Artificial intelligence has a memory problem.
Once a large language model learns something during training, removing that knowledge is incredibly difficult. Developers have traditionally relied on safety guardrails, content filters, and reinforcement learning to stop models from producing dangerous information. But those approaches don’t actually erase what the AI has learned—they simply try to prevent it from revealing it.
Now, Anthropic and research partner AE Studio have introduced a different idea.
Instead of teaching an AI everything in one giant neural network, their new technique—called Gradient-Routed Auxiliary Modules (GRAM)—stores certain high-risk knowledge inside separate, removable components. If those components are deleted, the model behaves almost as if it never learned that information in the first place.
If the approach proves effective at larger scales, it could reshape how future AI systems are deployed for governments, enterprises, researchers, and the public.
Table of Contents
What Is Anthropic’s GRAM?
GRAM stands for Gradient-Routed Auxiliary Modules.
Think of a language model as a massive digital library.
Normally, every new piece of information gets mixed into the same collection of books. Removing one subject later becomes nearly impossible because the knowledge is spread throughout the library.
GRAM changes that idea.
Instead of mixing everything together, the model stores sensitive subjects inside dedicated “rooms.” Those rooms contain knowledge related to areas such as:
- Virology
- Cybersecurity
- Nuclear science
- Specialized programming
If an organization doesn’t need access to one of those subjects, developers can simply remove the corresponding module.
The rest of the AI continues working normally.
Why This Is Different From Current AI Safety Methods
Today’s AI companies mainly rely on three approaches:
- Safety filters
- Reinforcement learning
- Refusal policies
These methods attempt to stop the AI from answering harmful questions.
However, researchers have repeatedly shown that determined users can sometimes bypass these safeguards through prompt engineering or jailbreak techniques.
GRAM tackles the problem much earlier.
Instead of asking the AI to hide dangerous knowledge, it aims to prevent that knowledge from existing inside the deployed model at all.
That represents a fundamentally different philosophy for AI safety.
How Does GRAM Work?
During training, researchers assign sensitive information to dedicated auxiliary neural modules.
Whenever the model encounters data from a protected category, those specialized modules learn alongside the model’s general capabilities.
After training, developers can choose whether to include or remove each module.
According to the researchers, deleting a module removes the associated capability while leaving unrelated skills largely unaffected.
For example, an AI could retain its writing ability, reasoning skills, and coding assistance while losing specialized expertise in advanced virology.
Why Researchers Believe This Matters
One of the biggest challenges in AI development is balancing usefulness with safety.
Scientists, hospitals, cybersecurity teams, and government agencies often require advanced technical knowledge that would be inappropriate for general public deployment.
GRAM could make it possible to distribute different versions of the same AI model.
For example:
- Medical researchers could receive models containing advanced biological knowledge.
- Security professionals could access cybersecurity expertise.
- General consumers could use versions where high-risk domains have been removed.
Instead of maintaining entirely separate AI systems, companies could theoretically enable or disable specific knowledge modules.
Does GRAM Actually Work?
Anthropic’s research suggests promising early results.
The team tested models ranging from tens of millions to several billion parameters.
According to the published findings, removing a GRAM module reduced the model’s performance on that specific domain to nearly the same level as if it had never been trained on the information.
At the same time, general language understanding and reasoning remained close to models trained with the complete dataset.
The researchers also reported that GRAM appeared more resistant to adversarial fine-tuning than existing “machine unlearning” methods, which often suppress knowledge rather than removing it completely.
These results are encouraging, but they come with an important caveat.
The Biggest Questions Still Remain
Despite the excitement, GRAM is far from a finished solution.
The research has not yet been implemented in Anthropic’s production Claude models.
Several important questions remain unanswered.
Can it scale?
Modern frontier AI models contain hundreds of billions of parameters.
A method that works on smaller research models may behave differently at much larger scales.
Can knowledge really be separated?
Subjects like biology and virology overlap significantly.
Removing only the dangerous aspects without affecting legitimate scientific understanding could prove extremely difficult.
Will attackers find new ways around it?
History suggests that every AI safety improvement eventually faces attempts at circumvention.
Whether GRAM remains robust against future attacks remains to be seen.
Why This Could Change AI Policy
Governments worldwide are debating how advanced AI systems should be regulated.
Some policymakers have proposed restricting access to entire models because they contain potentially dangerous capabilities.
GRAM introduces another possibility.
Instead of banning complete systems, regulators could require companies to remove certain high-risk knowledge before public deployment.
Organizations with legitimate needs—such as universities, pharmaceutical researchers, or government agencies—could receive versions where those modules remain available under controlled access.
If successful, that approach could offer a more flexible balance between innovation and security.
Could Other AI Companies Adopt Similar Technology?
Although GRAM comes from Anthropic and AE Studio, the underlying concept is not limited to one company.
If the research continues to show positive results, similar modular training techniques could eventually appear across the AI industry.
Companies developing foundation models—including OpenAI, Google DeepMind, Meta, and others—are all exploring ways to improve AI safety without reducing model usefulness.
Whether modular knowledge becomes a common design principle will depend on future research and real-world testing.
Final Thoughts
Anthropic’s GRAM isn’t just another AI safety announcement.
It challenges one of the biggest assumptions in machine learning—that once an AI learns something, that knowledge is effectively permanent.
If researchers can reliably isolate and remove dangerous capabilities without harming general intelligence, AI deployment could become far more customizable than it is today.
For now, GRAM remains an experimental research project rather than a production-ready solution.
But if future studies confirm its effectiveness at the scale of today’s most advanced models, it could become one of the most significant developments in AI safety in years.
Recommended article :