Licensing and data terms

What you will be able to do

Find out what you may do with a model, and what its provider may do with your customers’ tickets, before you build on it. A model that works well on your tickets is not yet a model you may use.

Before you startWhat to know and to have readyAbout 35 minutes

What you need to know

• The three ways to get a model (hosted API, managed host for open weights, open weights on your own machine), from Hosted and self-hosted choices.

What you need

• A browser. You read two official pages in the guided practice.

Terms change often: read the current version of each page before you decide.

This is not legal advice. The lesson teaches you which questions to ask and where to find the answers. A qualified person must confirm the answers for a real project.

Two questions before you sign a lease

Before you rent a flat, you read the lease for two kinds of answers. First: what may I do here? May I sublet a room, or run a small business from it? Second: what may the landlord do? May the landlord come in when you are out, or keep a copy of your key?

A language model comes with the same two kinds of question. What may you do with the model? And what may the company that runs it do with the text you send?

The problem: “it runs, so we can use it”

Tomás has Gemma 3 4B running on the Mac mini. “It was free to download, and it answers our tickets. We are done.”

Omar is not sure: “Free for a person who tries it at home, or free for a shop that sells garden tools?” The download did not ask Larkfield to sign anything. But a licence comes with the weights, and Ollama shows it with ollama show gemma3:4b --license. Its first lines:

Gemma Terms of Use

Last modified: February 21, 2024

By using, reproducing, modifying, distributing, performing or displaying any portion or element of Gemma, ...

So using Gemma means agreeing to a document that nobody at Larkfield has read. And this copy is not even the current one: Google’s page says “Last modified: April 1, 2026”. The failure case at the end shows what changed.

The hosted side has the same problem in another form. Suppose Amira tests a reply with the free quota of the Gemini API, using a real ticket with a customer’s name and order number. The Gemini API Additional Terms of Service (page last updated 2026-04-28) say this about the free quota:

When you use Unpaid Services, including, for example, Google AI Studio and the unpaid quota on Gemini API, Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies […]. To help with quality and improve our products, human reviewers may read, annotate, and process your API input and output.

And a few lines later: “Do not submit sensitive, confidential, or personal information to the Unpaid Services.” The model is the same model that Larkfield would pay for. The rules for the ticket are not.

The idea: three documents, three questions

A model choice needs three separate answers, from three separate places:

1. Capability: what the model can do on your task. No document answers this; your own test does (Module 4).
2. The model licence: what you may do with the weights. A licence is the document in which the owner of the weights gives you permission to use, change and share them, with conditions.
3. The service terms: what the company that runs the model does with your data. Service terms (or data terms) are the contract and policies of a hosted API or a managed host. They say how long it keeps your prompts, whether it trains on them, and where it processes them.

The picture puts the three documents against the three ways to get a model from the last lesson.

Three documents, three questions. Your test answers the first row in every column. Which documents you must read depends on how you get the model.

Text description of the diagram

A grid of three rows and three columns.

• The rows are the three questions: model capability (“What can it do on our tickets?”), the model licence (“What may we do with the weights?”) and the service terms (“What happens to our data?”).
• The columns are the three ways to get a model: hosted API, managed host for open weights, and open weights on your own machine.
• Capability: “your test” in all three columns, in dark blue (you find out yourself).
• Model licence: for the hosted API, “No weights to licence. A usage policy applies”, with a dotted border (does not apply). For the managed host and your own machine: “the weights’ licence”, a document you read.
• Service terms: “the provider’s terms” for the hosted API and “the host’s terms” for the managed host, both documents you read. For your own machine: “No provider: your own rules”, with a dotted border.

Keep the rows apart. A good score on your tickets says nothing about retention. A generous licence says nothing about quality. And the same open model can sit in two columns. Gemma on the Mac mini has a licence and no service terms; Gemma on a managed host has both.

The six questions to check

Grace writes down the questions that Larkfield must answer for every candidate. Each one comes from a real risk:

• Commercial use. May a business use the model to serve its customers? Some open-model licences allow only research or non-commercial use.
• Training on your data. May the provider use your prompts and outputs to improve its models? Then a customer’s ticket can influence a future model.
• Retention. How long does the provider keep your prompts and outputs, and who can read them?
• Region. In which countries is the data processed and stored? Larkfield’s customers are in Canada, and Quebec has its own privacy law for personal information.
• Redistribution of weights. May you copy the weights to another company, or offer the model to others as a service? On what conditions?
• Acceptable use. Which uses does the licence or the provider forbid? An acceptable-use policy (also called a usage policy or prohibited-use policy) lists them. A breach can end your right to use the model.

Most questions belong to one document. Training, retention and region are in the service terms. Commercial use and redistribution are in the licence. Acceptable use can be in both.

Worked example 1: an open-model licence

Amira reads the official Gemma Terms of Use (“Last modified: April 1, 2026”) with Grace’s six questions. Each answer quotes the text.

