tasks, and then make those dials harder to change during new training, while letting other dials change more freely. A well-known version of this idea is called elastic weight consolidation, but the plain-language picture is enough: imagine marking the most important pages of your notebook with a sticky note that says “do not erase”. The system can still learn, but it must learn around those protected parts. This can reduce forgetting, but it is not perfect. If the new task really needs to change a protected part, the system may learn more slowly, or may need a different approach.
The third idea is to give each task its own space. Instead of forcing every task to share exactly the same dials, the system gives new tasks some new, separate parts to use, while keeping old parts mostly untouched. This is like adding a new section to a notebook, rather than writing over old pages. It can work very well for keeping old skills intact, but it has a cost: the system grows larger with each new task, and at some point it becomes too large, too slow, or too expensive to keep growing forever. Researchers therefore try to share parts where sharing is safe, and separate parts where sharing would cause forgetting.
The fourth idea is the simplest, and often the most reliable: retrain from scratch, or from a broad base, on a mix of old and new data. Instead of teaching an old model only the new task, the company trains a new version on a large, mixed set of examples that includes both old and new tasks. This is like a student starting a new revision course that covers the whole subject, not only the latest chapter. It avoids much of the forgetting problem, because the system never learns the new task in isolation. The downside is cost. Retraining a large modern model from scratch can take a huge amount of computing power, time, and money, and it may not be practical every time a small update is needed.
What This Means for Everyday Users
For most people, the technical details matter less than the practical question: what should I expect, and what should I do, when AI seems to forget?
First, it helps to keep short-term and long-term forgetting apart. If a chatbot forgets a detail from earlier in a long chat, the simplest fix is often to repeat the important detail, or to start a fresh chat with a clear summary at the top. You can say, for example, “We are writing a story about a sailor named Maya, who lives in a coastal town, and the story should be funny.” Giving the system a short summary at the start of a new chat often works better than expecting it to remember every detail from a very long earlier chat.
Second, it helps to save important work outside the chatbot. If you are writing a story, planning a trip, or working on a project with an AI assistant, keep your own copy of the key facts, names, dates, and decisions. Do not rely on the chatbot as your only memory. This is good advice even if the chatbot worked perfectly, because chats can be lost, accounts can change, and products can be updated.
Third, it helps to test important answers, especially when they matter. If an AI system gives you medical, legal, financial, or safety information, check it against a trusted source, or ask a qualified professional. Forgetting is one reason AI answers can be wrong, but it is not the only reason. AI systems can also make mistakes, mix up similar facts, or give confident-sounding answers that are out of date. A system that has forgotten an old fact may not warn you that it has forgotten. It may simply give a wrong or incomplete answer with the same confidence as a right one.
Fourth, it helps to understand why products change after updates. If a favourite AI tool suddenly behaves differently, it may have been retrained, or its rules may have changed, or both. If the change matters to you, look for the company’s update notes, or ask its support team what changed. You may not get a detailed technical answer, but you can at least find out whether the change was deliberate, or whether it is a bug that should be reported.
Fifth, it helps to be patient with the idea of progress. AI systems are improving, and researchers are making real progress on forgetting. Larger models, better training mixes, and better testing have reduced the problem in many cases. But progress does not mean the problem has disappeared. A system that seems to remember everything today may still forget after a future update, and a system that forgets today may be improved tomorrow. Expecting AI to remember perfectly, like a perfect human memory, sets up disappointment. Expecting it to be useful, while checking important details yourself, is a more realistic approach.
Limits and Honest Unknowns
A fair article about AI forgetting should end with what we do not know, as well as what we do know. Researchers have learned a lot since 1989, but several important questions remain open.
Another open question is how to handle many tasks over a long time. Most laboratory tests use a small number of tasks, such as two, five, or ten. Real products may need to handle hundreds of tasks, over months or years, with new data arriving all the time. Methods that work for a few tasks may become too slow, too large, or too expensive when the number of tasks grows. Researchers are still working out how to scale continual learning to real-world size.
None of this means AI is useless, or that forgetting makes AI untrustworthy in every case. It means AI is a tool with known limits, and that using it well means understanding those limits. The best AI products are those that are tested carefully, updated transparently, and used by people who check important answers, save important work, and keep learning themselves.
Frequently Asked Questions
Can AI forget what it learned?
Yes. When a neural network is trained on a new task, especially if it is trained only on new examples, it can lose much of what it learned from earlier tasks. Researchers call the most dramatic version of this catastrophic forgetting. The effect was shown clearly by McCloskey and Cohen in 1989, and it remains a real challenge today, even though modern methods reduce it.
Is a chatbot forgetting in a chat the same as catastrophic forgetting?
Usually, no. Forgetting in a long chat is often caused by limited short-term context: the chatbot can only keep a certain amount of recent conversation in view, so older details fall away. Catastrophic forgetting, by contrast, happens when the model’s stored patterns themselves change during retraining, so it becomes worse at an old task even in a fresh conversation.
Why does AI forget more easily than people do?
People and AI store knowledge differently. Human memory uses many systems, and learning a new skill does not usually erase an old skill completely. AI knowledge is often stored as shared numbers across the whole system, so changing those numbers to learn something new can overwrite old knowledge. Human forgetting is usually gradual. AI forgetting, when it happens during sequential training, can be sudden and large.
How do researchers stop AI from forgetting?
They use several ideas, often together. Replay shows old examples again along with new ones. Protecting important parts makes the most useful dials harder to change. Giving each task its own space reduces sharing that causes overwriting. Retraining from scratch on a broad mix of old and new data avoids learning the new task in isolation, but it is expensive. No method removes forgetting completely in all cases.
Should I trust an AI system that might forget?
Use it as a helpful tool, not as a perfect memory. Save important facts outside the chatbot, repeat key details in long chats, and check important answers against trusted sources, especially for health, money, law, or safety. If a product changes after an update, check its update notes, and report clear mistakes. AI can be very useful, even though it can forget.
Closing: Forgetting Is a Clue, Not Just a Flaw
So, can AI forget what it learned? Yes, and the fact that it can forget tells us something important about how AI works. AI does not remember like a person. It learns patterns, stored as numbers, and those patterns can be reshaped by new training. When the reshaping is careful, and old examples are kept in the mix, the system can learn new things while keeping old ones. When the reshaping is careless, or when new training is done in isolation, the old patterns can fade fast.
That is why researchers talk about a tug of war between stability and plasticity, between remembering and learning. Human brains handle that tug of war well, though not perfectly. AI systems are still learning how to handle it, and every improvement brings new trade-offs in cost, size, privacy, and speed.
For everyday users, the lesson is simple and practical. Enjoy what AI can do, but do not treat it as a perfect memory. Keep your own notes, check important answers, and expect products to change after updates. For researchers and companies, the lesson is harder: build systems that can keep learning over time, test them on old tasks as well as new ones, and be honest about what they have forgotten, as well as what they have learned.
