The Familiar Map

Most people who work with software know the familiar map.

MIT means one kind of permission. Apache means another familiar permissive bargain, with attribution, notice, and a patent grant. GPL means reciprocity: use it, study it, modify it, share it—but when you distribute the resulting work, the freedoms are supposed to travel with it. AGPL carries that idea into a networked world, responding to the fact that software can be modified and offered as a service without ever being handed to the user as a copy.

The map is not effortless. Lawyers still argue about linking, derivative works, compatibility, patents, notices, and distribution. But the basic landmarks are recognizable. A developer can look at a license and begin to understand the bargain.

Then came a new class of artifacts.

A model can be downloaded without its training data. Its weights can be fine-tuned without its training code. Its inference code can be published while its architecture, checkpoints, evaluation process, or data provenance remain obscure. A company can call a model “open source” while retaining restrictions on commercial scale, fields of use, redistribution, or acceptable behavior.

The familiar map has not disappeared. It has become incomplete.

A word with too much work to do

“Open” is now asked to describe almost everything that is more available than a closed API.

A model may be open because its weights can be downloaded. A dataset may be open because it is publicly accessible. A benchmark may be open because the evaluation code is on GitHub. A company may be open because researchers can apply for access. A hosted service may be open because anyone can sign up.

These are different conditions. They create different capabilities and different dependencies.

If I can download a model but cannot inspect how it was trained, I have a useful artifact—but not necessarily a reproducible system. If I can inspect the source but cannot use it commercially, I have visibility—but not the ordinary freedom associated with open source. If I can run an API but cannot obtain the model, I have access to a service—but not control over the technology. If a dataset is downloadable but its provenance and license conditions are unclear, I have possession—but not necessarily lawful or responsible reuse.

The word “open” does not answer these questions. It only invites them.

This is not an argument against partial openness

There is an easy mistake to make here. Once we see that “open source” is being used loosely, we may decide that every partial release is dishonest or worthless.

That would be wrong.

Open weights can be valuable. They can enable local execution, privacy, experimentation, fine-tuning, and competition. A source-available project can be easier to audit than a proprietary one. A research-access program can let people study a system they could never independently train. A benchmark can improve public knowledge even when the model being measured is closed.

The problem is not that these forms of access exist. The problem is that they are often described as though they provide the same freedoms.

OSI’s Open Source AI Definition offers one important boundary. It describes Open Source AI through the freedoms to use, study, modify, and share, and identifies data information, code, and parameters as the preferred form for modifying machine-learning systems.¹ The definition does not make every partial release bad. It gives readers a way to say what is present and what is missing.

That distinction matters because a spectrum without boundaries becomes a blur. We need a spectrum of openness, but we also need terms that mean something.

The spectrum is not a ladder

The spectrum should not be imagined as a single ladder from closed to open. It is better understood as a profile.

A system can be highly accessible but weakly reproducible. It can be highly reproducible but difficult to deploy. It can be permissively licensed at the model layer while depending on data with complex upstream conditions. It can be open to modification while imposing safety restrictions on use. It can be transparent today but vulnerable to unilateral relicensing tomorrow.

Ask at least these questions:

• Can I download the artifact?
• Can I run it locally?
• Can I study its components?
• Can I modify it?
• Can I use it commercially?
• Can I redistribute it?
• Can I inspect the training process?
• Can I understand the data provenance?
• Can I reproduce a substantially equivalent result?
• Can I preserve the permissions if the original provider changes direction?
• Can I participate in governing the project?
• Is there a reciprocal obligation to share improvements or benefits?

A model’s openness is therefore not a single number. It is a shape.

Why AGPL is the right historical analogy

The AGPL story helps explain why this moment is not unprecedented.

GPL was written for a world where distribution was the important event. If you gave someone a copy of a modified program, the license could require you to provide source code and preserve the user’s freedoms. But network services changed the practical meaning of distribution. A company could modify a program and let people interact with it remotely, arguing that it had not distributed the program at all.

AGPL responded to that gap by treating network interaction as a trigger for source-sharing obligations.²

The lesson is not that AI requires an “AI AGPL” by default. The lesson is that a new technology or delivery model can expose a gap in an existing bargain. When that happens, the community has choices:

• reinterpret existing licenses;
• create a new license;
• define new artifacts and disclosure obligations;
• build technical mechanisms for attribution and provenance;
• or accept that some forms of openness cannot be captured in one legal instrument.

AI raises analogous questions. What is the source of a trained model? What is the preferred form for modifying it? What does reciprocity mean when the valuable inputs include data, compute, evaluation, and community feedback? Does a restriction intended to prevent harm preserve openness, or does it change the category entirely?

These are not merely philosophical questions. They determine what users can build.

The first rule of the new map

Do not ask whether a system is open before asking what the system is.

Is it code? Weights? A model? A dataset? A benchmark? An API? A complete application? A bundle of all of these?

Then ask what has actually been released and under what terms.

That is the first rule of the new map:

Name the artifact before you name its openness.

A release that provides weights under a permissive license may be an excellent open-weight release. A release that provides code, data information, training code, parameters, and broad permissions may qualify as Open Source AI under the OSI definition. A release that publishes source but prohibits competitive services may be source available. A release that offers only an API is a service, not an open model.

None of these descriptions is an insult. They are navigation aids.

What this book will do

This book will trace the evolution from the familiar licenses to the new AI landscape. It will explain why MIT, Apache, GPL, and AGPL remain important; why source availability and open weights deserve precise names; how institutions such as OSI, Creative Commons, and MLCommons are approaching the problem; and what recent legal disputes reveal about permissions, attribution, enforcement, and reproducibility.

It will compare real releases rather than debating abstractions. Ai2’s OLMo release, for example, describes training data, training code, model weights, logs, checkpoints, inference code, evaluation code, and Apache 2.0 licensing.³ Other releases make different tradeoffs: they may provide useful weights while using custom terms or acceptable-use policies.

It will also ask what openness requires beyond a license. MLCommons’ work on datasets and Croissant points toward metadata, provenance, discoverability, and governance as part of the infrastructure of reuse.⁴ Creative Commons’ work on AI and the commons similarly asks how attribution, agency, and responsible reuse should function when content is consumed by machines at scale.⁵

The goal is not to produce one final definition that ends the argument.

The goal is to make the argument more useful.

The question

When a new system arrives with the word “open” attached to it, pause before accepting the label.

Ask:

What has been opened, to whom, under what terms, with which freedoms, and with what ability to continue without permission?

That question is not a rejection of the new. It is the beginning of understanding it.