AI Model Licenses Explained: What Kids Can Legally Do

An AI model license is the legal agreement that decides what you can do with a model you’ve downloaded: run it, modify it, redistribute it, build a product on it, or none of the above. It is not the same thing as whether you can get the file.

Google’s Gemma family passed one billion downloads and, per Google’s August 2026 roundup, “supports environments from phones and edge infrastructure to space.” That number makes it sound maximally open. But Gemma is distributed under Google’s own Terms of Use, not a standard open-source license, and the Open Source Initiative’s definition of Open Source AI requires things most “open” models don’t provide.

If your kid downloads a model for a science fair project, a school assignment, or a first startup, the license is the difference between fine and a problem. It’s also, unexpectedly, one of the best civics lessons available right now.

Key Takeaways

• “Open weights” means you can download the parameters. “Open source” is a stricter standard requiring the training code, parameter values, and enough data information that a skilled person could build a substantially equivalent system.
• Gemma’s terms permit commercial use, redistribution, fine-tuning, and derivative models, but require you to pass the use restrictions to downstream users, include the Agreement, mark modified files, and include a specific Notice.
• Gemma’s terms are also subject to a separate Prohibited Use Policy, and Google reserves the right to restrict usage remotely.
• Truly permissive licenses exist: Apache 2.0 and MIT allow essentially anything with attribution. Some Gemma versions use Apache 2.0, which is a meaningfully different arrangement.
• The teaching frame: a license is a promise about what happens next, and reading one is a skill a 13-year-old can genuinely learn in an afternoon.

What the Open Source Initiative actually requires

The Open Source Initiative has published an Open Source AI Definition, version 1.0, and it’s short enough to read with a teenager.

It requires four freedoms: the freedom to “use the system for any purpose and without having to ask for permission,” to “study how the system works and inspect its components,” to “modify the system for any purpose, including to change its output,” and to “share the system for others to use with or without modifications.”

And it specifies three components that must be available:

Data information. “Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.”

Code. “The complete source code used to train and run the system.”

Parameters. “The model parameters, such as weights or other configuration settings.”

That second requirement is the one most so-called open models fail. Publishing weights is now common. Publishing the complete training code and detailed data information is rare, because the data is legally fraught and the training pipeline is a competitive asset. It’s also why labs guard against having their models copied indirectly, a concern named explicitly in Anthropic’s June 9, 2026 Fable 5 announcement.

So there’s a spectrum, and the vocabulary matters:

Closed API. You send requests, you get responses, you never touch the model. Claude, GPT, and Gemini’s frontier models work this way.

Open weights, restricted license. You download the parameters and agree to a custom contract with conditions. Gemma’s Terms of Use fit here.

Open weights, permissive license. Apache 2.0 or MIT. You can do nearly anything, including commercial use, with attribution.

Open source per the OSI definition. Weights plus complete training code plus data information. Genuinely rare.

What Gemma’s terms actually say

I read the Gemma Terms of Use so your family doesn’t have to, and the structure is worth knowing.

What’s permitted: commercial use is allowed, redistribution is allowed if you meet the conditions in section 3.1, fine-tuning falls under permitted modifications, and “Model Derivatives” are explicitly addressed. That’s a genuinely useful set of permissions, and it’s why Gemma got to a billion downloads.

What’s required when you redistribute:

1. Include the use restrictions as “an enforceable provision” in your downstream agreements.
2. Provide recipients a copy of the Agreement.
3. Mark modified files with “prominent notices.”
4. Include a Notice file stating: “Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.”

That first condition is the one that distinguishes this from open source. The restrictions travel. Anyone you give your fine-tuned model to inherits them, and you’re responsible for passing them along.

There’s also a separate Prohibited Use Policy that applies, a requirement not to violate applicable law, and a clause under which Google reserves rights to restrict usage remotely.

One important nuance: the Gemma terms document references a separate Gemma 4 license under Apache 2.0, so different versions carry different terms. That’s an argument for checking the license of the specific version rather than assuming “Gemma is licensed like X.”

Nothing here is sinister. It’s a company distributing something valuable with conditions attached. But “one billion downloads” and “open source” are different claims, and a kid who can tell them apart has learned something real.

Why a kid should care before their first project

Four concrete situations where this decides something:

A science fair project. Most fairs require you to document tools and sources. A license that requires a Notice file means you include it. Easy, and it’s a disqualification risk if you skip it.

Publishing a fine-tuned model. The moment your kid uploads a model they modified, redistribution conditions apply. Notice file, passed-along restrictions, marked modifications.

A first business. Commercial use permitted under Gemma’s terms is a real green light, and the downstream-restriction requirement is a real obligation. Both matter to a 17-year-old building something.

School and district deployments. When a district picks a model, the license shapes whether they can self-host, modify, and control data. Contrast that with a closed API, where published per-token pricing is the whole commercial relationship and the weights never leave the provider. That’s a big part of why open-weight models matter for education at all. Our piece on Gemma’s billion downloads and edge deployment covers the reach, and open-source vs. proprietary AI covers the trade-offs.

There’s a fifth reason, less practical and more valuable: reading a license is reading a contract, and most people never learn how. A one-page license with four numbered conditions is the friendliest possible entry point.

How to Teach Your Kid About AI Model Licenses

Licenses are rules about sharing, which every kid has strong opinions about already.

Ages 5–8: The lending rules

Have your kid lend a toy to a sibling or friend and set the rules out loud first: Can they take it home? Can they change it? Can they lend it to someone else? Do they have to say who it came from? Write the rules on paper and have both sign. Then test one: the borrower lends it to a third kid. Was that allowed? Say: “That’s a license. Grown-ups write these for computer programs, and the rules work the same way.”