Commercial use. No clause limits Gemma to research or to non-commercial use. The limits are in section 3.2: “You must not use any of the Gemma Services: for the restricted uses set forth in the Gemma Prohibited Use Policy […]; or in violation of applicable laws and regulations.” Larkfield’s support replies are not on that policy’s list.

Training, retention, region. These do not apply while Gemma runs on the Mac mini. No company receives the tickets. Larkfield’s own rules decide how long it keeps them.

Redistribution. Section 3.1 allows it on conditions. Two of them: “You must provide all third party recipients of Gemma or Model Derivatives a copy of this Agreement”, and “You must cause any modified files to carry prominent notices stating that you modified the files.” Larkfield does not plan to share the weights, so these conditions do not apply today.

Acceptable use, and one surprise. Section 3.2 continues: “Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement.” Open weights on your own machine are not free of the owner’s rules.

Outputs. Section 3.3: “Google claims no rights in Outputs you generate using Gemma. You and your users are solely responsible for Outputs and their subsequent uses.” So the replies belong to Larkfield, and so does the responsibility for a wrong reply.

Compare a standard open-source licence. Mistral 7B and Qwen3 4B come with the Apache License 2.0 (ollama show mistral:7b --license). It gives “a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work”. It has no list of forbidden uses. Its main condition for redistribution: “You must give any other recipients of the Work or Derivative Works a copy of this License”. “Open weights” is not one licence: read the one that comes with each model.

Worked example 2: a provider’s data terms

Now the hosted side. The answer to “do they train on our tickets?” can depend on how you pay, not on which model you call. The Gemini API terms have two sets of rules. The flow shows which set applies to a ticket.

Which data rules apply to a ticket sent to the Gemini API, from the Gemini API Additional Terms (last updated 2026-04-28). Larkfield is in Canada, so the free quota means the unpaid rules.

Text description of the diagram

A flow chart with two questions.

1. A ticket is sent to the Gemini API. First question: is it a paid service, that is, a Cloud project with an active billing account?
2. If yes: the paid rules apply. Google does not use the prompts and responses to improve its products, and it logs them for a limited time.
3. If no, second question: are the users in the European Economic Area, Switzerland or the United Kingdom?
4. If yes: the paid rules apply, even on the free quota.
5. If no: the unpaid rules apply. Google uses the content to improve its products, and human reviewers may read it.

The paid rules, in the terms’ own words: “When you use Paid Services, including, for example, the paid quota of the Gemini API, Google doesn’t use your prompts […] or responses to improve our products”. And: “For Paid Services, Google logs prompts and responses for a limited period of time, solely for detecting and preventing violations of the Prohibited Use Policy […]”. On region: “This data may be stored transiently or cached in any country in which Google or its agents maintain facilities.”

A second provider gives the retention as a number. Anthropic’s privacy page How long do you store my organization’s data? (dated July 1, 2026) says: “For Anthropic API users, we automatically delete inputs and outputs on our backend within 30 days of receipt or generation”. The exceptions on the same page matter. They include a service that you keep data in (such as the Files API), a separate “zero data retention agreement”, and the law. For a flagged request: “We retain inputs and outputs for up to 2 years”.

Region can also change the price. Anthropic’s pricing page says that US-only processing with the inference_geo setting “incurs a 1.1x multiplier on all token pricing categories”. A data rule can be a cost, too.

Grace’s terms card

Grace puts the answers for three candidates on one card. Each cell is what the official pages said.

Question	Gemma 3 4B on the Mac mini	Claude API	Gemini API, paid
Commercial use	Allowed; Prohibited Use Policy applies	Allowed; Usage Policy applies	Allowed; Prohibited Use Policy applies
Training on our tickets	No provider receives them	Not on the page we read: check the commercial terms	“doesn’t use your prompts […] to improve our products”
Retention	Our own rule	Deleted “within 30 days”, with exceptions	“a limited period of time”
Region	Larkfield’s own machine	Default region not on the pages we read: check. US-only costs 1.1× the price	“any country in which Google or its agents maintain facilities”
Redistribution of weights	Allowed with conditions (3.1)	No weights	No weights
Acceptable use	3.2, and Google may “restrict (remotely or otherwise)”	Usage Policy	Prohibited Use Policy

Look at the open answers in the Claude column. The pages we read do not say whether Anthropic trains on API data, or where it processes data by default, so the card says “check”. A card with an honest gap is better than a guess. The gaps go on the list for Larkfield’s lawyer, with the questions that only a lawyer can answer, such as Quebec’s privacy law.

When to check, and what it costs

Check the terms before you send the first real ticket, not after the comparison. A model that fails a must-have (for example “no training on customer tickets”) leaves the race before you spend time on its scores. Module 5 calls these must-haves.

Check them again when anything changes: a new model version, a new price plan, a move to a paid plan, or a new date on the page. Record the date of every page you read, and keep a copy.

