Keep client work out of AI training: what to ask vendors before you upload

Clients now ask this on the first call, and many put it into the contract. It is a fair question. If you upload an unreleased product photo to an AI tool, you want to know where that file goes and whether it helps the vendor build its next model.

The answer is different for every tool and every plan. This page gives a method you can reuse: read your contract, ask the vendor in writing, sort your client material into three levels, and write a one page policy. I use Higgsfield as a worked example because its terms and team help page are public, and I show what other vendors say too. This is general information, not legal advice. Check local law and ask a lawyer about any contract that matters.

This page may contain affiliate links from which Blur Studio earns commissions.

The question a client asks, and the answer that loses the account

The answer that loses the account is a confident one you cannot back up: "No, never, it is a business plan." Business plan names tell you almost nothing. Some vendors promise no training on every paid tier. Others promise it only on a top tier sold by quote. A few allow training unless you opt out.

The safe answer has three parts: what the contract with the client says, what the vendor's written terms say for your plan, and what you do to limit exposure. The rest of this page builds those three parts.

What "trains on my data" actually means

People use one phrase for three different things. Ask about each.

• Training. Does the vendor use your uploads and outputs to build or improve its AI models? Training means the model learns from your content. Once a model has learned, you cannot take that back.
• Retention. How long does the vendor keep your files, prompts and outputs, and what happens to backups after you delete?
• Human review. Can staff or contractors of the vendor read your content, for safety checks or support?

A confidentiality duty (the kind in an NDA, a non-disclosure agreement) attaches to the information, not the software. If the client's information is under NDA, it stays under NDA when you paste it into a tool. That is why the vendor's terms matter even when the output looks harmless.

Read your own client contract first

This is the cheapest step, and it is the one most vendor pages skip. Open the master services agreement (MSA) or the NDA for the client and look for:

1. A clause about subcontractors or sub-processors. An AI tool may count as one, which can mean you need the client's consent.
2. A clause about using third party tools with client data.
3. A clause about confidential information and how it must be stored.
4. Any ban or approval step for AI tools by name.
5. Who is responsible if a vendor has a breach.

If the contract is silent, that is not permission. It means the question has not been asked yet, so ask the client before the first upload, and keep the answer in writing.

Consumer, business and enterprise: how tiers differ

As a rule of thumb, free and personal plans have the loosest terms, business plans sit in the middle, and enterprise contracts usually carry the strongest promises. That is a pattern, not a law. Read the actual page for the plan you pay for.

OpenAI is a clear public example. Its enterprise privacy page (updated 8 January 2026) says data from ChatGPT Business, Enterprise, Edu and its API is not used for training by default, while data from the consumer version of ChatGPT may be. It also says business workspace admins can view, export and delete members' conversations. So a business tier can be more private from the vendor and still open to your own admins.

The same point applies to staff accounts. A designer's personal login is not covered by any enterprise promise you signed. Your policy has to say work goes in company accounts only. That ties to the page on shared AI accounts and bills.

What Higgsfield's terms say

From Higgsfield's terms of use (last updated 23 October 2026) and its team help article (dated 2 August 2026), in my words and not as a legal reading:

• Section 4.4 lets Higgsfield use your content, inputs and outputs to train, develop and improve its AI models.
• The exception is an Enterprise agreement. Under one, Higgsfield says it does not use the customer's content to train or improve its models and treats it as confidential. The help center says this has been written into section 4.4 since July 2026.
• The help center plan table lists no data training as No for Team, No for Scale and Yes for Enterprise. Many buyers assume "business plan" means private. On Higgsfield, paying for Team does not change training use.
• Deleting content or the account stops that use going forward. Models already trained cannot be changed.
• Workspace administrators can access and manage members' content, so workspace content is not private from admins on any plan.
• Higgsfield says in section 4.4 that it will not use your private content, or private content you create for or on behalf of a client, in its own marketing without your consent. Content you post publicly on the service is treated differently, so keep client work private.
• Section 4.2 says you must hold the rights, releases and consents for any name, likeness or voice in your content. Section 5.2 bars using outputs to train other AI models, with a carve-out for models trained with your own inputs.

Enterprise is sold by quote and the public pages do not give a price. So for confidential client work on Higgsfield, there are three honest paths: ask for an Enterprise agreement, keep client material out and use text prompts, or use another tool whose terms fit. For wider safety questions read the safety and legitimacy check.

One trap deserves its own line. A reference image is an input. If you upload a client's unreleased product photo or logo as a reference, it falls under those input terms. Until the contract question is settled, rewrite the idea as a text prompt or use a stand-in image you are free to share.

What other tools say, and what is not confirmed

• Runway, an AI video tool for creative teams, lists "no training on your data", custom single sign-on, SOC 2 and ISO 27001 certifications and legal indemnification for its Enterprise tier on its Team plan page. Those items are not on the Team plan list. So Runway is clearer in public than Higgsfield about what the top tier includes, but the same pattern holds: the promise sits in the tier sold by quote.
• OpenAI publishes a default no-training commitment for its business tiers and its API, without an enterprise contract. For a team that wants a public commitment without a sales call, that is a real advantage, though it covers OpenAI's own products and API, not other vendors' tools.
• Adobe Firefly, Adobe's image and video generator, markets itself as commercially safe and has a teams plan. Its FAQ (last updated 13 May 2026) lists the question whether your content is automatically used to train Firefly, but the answer sits in a collapsed section that did not show in the page text I read. So I am not stating an answer here. Read that FAQ and ask Adobe in writing.

