Guides · ← All guides
Who Owns What You Type Into an AI Tool? Data-Use Rights Explained
Last updated: October 13, 2026 · 10 min read
This guide is informational, not legal advice.
Picture this (a hypothetical, but a common one): a marketing agency pastes a client's unreleased product brief into a free AI chatbot to brainstorm taglines. Six months later, the model confidently describes a "new" product category that looks suspiciously like that brief. Nobody can prove where the model learned it. The agency can't un-paste the brief either.
This is the question every business using AI tools should be able to answer before the first prompt, not after: who owns what you typed, and what can the vendor do with it? The answer isn't on the marketing page. It's scattered across the terms of service, the privacy policy, and a settings screen you've never opened — and vendors' answers differ enormously.
Ownership is the wrong first question
Nearly every AI vendor's terms say you retain ownership of your inputs. That sentence feels reassuring. It shouldn't be, because it's not the operative one.
The operative question is what rights you grant the vendor in return. Every vendor needs some license to your inputs just to function — they have to receive your prompt, run it through the model, and send you the answer. A license limited to "operating and providing the service" is narrow and normal. It covers exactly what you'd expect.
The trouble starts when the license keeps going. Watch for grants that extend to "developing, improving, or training our products and services." That's a training-data clause wearing a license costume. The words "improve" and "develop" are doing heavy lifting in a lot of AI terms right now. When you see them, slow down and read the whole paragraph. Ask: improve how? With my data? For which products? Can I turn it off?
Also check the adjectives attached to the license: "worldwide, perpetual, irrevocable, sublicensable, royalty-free." Those are standard in social media terms. In AI vendor terms, a perpetual irrevocable sublicense to your inputs is a much bigger deal than it sounds, because the vendor's product is a model that may absorb patterns from your data indefinitely.
Training-data use: the clause that matters most
Nothing in an AI vendor's terms affects you more than whether your inputs train future models. And the industry's track record here is a case study in why you can't set this once and forget it.
In May 2024, Slack's privacy principles stated that its systems analyze customer data — messages, content, files — to develop AI and machine learning models. The catch that set off a week of outrage: opting out wasn't a toggle. A workspace owner had to email Slack's customer experience team with the subject line "Slack Global model opt-out request" to get the workspace excluded. Training was the default; privacy was the paperwork. Slack later clarified that customer data wasn't used to train generative AI products, only traditional ML features like emoji recommendations — but the structure of the deal (opt-out by email, train-by-default) is the pattern to memorize. You'll see it again.
A year earlier, Zoom quietly updated its terms in March 2023 with language granting itself broad rights over customer content, including use for training and tuning AI models. Nobody noticed until August 2023, when the changes surfaced publicly and triggered a backlash. Zoom then revised its terms twice in one week — first adding "without your consent" language, then removing the AI-training permissions from customer content terms altogether. The lesson isn't that Zoom was uniquely careless. It's that a company with one of the most-read terms documents in the world changed its AI data terms without meaningful notice, and the only reason anyone found out was public scrutiny.
So when you audit a vendor, pin down the answer in one of three buckets:
- Clear no: the terms state inputs are not used for training, with no weasel words. Business and API tiers from major vendors usually land here — but verify, because "usually" isn't a clause.
- Yes, with an opt-out: training is on unless you act. This is where you find out whether the opt-out is a settings toggle (good), an email to support (Slack-style friction), or something buried in an enterprise amendment.
- Ambiguous: "we may use data to improve our services." No mention of training, no mention of opt-out. Treat this as a yes until proven otherwise.
The free-tier trap
There's a pricing pattern across the AI industry that nobody advertises: the free tier and the paid tier are often different products wearing the same logo. Free tiers frequently reserve the right to train on inputs; paid and business tiers frequently don't. The price gap isn't just about features or rate limits. Part of what you're buying is a different data deal.
