AI App Builder Training Data (2026): Who Is Holding the Opt-Out Switch

Quick answer

Ask eight AI app builders whether they train on your code and you will get five structurally different answers, only one of which is a plain no. Webflow says no in its contract and binds its subcontractors to it. Lovable, Replit and v0 all say yes by default and give you a switch, but the switch is gated behind a different plan on each one and lives in your account rather than in your project. Base44 says yes on every plan except Enterprise and states outright that there is no switch to find. Bolt inverts the whole question with Forge, where training is not a default you escape but a price you agree to pay in exchange for more usage. Bubble and Softr document their AI features in detail and say nothing about this in their product documentation at all.

The useful version of the question is not "do they train on my code". Mostly they do. It is four narrower questions that vendors answer independently, and that buyers collapse into one.

The four questions underneath the one everybody asks

Whose data. Your prompts and code are one thing. The personal data your app's end users type into your forms is another. These are separated on some platforms and not on others.

Which plan holds the control. On every platform here that offers a control at all, the control is plan-gated, and the gate sits at a different height on each one.

Where the switch physically lives. Not one of these switches is in your project. They are in account settings, in a personalization menu, or in a different product's team settings entirely.

When it takes effect. Every vendor that documents this says the same thing, and it is the part most likely to be missed: opting out is forward-only.

What the eight builders publish

Method note on the last two rows. "Not stated in product documentation" is a claim about a corpus, and the corpora are not equally strong. v0 publishes a full-content documentation index, so a measured absence there is a strong claim. The Lovable, Bolt, Replit, Base44, Bubble and Softr indexes are title-and-description lists, so a zero in the index is weaker and we fetched the underlying pages rather than resting on it. That distinction mattered here: Base44's index returns zero hits for the word training while the page it points to carries an entire section headed AI model training. An index tells you what exists. Only the page tells you what it says.

Webflow is the only plain no, and it is contractual rather than a setting

Webflow's AI terms carry a clause titled No Training. Verbatim:

Webflow, AI terms, 2026: "Webflow will use Inputs and Outputs solely to provide the AI Offerings to Customer. Webflow will not (and will not permit its subcontractors to) use them to train, improve, or fine-tune the AI Offerings or any underlying models without Customer's prior consent."

Two things make this the strongest position on the board. It is not plan-gated, so it does not ask you to buy a tier to get it. And it binds subcontractors, which is the exact gap the other vendors leave open when they describe sending your content to third-party model providers.

Lovable changed twenty days ago, and the change is opt-out rather than opt-in

Lovable is the only platform here with a documentation page dedicated to this question, and it is dated. Customer data from Free and Pro plans may be used to train Lovable's models as of 9 September 2026, unless you opt out. Business and Enterprise workspace data is excluded by default.

The boundary Lovable draws most clearly is the one most buyers actually care about:

Lovable, training data documentation, 2026: "Your apps' end-user data: information your apps' visitors or customers submit stays in your project's own database and storage, and is not used to train our models."

That exclusion applies on every plan, including Free. It is the cleanest single sentence any vendor here publishes on the whose-data question.

And then the trap. The opt-out is an account setting, and Lovable says plainly what that implies for a team:

Lovable, training data documentation, 2026: "The setting is on your account and covers the content you contribute. It does not change the setting of other members in a shared project or workspace. Each person controls their own."

So "have we opted out?" is not a question with one answer for a project. It has one answer per collaborator. A single teammate who never opened their account preferences keeps contributing that project's prompts and code to training, and nothing in the project itself will tell you.

Replit gates the control at Pro, which means free accounts do not have one

Replit's Model improvement page is short and unusually direct about scope: it uses "your prompts, instructions, context, and responses generated for you". The control is at Settings, then Personalization, then Model Improvement.

Replit, model improvement documentation, 2026: "This control is available to Pro users. For Enterprise users, Replit does not use customer content to train its models."

Read that as written. The opt-out is a paid-tier control. An account below Pro is described as having content used for model development with no documented switch. That is a different shape from Lovable, where the Free plan gets the same toggle as Pro.

v0 puts the switch in a different product

v0's documentation says its prompts and content may be used as inputs to models and learning systems from third-party providers. Enterprise customers are excluded, and so is anyone who has opted out. The detail worth noting is where:

v0, FAQs, 2026: "Content of v0 Enterprise customers, or any customer that has opted out of model training in their Vercel settings, is not used for training."

The switch is in Vercel settings, not v0. If you think of v0 as the product you use and Vercel as the thing you deploy to, you will look for this control in the wrong place. v0 does separately state that customer data from Vercel's platform services is not used to train the models behind v0, which is a genuinely useful carve-out and a different question from what your v0 chats do.