I have left out claims from third party trackers about other vendors' consumer terms, because only the vendor's own terms count. Always go to the vendor's page.

Eight questions to send a vendor in writing

Copy these into an email to the vendor's sales or support team. Keep the reply, because a chat answer is not a contract.

1. On my plan, is customer content used for training or fine tuning, yes or no?
2. Where is that written, with the clause number or page link?
3. How long are inputs and outputs kept, and can I set that?
4. Do humans review content, and which contractors or teams?
5. Which third party model providers receive my content?
6. What happens on deletion, including backups, and how long does it take?
7. Do you sign a data processing agreement (DPA, a contract about how personal data is handled)?
8. If there is a breach, who is told and how fast?

Question 5 matters for platforms that bundle many models. A platform may pass your content to the company that makes the model you picked, and that company has its own terms. Ask for a list.

Beyond training: retention, logs, certificates and the exit test

Independent audits of AI tool privacy, such as one by the privacy company Sonomos (a vendor with its own product, so read it as one view), make the same point I made above: "no training" says nothing about how long files are kept, who can read them, or which other companies receive them. A tool can pass on training and still keep prompts for a set window for abuse checks. Four extra checks for larger clients:

• Retention with training off. Ask whether any copy is kept for abuse monitoring, and for how many days.
• Audit logs. If a file goes somewhere it should not, can an admin see who uploaded what and when? Personal and free plans often have no logs at all. Ask which plan adds them.
• Certificates. SOC 2 and ISO 27001 mean an outside auditor checked the vendor's security practices. Look on the vendor's trust or security page. A missing certificate is not proof of a problem, but it means you rely on the vendor's word.
• The exit test. Before you roll a tool out, make a test account, upload a harmless file, then ask for deletion and note how long the vendor takes to confirm. If nobody can tell you what happens to your data after you cancel, treat that as a warning.

Terms change. Put a reminder in the calendar every 90 days to re-open the terms page for each tool on your list and note the date. Also remember that a tool you run on your own computer or your own servers has no vendor training question at all, though it costs more effort to set up. That route suits red-level work when no hosted tool can give you a written promise.

Two shortcuts that do not hold up

• A settings toggle. Some tools let a user switch training off in a menu. That is weaker than a clause in the contract, because staff can forget it, a new login starts with the default, and the vendor can change the menu. Ask whether the protection is in writing and on by default.
• Removing names. Blanking a client name does not make a product shot, a script or a face anonymous. A launch date, a prototype shape or a price can still identify the client. Treat redaction as an extra step, not a substitute for the vendor's written answer.

A three level client data policy

Level	Examples	Rule
Green	Public brand assets, already published ads, logos on the client's website	May go in any tool you are allowed to use for work, on a company account.
Amber	Unreleased campaign ideas, draft scripts, internal photos	Only in tools or tiers with a written no-training promise. Otherwise rewrite as text prompts.
Red	NDA products, people's faces or voices, financials, health or legal material	Never uploaded as a reference unless the client signs off in writing for that tool.

The policy only works if people can classify quickly. Add a line to your job brief: "Data level: green, amber or red." The producer reads it before any upload.

A 7 day check of your current tools

A plan you can run this week:

1. Day 1. List every AI tool anyone has used on client work, including personal accounts. Ask staff directly.
2. Days 2 and 3. For each, note the plan and open the terms page. Record the training clause and its link.
3. Day 4. Send the eight questions to the three tools you use most.
4. Day 5. Sort the last month's client files into green, amber and red.
5. Days 6 and 7. Write the policy and send it to staff.

Pass if every tool has a clause link on file, every amber and red file has a known tool and plan, and nobody is using a personal login for client work. Fail if any tool has no written answer, or if any red file went into a tool you cannot show is covered. A fail means stop uploading to that tool until you have it in writing, and tell the client if the contract requires it.

Who should skip generative tools on client files

• Work under a strict NDA where the client has not agreed to any third party tool.
• Regulated material, such as health, legal or financial records.
• Pre-release products the client has not shown anyone.
• Any case where the only tool you like offers a no-training promise only through a quote you cannot get before the deadline. Use stock material, stand-ins or a traditional shoot for that job.

Skipping the tool for one job costs little. A breach claim costs far more.

How I'd approach this

I would do the contract read this week, before any vendor email. Then I would pick one or two tools and get their answers in writing, and I would run everything amber and red by text prompts until those answers arrive. I would not promise a client "never trained" for any plan unless I can point to a clause that says so for that plan.

For image and video work on Higgsfield, I would use it for green material and for amber material rewritten as prompts, and I would ask about an Enterprise agreement only if a client's work needs it. For the credit side of the same decision see the pricing explainer. If clients also ask whether AI use must be disclosed, that is a separate question, covered in the rights checklist for paid ads.

If you want this policy to live somewhere your team will follow it, Blur Studio builds the repeatable workflow, client approval portal or shared asset library around these tools, with a data level on every job. You can read more at our business page.