This matters because of how AI tools actually spread inside companies. Nobody procures the free tier. A designer signs up on a Friday afternoon to mock up a campaign. A developer pastes a stack trace at 11 p.m. to debug faster. Each of them just accepted terms that the legal team never reviewed — on the tier with the weakest data protections the vendor offers. If your team uses AI tools at all, assume free-tier accounts exist that you don't know about. We'll come back to this in our shadow-AI guide.
Retention and deletion: the second clause
Even vendors that don't train on your data typically keep it for a while. Common windows are 30 to 90 days, for abuse monitoring, debugging, and legal compliance. What you want in the terms is a number, a reason, and a deletion mechanism.
Read the deletion promise carefully. There's a meaningful gap between "we delete your data within 30 days of your request" and "we will stop displaying your data." The first is deletion. The second is hiding. Also watch for the carve-outs: many policies promise deletion "except where we are required to retain it for legal or security purposes" — which can swallow the promise if the exceptions are broad enough.
One more thing to check: the gap between account deletion and data deletion. "Delete your account" and "delete your data" are not the same promise, and some vendors' terms make them sound like one.
Human review: strangers reading your prompts
Many vendors reserve the right to have humans review inputs — usually content flagged by automated abuse systems, sometimes random samples for quality evaluation. The terms phrase this as "we may review content to ensure compliance with our policies" or similar.
Translate that into operational terms: if your team pastes customer records, patient notes, unfiled patents, or source code into a chat interface, a contractor at the vendor may read it. For low-sensitivity work, that's an acceptable trade. For regulated or confidential workflows, it's a reason to prefer vendors and tiers that minimize human review — and to treat the AI chatbox with the same caution you'd apply to emailing the vendor directly. If you wouldn't email it to them, don't paste it.
Who owns the outputs?
The model generates a report, a design, a block of code. Who owns that? Terms vary more here than on inputs, and the answer has three common shapes:
- Assigned to you: the vendor assigns all right, title, and interest in outputs to you. Cleanest for businesses.
- Licensed to you: you get a license to use the outputs, but the vendor retains underlying rights. Usually fine in practice, worth knowing.
- Yours, with disclaimers: you own the outputs, but the vendor disclaims any warranty that they don't infringe someone else's rights — and notes that identical or similar outputs may be generated for other users.
That last disclaimer is doing real work. Models don't have memories of individual prompts, but they do produce similar outputs for similar inputs. If two companies ask for a logo concept in the same style, they can get uncomfortably similar results. The terms are telling you, in advance, that this is your problem to manage.
FAQ
Do I own my prompts?
Almost always, yes — on paper. What matters is the license you grant back to the vendor. Ownership without a narrow license grant is a hollow right.
If I delete my account, is my data gone?
Not necessarily. Check whether the terms promise data deletion on a timeline, or only account closure. And check the retention exceptions — legal and security carve-outs can extend storage well beyond the headline number.
Can a vendor change its training policy after I sign up?
Yes, and this has happened repeatedly. Most terms allow the vendor to update policies with notice (sometimes just a "last updated" date change), and continued use counts as acceptance. This is why one-time reviews aren't enough — the terms need watching, not just reading.
Is the paid tier always safer than free?
Not always, but often enough that it's the first thing to check. The pattern — free trains, paid doesn't — is common but not universal. Some vendors train on paid-tier data too, and some free tiers are clean. Read the clause for the tier you're actually on.
What to do this week
- List the AI tools your team actually uses. Include the free ones people signed up for themselves — ask around, because the official list is never complete.
- For each tool, find the training-data clause and the retention clause. If you can't locate both in five minutes, treat that as a red flag about the vendor's transparency.
- Switch sensitive work to paid or business tiers where the data terms are stronger, and turn on every available training opt-out today.
- Save a dated copy of each vendor's current terms. Without a baseline, you'll never be able to prove what changed or when.
- Set up change tracking on each vendor's terms pages, so the next quiet policy flip lands in your inbox instead of hiding in a "last updated" date.