A ‘Delete Button’ for the AI Brain? If AI Knows These Things, Could It Become Dangerous?

— Surya Prakash Josyula

Imagine a new virus is spreading across the world.

Scientists are working day and night to develop a medicine for it. At the same time, another person is sitting in front of an AI system, searching for more dangerous information about the same virus.

Both are asking for the same information.

But in one person’s hands, it can save lives. In another person’s hands, it can become a deadly weapon.

A knife can save a life in a doctor’s hands. The same knife can take a life in the hands of a criminal. Information is no different. The same AI knowledge can help a scientist develop a cure, but it can also become dangerous if used with bad intentions.

This is the biggest challenge the AI industry is facing today. That is why researchers are now working not only on teaching AI new things, but also on making it “forget” certain dangerous knowledge when needed.

How Is This ‘Delete Surgery’ Possible?

Normally, if developers wanted to remove one dangerous topic from an AI model, they had to retrain the entire model from the beginning. This process costs millions of dollars and takes several months.

Now, researchers at Anthropic have developed an experimental method that stores sensitive information separately and removes it whenever necessary.

For this, they created a new training method called GRAM.

Think of it like deleting just one folder instead of formatting your entire computer. Or removing one dangerous book from a library instead of burning down the whole library.

The ‘Dual-Use’ Problem Explained with Virology

Virology, the science of viruses, is a perfect example of why this technology is needed.

Virology helps scientists develop medicines and life-saving vaccines. But the same knowledge could also be misused to create dangerous viruses.

This is called dual-use knowledge—information that can be used for both good and bad purposes.

With the GRAM method, such sensitive knowledge is stored in separate neural modules while training the AI. If required, only that specific module can be removed.

The Most Interesting Part

The most surprising part is that after removing the dangerous information, the AI continued to perform almost the same in other tasks.

Researchers tested this on an AI model with 5 billion parameters. They found that once the module was removed, it was almost as if the AI had never been trained on that information, while its overall performance remained almost unchanged.

Quick FAQ

What is GRAM?

It is a new AI training method that stores sensitive information separately.

Why was it developed?

To remove dangerous knowledge and make AI safer.

Will AI performance decrease?

According to the researchers, their tests showed no significant drop in the AI’s overall performance.

Benefits of This Technology
It can reduce the misuse of dangerous information.
It can make AI safety easier to improve in the future.
Companies may not have to retrain an entire AI model for small changes.
It could help AI companies comply more easily with future government regulations.
The Biggest Question: Who Decides What Is Dangerous?

While many experts have welcomed this technology, it has also raised an important question.

Who decides what should be considered “dangerous information”?

Will it be limited only to scientifically harmful topics? Or could governments or companies one day remove information they simply do not like?

This has now become a major debate in the AI world.

Researchers also point out that this work is still at an early stage. It may take time before such methods are widely used in large language models.

Although experts believe similar safety techniques may appear in AI systems like ChatGPT and Gemini in the future, there have been no official announcements so far.

Bottom Line

Until now, the biggest question was:

“How much can AI learn?”

In the coming years, the bigger question may become:

“What should AI remember… and what should it forget?”