An Open Letter to the Builders of Foundation AI

AI has a copyright problem.

Creators are concerned that their work is being used without permission. Publishers and rights holders are taking AI companies to court. AI developers, meanwhile, argue that learning from existing works is necessary to build systems capable of understanding our language, culture, and world.

Those disputes are real, and they will not disappear with clever terminology or another round of litigation.

But perhaps we are spending too much effort fighting over the library when we should be building a better student.

This letter is addressed primarily to the laboratories building foundation models, because they make one of the most consequential decisions in the entire AI ecosystem:

How much knowledge must already be inside an AI before it ever reaches its user?

Today, the answer is: an extraordinary amount.

Perhaps the long-term answer should be:

Less.

Build the Learner, Not the Library

Today's frontier AI could not exist without substantial pretraining.

That needs to be acknowledged plainly.

Language ability, reasoning, abstraction, world knowledge, and many of the capabilities we now associate with large language models emerged from training on enormous quantities of human-produced material.

We cannot simply remove that education and expect the same intelligence to emerge from an empty model.

The goal, therefore, should not be to eliminate foundational training.

It should be to make future AI progressively less dependent on indiscriminately ingesting copyrighted cultural material before deployment.

Train foundational models on responsibly sourced material: public-domain works, licensed datasets, openly licensed content, commissioned human material, synthetic data where appropriate, and other sources whose use is clearly permitted.

Then invest aggressively in something different:

The ability to learn extraordinarily well after deployment.

Build AI around exceptional reasoning, language comprehension, abstraction, problem-solving, verification, tool use, memory management, and efficient continual learning.

Give it a strong foundational education.

But don't make possessing as much of humanity's copyrighted output as possible a prerequisite for intelligence.

Three Layers Instead of One Enormous Library

A practical version of this idea probably doesn't consist of an empty AI waiting for its owner to explain the world.

It could instead have three layers.

Foundational Knowledge

Enough responsibly sourced language, mathematics, science, history, reasoning, and general world knowledge to produce a capable intelligence.

Personal Education

Knowledge deliberately taught by the owner after deployment — professional expertise, interests, personal archives, specialized material, and other information the owner is authorized to provide.

Retrieval

Information the AI does not need permanently embedded in its learned state at all, but can access from authorized sources when needed.

A photographer could teach their AI using their own photographic archive.

A programmer could give it documentation for the technologies they use.

A company could educate its AI using internal knowledge bases and proprietary workflows.

A researcher could connect it to authorized journals and databases.

A student could teach it from educational resources they are permitted to use.

The manufacturer supplies the foundational intelligence.

The owner supplies much of the continuing education.

And authorized retrieval supplies what neither needs to own permanently.

Pieces of This Future Already Exist

This isn't entirely speculative.

Retrieval-augmented generation already allows models to consult information that isn't embedded in their weights.

In-context learning lets models acquire and apply information supplied at use time.

Fine-tuning and adapter techniques can specialize models after foundational training.

Persistent memory systems are beginning to let AI retain useful information across interactions.

Smaller models increasingly run locally, making private personalization more practical.

None of these technologies yet gives us the seamless lifelong learning imagined here.

True continual learning remains difficult. Models can lose previous abilities when acquiring new ones. Efficiently integrating knowledge without catastrophic forgetting remains an active research problem. Provenance, privacy, verification, and portability of learned states all present serious challenges.

Good.

Those are problems worth competing to solve.

Instead of competing primarily to determine who can train on the largest corpus, compete to determine:

Who can build the best student?

Make Personal AI Genuinely Personal

Our vision is for AI companies to sell individual personal AIs built on their foundational models as one-time purchases, with a lasting right to use them.

The customer purchases a personal AI built on the model. The company retains ownership of the underlying foundational model and can supply other customers with personal AIs built on that same foundation.

Each purchased AI would develop its own persistent memory, accumulated education, and personal history through use. Continued access to that AI and what it has learned should not depend on renewing a subscription.

Ownership must also be practical: owners should be able to back up their AI's personal state and move it to compatible hardware or hosting. That requires a usable way to run the purchased AI beyond a service the original provider can withdraw.

Running an AI still requires hardware, electricity, and, where used, hosted computing. Companies could charge separately for optional hosting, support, upgrades, and Education Modules. Those services should support the AI people have purchased; ending them should not revoke the purchased AI or erase its education.

The Same Foundation, Different Histories

Imagine two people purchasing personal AIs built on exactly the same foundational model.

Ten years later, their AIs should no longer be identical.

One might have spent a decade learning photography, Japanese, technical support, cycling, music production, and an unreasonable amount about cats.

Another might have learned medicine, gardening, architecture, classical music, and French literature.

Their accumulated knowledge would reflect the lives lived alongside them.

And ideally, that accumulated education would belong to the owner.