The cost is a few hours of careful reading per candidate, and some legal review for the final choice. The cost of skipping it is higher: customer data in a training set, or a product built on weights that you may not use.

Guided practice: read one licence and one data page

Read the official pages, not a summary of them.

1. Open the Gemma Terms of Use. Write down the “Last modified” date.
2. Find section 3.1. Write down one condition for sharing the weights with another company.
3. Find section 3.3. Write down who owns the replies that Gemma writes.
4. Open the Gemini API Additional Terms. Write down the “Last updated” date at the bottom of the page.
5. Find “Unpaid Services”. Write down whether people may read the prompts.
6. Find “Paid Services”. Write down how the paid rules are different.

If a date is newer than the dates in this lesson, the terms changed. Then trust the page, not the lesson.

Show the answers

1. “Last modified: April 1, 2026”.
2. Any one of: give every recipient a copy of the agreement; mark the modified files; put the use restrictions of section 3.2 in your own agreement; add a “Notice” file.
3. Larkfield does: “Google claims no rights in Outputs you generate using Gemma.” Larkfield is also responsible for them.
4. “Last updated 2026-04-28 UTC.”
5. Yes: “human reviewers may read, annotate, and process your API input and output.”
6. Google does not use the prompts and responses to improve its products. It logs them for a limited time to find policy violations.

Your turn: an Apache 2.0 model

Larkfield’s warehouse team wants to run Mistral 7B on its own server, and later give a changed copy to a partner company. Use the Apache License 2.0 quotes in this lesson. Answer three of the six questions: commercial use, redistribution and acceptable use. Then say which document you must read if the team uses Mistral’s own hosted API instead.

Show a model answer

Commercial use: allowed. The licence gives a “no-charge, royalty-free” right to use and distribute the work, with no limit to research. Redistribution: allowed. The team must give the partner a copy of the licence and mark the files it changed. Acceptable use: the Apache License 2.0 has no list of forbidden uses, but the law still applies. With Mistral’s hosted API, the team does not receive the weights. It must read Mistral’s service terms instead: training on the data, retention, region and the usage policy.

Common mistake

The licence on the disk is older than the licence on the web

Symptom: Amira answers the six questions for Gemma from ollama show gemma3:4b --license. Her notes say: “Section 4.1: we must make reasonable efforts to use the latest version of Gemma. So we must upgrade the model each time Google releases one.”

Cause: the copy on the disk is as old as the package. It says “Last modified: February 21, 2024”. In that version, section 4.1 says: “Google may update Gemma from time to time, and you must make reasonable efforts to use the latest version of Gemma”. The official page, “Last modified: April 1, 2026”, says only: “Google may update Gemma from time to time.” The duty to upgrade is gone. Her notes describe a contract that no longer applies.

Fix: read the official page. Write its date and URL in your notes, and keep a copy (a saved PDF) with the decision. When the date on the page changes, read it again.

Consequence in AI work: the terms of models and APIs change more often than most contracts. A decision record without the dates of the terms cannot show which rules the team agreed to.

Knowledge check

Check your understanding

Omar says: 'Claude Sonnet 5.5 wrote good replies to our 24 test tickets, so it is safe to send it customer tickets.' Which of the three documents does his evidence come from, and which question is still open?

Show the answer

His evidence is about capability: what the model can do on Larkfield’s tickets, from Larkfield’s own test. Whether it is safe to send tickets is a question for the service terms: whether the provider trains on the data, how long it keeps it, and where it processes it. A model can score well and still have data terms that Larkfield cannot accept. The two questions need two kinds of evidence.

Check your understanding

Predict: Larkfield moves its Gemini API tests from the free quota to a Cloud project with active billing. Which of the six answers change, and which stay the same?

Show the answer

Training on our tickets changes: on the paid service, Google does not use the prompts and responses to improve its products. Retention changes: the paid service logs them only for a limited time, to find policy violations, and no human reviewers read them to improve products. Region stays the same: the data may be cached in any country where Google has facilities. Acceptable use stays the same: the Prohibited Use Policy applies to both. Commercial use stays the same. Redistribution of weights does not apply: there are no weights. The model is the same; only the data rules change.

Recap

• Three questions need three kinds of evidence: capability (your own test), the model licence (what you may do with the weights), and the service terms (what the provider does with your data).
• Check six things for each candidate: commercial use, training on your data, retention, region, redistribution of weights, and acceptable use.
• “Open weights” is not one licence. Gemma’s terms forbid the uses in a policy and let Google restrict use; the Apache License 2.0 has no such list.
• The same hosted model can have different data rules: the Gemini API free quota may use your prompts to improve products; the paid service does not.
• Record the date and URL of every page you read, and read the official page, not a copy on your disk. This is not legal advice: a qualified person confirms the final answers.

Next, the costs that you can measure: speed, memory and the size of the model.

Next: Resource trade-offs