Ages 9–12: Sort the licenses

Make four cards: “you can only use it, not download it,” “you can download it but must follow extra rules and pass them on,” “you can do almost anything if you say where it came from,” and “you get everything including the recipe.” Then give them four scenarios and have them match. Tell them the real names afterward: closed API, open weights with a custom license, Apache 2.0 or MIT, and OSI open source. Kids handle this faster than adults because they haven’t learned to be intimidated by legal language.

Ages 13+: Read a real license and write a summary

Have your teen open the Gemma Terms of Use and produce a one-page summary answering: Can I use this commercially? Can I redistribute it? What exactly must I include if I do? What policy applies on top? Then have them read the OSI Open Source AI Definition and answer whether Gemma meets it, with a reason. That’s a genuine legal-reading exercise with a real document, and it takes about an hour.

The question to ask: “If you made something using this and gave it to a friend, what would you have to tell them?”

License to allowed uses: a comparison

License type	Download weights?	Commercial use?	Redistribute?	Must pass on restrictions?	Meets OSI Open Source AI?
Closed API (Claude, GPT, Gemini frontier)	No	Per terms of service	No	N/A	No
Gemma Terms of Use	Yes	Yes	Yes, with conditions	Yes, as an enforceable provision	No; lacks full code and data information
Apache 2.0 (incl. some Gemma versions)	Yes	Yes	Yes	No; attribution and notice only	Only if code and data information also published
MIT	Yes	Yes	Yes	No; keep the copyright notice	Only if code and data information also published
Custom “community” licenses	Yes	Usually, with thresholds	Yes, with conditions	Usually yes	Generally no
OSI Open Source AI	Yes	Yes	Yes	No	Yes, by definition

Two things to notice. Downloading the weights and being open source are separate columns, and most models that people call “open” sit in the middle rows. Also, license terms change between versions of the same model family, which is why “check the version” is the actual advice.

What to do at home

Check the license before the download, not after

Thirty seconds on the model card or terms page. If your kid can’t find a license, that itself is information. This habit prevents the specific bad outcome where a project is built and then can’t be shared.

Teach the four-freedoms test

Can you use it for anything? Study it? Modify it? Share it? Those four questions from the OSI definition let a kid classify any license quickly, and they work for software generally, not just AI.

Keep the Notice file requirement in mind for projects

If your kid publishes anything built on Gemma, the terms require a specific Notice. Making that a checklist item is thirty seconds and avoids a real problem.

Use license reading as contract-reading practice

This is the sleeper benefit. Your kid will sign dozens of agreements in the next decade, and almost none of them will be as short or as clearly structured as a model license. Practicing on this one is unusually good training.

What not to do

Don’t tell your kid a model is “open source” because the weights are downloadable. The OSI’s definition requires the training code and data information too, and most popular open-weight models don’t meet it. Precision here isn’t pedantry; it’s what lets a kid predict what they’re actually allowed to do.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can state the difference between open weights and open source, and knows to check the license before downloading.
• Month 2 red flags: They’ve published something built on a licensed model without the required notice. Or they use “open source” for anything free.
• Month 3 self-check: Hand them a model card and ask what they’d be allowed to do with it. A specific answer citing the license terms means the habit stuck.

Frequently Asked Questions

Is Gemma open source?

Not by the Open Source Initiative’s definition. The Gemma Terms of Use are Google’s own proprietary terms, and OSI’s Open Source AI Definition also requires the complete training code and detailed data information, not just the parameters. Note that the terms reference a separate Gemma 4 license under Apache 2.0, so check the specific version.

Can my kid use an open-weights model for a business?

Gemma’s terms permit commercial use. But redistribution carries conditions: you must include the use restrictions as an enforceable provision downstream, provide the Agreement, mark modified files, and include the specified Notice. Read the terms for the specific model and version before building on it.

What’s the difference between open weights and open source?

Open weights means the parameter file is downloadable. Open source, per OSI version 1.0, additionally requires the complete source code used to train and run the system and sufficiently detailed data information that a skilled person could build a substantially equivalent system.

Do these licenses apply to using ChatGPT or Claude?

No. Those are accessed through an API or an app under terms of service, not distributed as downloadable weights. Different legal structure entirely: you’re using a service rather than possessing a model.

Why do companies release open-weight models at all?

Several reasons: ecosystem building, running on devices where an API call isn’t possible, letting institutions self-host for privacy, and research adoption. Gemma’s reach from phones to edge infrastructure to space, per Google’s August 2026 roundup, is the payoff for that strategy.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Google AI. “Gemma Terms of Use.” https://ai.google.dev/gemma/terms
2. Open Source Initiative. “The Open Source AI Definition, version 1.0.” https://opensource.org/ai/open-source-ai-definition
3. Google. (2026, August). “Google AI updates, August 2026.” The Keyword. https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
4. Apache Software Foundation. “Apache License, Version 2.0.” https://www.apache.org/licenses/LICENSE-2.0
5. Anthropic. (2026, June 9). “Claude Fable 5 and Claude Mythos 5.” https://www.anthropic.com/news/claude-fable-5-mythos-5
6. Anthropic. (2026). “Pricing.” Claude Platform Documentation. https://platform.claude.com/docs/en/about-claude/pricing
7. Hinton, G., Vinyals, O., & Dean, J. (2015). “Distilling the Knowledge in a Neural Network.” arXiv:1503.02531. https://arxiv.org/abs/1503.02531