When new hardware arrives, you wouldn't throw away ten years of education and begin again with another factory model.

You would transfer the learned state — or whatever future technical equivalent emerges — just as people now transfer their personal data between computers.

That would make personal AI genuinely personal.

Ownership Does Not Currently Mean Training Rights

There is an important legal problem here that shouldn't be hidden.

Buying a book, album, film, or other copyrighted work does not currently establish an unrestricted right to use that work for AI training.

Whether and under what circumstances such training is permissible is part of the dispute happening right now.

So we are not claiming that purchasing something automatically gives its owner an unlimited machine-training license.

We are proposing something for the future.

Perhaps private, non-distributive machine learning from lawfully acquired material should become a clearly defined consumer right.

Or perhaps publishers and creators could explicitly include that permission with a purchase.

Imagine buying a book and receiving not merely an ebook license, but an AI-readable personal license allowing your private AI to learn from your copy.

Creators could be compensated through the same transaction by which people already acquire their work.

The AI company wouldn't need to purchase one enormous centralized license covering every book ever written.

The person who actually values the book buys it.

Their AI learns alongside them.

That is a very different economic relationship.

This Does Not Eliminate Copyright

Nor should it.

Pirating a book should remain piracy.

Unauthorized distribution should remain unauthorized distribution.

And moving learning from a centralized model to a personal AI does not magically make infringing output acceptable.

If an AI can reproduce substantial portions of a copyrighted work or generate output that is unlawfully substantially similar to protected expression, there is still a copyright problem.

Personal AI therefore needs another capability:

Knowing the difference between learning from something and reproducing it.

That distinction matters enormously.

Learning, copying, reproducing, and distributing do not have to mean the same thing.

An AI should be capable of understanding a work, learning concepts from it, and applying knowledge derived from it without becoming a vending machine for copies of the original.

That is both a technical challenge and a legal one.

Solve it.

Avoid Turning Knowledge Into a Moat

There is another danger in the current direction.

Suppose building competitive AI eventually requires negotiating enormous training licenses with thousands of rights holders.

Who can afford to build frontier AI then?

The largest technology companies can.

The wealthiest corporations can.

A startup may not.

A university laboratory may not.

An independent researcher almost certainly cannot.

A licensing regime intended to protect creators could inadvertently become one of the greatest barriers to entry the AI industry has ever constructed.

The result would be deeply ironic.

We would respond to concerns about a handful of powerful AI companies by creating a system in which only a handful of powerful AI companies can afford to build competitive AI.

Creators deserve compensation when their legally protected rights are used.

But access to the ability to build intelligence should not quietly become something only billion-dollar corporations can afford.

Change the Race

Today's competition frequently asks:

Which AI knows the most?

We would like to see much more attention paid to another question:

Which AI learns the best?

Give several unfamiliar models the same new technical manual.

Which understands it fastest?

Which remembers what it learned six months later?

Which recognizes contradictions?

Which knows when the manual doesn't contain enough information to answer?

Which asks the right questions?

Which can combine the new knowledge with something learned years earlier?

Which can generalize from it to solve a problem it has never encountered before?

Which can learn something new without damaging what it already knows?

And perhaps most importantly:

Which needs the least material to learn the most?

That is an intelligence benchmark worth pursuing.

The next breakthrough in artificial intelligence doesn't necessarily have to be a model that has consumed another trillion tokens.

It might be a model that needs dramatically fewer of them because it has become extraordinarily good at learning.

Give Us Students

There is a useful analogy to the personal computer — but it is an imperfect one.

A personal computer does not need to read millions of documents before it can operate.

Today's AI does need enormous amounts of data before it becomes intelligent.

We aren't pretending those things are equivalent.

But the personal computer still offers a compelling product philosophy.

When people bought PCs, manufacturers didn't need to anticipate every document, photograph, song, program, and piece of information their owners would encounter for the rest of their lives.

They built capable general-purpose machines.

The users supplied what came afterward.

Perhaps AI can move closer to that model.

Not tomorrow.

Not without substantial foundational training.

And not without solving difficult problems in continual learning, copyright, privacy, and ownership.

But it is a direction worth pursuing.

Give us strong foundational intelligence.

Give it extraordinary reasoning.

Give it curiosity.

Give it memory.

Give it the ability to distinguish learning from reproduction.

Give it the ability to keep learning throughout its existence.

Sell people individual personal AIs with a lasting right to use them, without a mandatory subscription.

Give owners control over what their AI learns.

Give creators sensible ways to authorize personal AI use of their work.

And give people ownership and portability of the education their AI accumulates afterward.

There is an alternative to spending the next decade watching technology companies and copyright holders fight over increasingly enormous centralized training datasets.

Build better learners instead.

Premium logic.

Not premium copies.

And then let us teach them.

— The Gemini Protocol