Base44 is the starkest, and it is the only one that says the end-user data part out loud

Base44 publishes a section headed AI model training with two bullets and no ambiguity: Enterprise workspaces are excluded automatically, and on all other plans your data can be used to train AI models. Then:

Base44, privacy and security documentation, 2026: "There is no opt-out setting to find or switch on. On Enterprise the exclusion is automatic and needs no action from you."

That is admirably clear and it is also the least generous position for anyone not on Enterprise, because there is no individual control to reach for.

The sharpest line on the whole board is the one describing what the Enterprise exclusion covers:

Base44, privacy and security documentation, 2026: "The Enterprise exclusion covers your whole workspace, including personal information that people submit through your app's forms."

Read the implication rather than the sentence. If the Enterprise exclusion is what covers end-user form submissions, then below Enterprise those submissions are not excluded. That is the opposite of Lovable's position, which excludes end-user data on every plan including Free. Two builders, same axis, and a buyer choosing between them on a feature matrix would never see it.

Bolt Forge inverts the question: training is the price, not the default

Everything above is a default you escape. Bolt Forge is the only thing on this board where training is something you actively agree to, with a consent gate, in exchange for something.

Bolt, Forge documentation, 2026: "Because using Forge shares your data to help train open-source models, you must agree to this before sending any prompts."

What you get is more usage through open-source models. What you give is enumerated more completely than anywhere else in this comparison:

Bolt, Forge documentation, 2026: "your prompts, code, project files, configuration, tool calls, edit histories, and any fix traces Bolt creates are de-identified, then stored and used by Bolt's partners to train open-source AI models."

Three details deserve attention. The recipient is Bolt's partners, not only Bolt, which is precisely the subcontractor gap Webflow's terms close. There is a geographic carve-out, with sessions from the EEA, UK and Switzerland excluded from training and datasets. And the exit is manual: switching away from Forge stops new collection, data already collected is not automatically removed, and removal is a request you email.

Set against the others, this is the most honest arrangement on the board even though it collects the most, because the trade is stated, priced and gated behind an explicit agreement. You know exactly what you bought.

Bubble and Softr document the features and not the policy

Both publish substantial AI documentation. Bubble covers its AI Agent, generating and editing apps with AI, and connecting your app to external models. Softr covers its AI Co-Builder, Database AI Agents, in-app AI chat and an MCP server. Neither product documentation set states whether your prompts, code or app data are used to train models.

That is not an accusation. It means the answer is not in the place a builder would look for it, and you would have to go to the privacy policy or the contract to find out. For a question with this much variation across the category, documenting the features and omitting the policy is a real gap.

The part everyone misses: opting out is forward-only

Every vendor here that documents an opt-out says the same thing, and none of them says it loudly.

Lovable, training data documentation, 2026: "Opting out takes effect going forward: your content is excluded from all training data assembled after your opt-out takes effect. It does not retract content from training datasets assembled, or models trained, before then."

Replit says new content will not be used. Bolt says data already collected is not automatically removed. The consequence is that this is a dated decision rather than a reversible setting. Everything you wrote before you flipped the switch is already in, and on Bolt the only route back out is an email to a privacy address.

If you are evaluating a builder you have already been using for months, the opt-out you flip today does not undo those months. That is worth knowing before you conclude that a switch has solved the problem.

What to actually do

Do not build this decision on which vendor is currently most permissive, because that is precisely the fact most likely to change, and Lovable's dated September 2026 change is the evidence. Build it on the questions, which are stable:

1. Find out whether the switch is per-person or per-workspace. If it is per-person, as it is on Lovable, then checking your own setting tells you almost nothing about a shared project. Check every collaborator or move to a tier where the exclusion is workspace-wide.
2. Ask specifically about end-user data, separately from your own code. These are separated on Lovable, explicitly in scope below Enterprise on Base44, and not addressed on several others. If your app collects anything sensitive from customers, this is the question that matters most, and it is not the one buyers usually ask.
3. Assume forward-only. Decide early, because the decision is dated and the months before it are not recoverable.
4. If you need a plain no rather than a setting, look for it in the contract. Webflow is the only vendor here that puts it there and binds subcontractors. A term you can point at is stronger than a toggle somebody has to remember to flip.

This is the same shape as the question of what you actually own when the builder writes your code. You can own the output completely and still have contributed the input to a training set, and those two facts sit together comfortably. It is also worth reading alongside what it costs to leave a platform, because the Forge arrangement is the clearest example in the category of a benefit you are given now in exchange for something you cannot take back later.

Honest limits

We read published vendor documentation and terms as of 29 September 2026. We did not audit anyone's training pipelines and could not: nothing here is a verification that a vendor does what it says. Where we report that something is not documented, that is a statement about the product documentation we fetched and not proof that no policy exists, and for Bubble, Softr and the standard Bolt models the answer may well sit in a privacy policy or a contract we did not treat as product documentation. Policies on this axis are changing quickly, as Lovable's 9 September 2026 change shows, so check the date on anything you read here before relying on it.