AI Terms of Service: What to Check Before You Sign

Meta released an 816-billion-parameter model this week under what it calls a licence with limited commercial terms, which allows many developers to use it for free. That phrasing is doing real work, and it is a good reminder that the word accompanying an AI tool, whether open or free or unlimited, is a marketing word until you read the clause behind it.

The six that matter

Clause	The question it answers	Answer you want
Training rights	Can our inputs train their model?	No, by default, on this plan
Output ownership	Who holds rights in what it produces?	Assigned to the customer, in writing
Commercial scope	Is our intended use permitted?	No threshold we expect to cross
Retention and location	How long, and stored where?	A stated period and a named region
Indemnity	Who pays if output infringes?	Provider covers it, with conditions you can meet
Exit	What do we keep if we leave?	Export in a usable format, stated window

Reading those six across your main tools is an afternoon, not a legal project. What makes it worth doing is that the answers differ sharply between providers and between plans at the same provider, so the generalisation you are carrying in your head is probably wrong about at least one of them.

Training rights, and the free-plan gap

The common pattern among major providers is that business and enterprise plans exclude customer inputs from model training by default, while consumer and free tiers do not. The gap between those two sets of terms is where most real exposure sits.

The exposure is rarely the plan you bought. It is the staff member who uses a personal free account because it is faster than requesting a licence, pasting in a client contract, and operating under terms nobody in the business has read. Your policy covers the tool you approved; their account is a different agreement.

Which makes provisioning a compliance control rather than an IT convenience. If using the approved account is harder than using a personal one, the terms you negotiated describe a tool your team has routed around, a problem covered in whether your business needs an AI policy.

Two different ownership questions

Most commercial providers now assign output rights to the customer. That settles the contractual question between you and the vendor, and it is the question the terms are capable of answering.

Whether the output is protectable by copyright is separate, depends on how much human authorship went into it, and varies by jurisdiction. A vendor saying the output is yours does not make it defensible against someone else copying it.

For material you intend to license, enforce or sell on, treat those as two questions and get advice on the second. The wider legal picture is in AI copyright lawsuits and what they mean for you.

Open weights is not open source

Released weights frequently come with a custom licence rather than a recognised open source one, and the conditions attached vary a great deal. Common restrictions include revenue or active-user thresholds above which different terms apply, limits on building competing products, and attribution requirements.

The threshold-based ones are worth particular attention because they are invisible while a project is small. A clause that applies above a certain scale is irrelevant during the pilot and decisive if the thing succeeds, which is the worst time to discover it.

Check the licence against the use you actually plan, including the version of that use two years out. This holds equally for open source libraries bundled into an AI tool, where the obligations travel with the code regardless of what the vendor's own terms say.

Retention, location and deletion

Three sub-questions hide inside this one. How long inputs are kept, which is often thirty days for abuse monitoring even where training is excluded. Which region they are processed and stored in. And whether deletion on request is actually offered, or only account closure.

For businesses handling personal information or operating under sector rules, the region answer is frequently the one that decides whether a tool is usable at all, which is the subject of AI data residency in Canada.

Indemnity has conditions

Several providers now offer indemnity covering copyright claims arising from model output, which is a meaningful protection and comes with conditions. Typical ones include using the provider's safety features, not disabling filters, and not deliberately prompting for infringing material.

Read the conditions rather than the headline, since an indemnity that assumes a configuration you have turned off is not protection you hold. Where output goes to customers, that configuration becomes part of your process rather than an IT setting.

What you keep when you leave

Exit terms are the most skipped and the most expensive to get wrong. Ask three things: can you export your data in a format something else can read, how long after cancellation can you still do it, and what happens to anything you fine-tuned or configured.

Fine-tuned models and custom configurations are frequently non-portable by design, so the work is a sunk cost rather than an asset. That changes the economics of committing to one provider, and it belongs in the same calculation as model hedging to avoid vendor lock-in.

Keep a dated copy

AI terms change more often than conventional software terms, usually announced by an email few people open. Save a dated copy of the version you reviewed for each tool you depend on, and re-read the training and retention clauses annually or whenever a provider ships a major product change.

A quarterly calendar reminder costs nothing and is among the cheapest controls available here. Where a vendor relationship is material enough to need more than that, the questions to put to them directly are in how to vet AI vendor claims and usage limits, and the governance wrapper is in an AI governance framework for a small business. This is general information rather than legal advice.

Frequently Asked Questions

What should I look for in an AI tool’s terms of service?

Six clauses carry most of the risk. Whether your inputs and outputs can be used to train the provider’s models. Who owns the output. Whether the licence permits your intended commercial use. How long data is retained and where. What indemnity, if any, covers you if the output infringes someone else’s rights. And what happens to your data and your fine-tuned work when you stop paying. Everything else in a typical agreement is closer to standard software boilerplate.

Does the free plan have different terms?

Very often, and the difference usually concerns training. A number of providers train on free-tier inputs and exclude paid business plans from training by default. That distinction matters more than the monthly fee, because a staff member using a personal free account for work is operating under different terms than the business account you reviewed. This is the most common gap between what a policy says and what is happening.

Who owns the output an AI tool produces?

Most commercial providers assign output rights to the customer in their terms, which settles the contractual question between you and the vendor. It does not settle whether the output is protectable by copyright, which depends on how much human authorship went into it and varies by jurisdiction. For material you intend to license, enforce or sell on, those are two separate questions and the terms only answer the first.

What does a limited commercial licence mean?

It means the right to use the tool or model commercially comes with conditions, which commonly include revenue or user thresholds above which different terms apply, restrictions on building competing products, or attribution requirements. Open weights in particular do not imply an open source licence. Read the specific licence rather than assuming, because the restriction that matters is usually the one that only becomes relevant once the project succeeds.

How often do AI terms change?

More often than conventional software terms, and typically with notice by email that few people read. Keep a dated copy of the terms you reviewed for each tool you depend on, and re-check the training and retention clauses annually or when a provider announces a major product change. A quarterly reminder costs nothing and has caught real changes for businesses we work with.