The model writes the code, but the law requires a human: who owns a startup built with Claude?

Anthropic, OpenAI and Cursor assign their rights in the output "if any", meaning only if such rights exist at all. A lot depends on the subscription. On personal plans (Claude Pro and Max, ChatGPT Plus, GitHub Copilot Pro), code and conversations go into training until the user switches this off. Business plans and the API exclude training, and the provider promises to defend the customer against claims from rights holders, though not for output that has been modified or combined with other code. App builders such as Lovable, Bolt and v0 give no such protection.

The most common trap is a personal subscription held by a founder or a freelancer. Any rights in the output, if they exist, belong to that individual rather than the company, and they can only reach the company through a separate assignment.

Claude Code, Cursor and Codex were trained on open-source code and sometimes reproduce chunks of it almost word for word, with no author or licence attached. In 2026 Black Duck found licence conflicts in 68% of the codebases it audited, up from 56% a year earlier.

For a startup this means two risks:

• Using someone else's code without meeting the terms of its licence is an infringement, and the rights holder can demand that you stop using it or pay damages.
• Copyleft licences (GPL, AGPL) can "infect" the product. These open-source licences require derivative works to be distributed under the same open terms. If such a fragment ends up in a product that is distributed, the company may have to release the source code of the whole product, and AGPL bites even when the product is simply available over a network, like any ordinary SaaS.

Dependency scanners, which check third-party libraries and their licences, miss this kind of code because the fragment does not point to any package. In September 2026 a US appeals court threw out the developers' main claim against GitHub over Copilot, but their claims for breach of open-source licences are still going.

Since copyright offers little to hold on to, the code can be protected as a trade secret, that is, as valuable information the company keeps confidential. For AI-generated code this is the most workable option, though it has its own pitfalls.

The first is secrecy. If the code was written through a personal account with training switched on, a court may decide the company took no reasonable steps to protect it. In January 2026 a federal court in California dismissed a trade secret claim against OpenAI after finding that, by developing her ideas in ChatGPT, the plaintiff had voluntarily disclosed them to the company.

The second is access. The repository, the API keys, the project in the app builder and the AI service account are often registered to the founder's personal email, and investors will check in due diligence who controls them.

In practice, then, trade secrecy is what protects the product today, and that comes down to who holds the keys. An assignment of "all rights" does not help much here either: you can only transfer rights that actually came into existence.

If the product was built with Claude Code, Codex or Cursor, the copyright position is stronger. The review will look at whose accounts and plans the code was written on, whether founders and contractors have signed IP assignments, and whether the codebase contains third-party fragments under copyleft licences.

If it was put together in an app builder such as Lovable, Bolt or Base44, the position is weaker. There is less evidence of human authorship, and the platform's terms decide who has rights in the code and the data. Lovable, for example, takes a perpetual licence to customer data for its own purposes.

Then there is liability for personal data. Escape checked 5,600 public apps built with tools like these and found 175 personal data leaks. The cause is usually the same as at Moltbook: a database with no access rules. For an investor this becomes a GDPR question, with fines attached.

That is why due diligence on a product built with an app builder such as Lovable more often leads to conditions in the deal: a security audit before closing, a plan for moving off the platform, specific warranties on personal data. And "I didn't write a single line of code" in a pitch is a good reason to order that audit before the term sheet.

Without copyright in a vibe-coded product, IP Box relief is in doubt. IP Box is a preferential tax regime that taxes income from qualifying intellectual property, such as patents and copyrighted software, at a reduced rate. The issue is particularly relevant in Cyprus. The Cypriot IP Box exempts up to 80% of the profit from a qualifying asset, so with corporate income tax at 15% the effective rate falls to 3%. European IP Box regimes do cover software, but they require copyrighted software, meaning a program protected by copyright.

If copyright does not protect the product, the company's right to the relief is open to question.

In Poland the 5% rate under PIT and CIT applies to income from "autorskie prawo do programu komputerowego" (copyright in a computer program), so the code itself has to be protected. The closest thing to guidance is an individual tax ruling issued by the National Tax Information service (KIS) on 18 June 2025 (0113-KDIPT2-3.4011.349.2025.2.AK). A programmer working on a B2B basis, who wrote code with an AI assistant, had his IP Box treatment confirmed. It is too early to call this established practice, though: the tax authority did not test whether the code was creative and did not look at the AI question at all. Neither Poland nor Cyprus has a case yet where the relief was withdrawn because of AI.

In venture deals the market's response shows up so far mainly in model documents. Back in 2023 the US NVCA added generative AI representations to its model stock purchase agreement (SPA), including one that the company has not used such tools to create material proprietary intellectual property. A vibe-coded startup can only give that representation with an exception in the disclosure schedule.

M&A practice is tougher. Law firms advise buyers to push for the seller's warranties on IP and data to be treated as fundamental: they survive longer than ordinary warranties, and liability under them can go up to the purchase price. Buyers also ask for specific indemnities, a promise by the seller to cover losses if a particular risk materialises. W&I insurers may exclude AI risks from cover, and then the seller carries them.

Technical audits are getting stricter as well. According to Black Duck, licence conflicts turn up in 94% of the M&A deals it audits. SIG, which runs technical audits for private equity funds, sees roughly twice as many security violations in AI-built projects.

To sum up, founders should address the following issues before an investment round to make due diligence easier:

• move AI tools to business plans and company accounts, with training switched off;
• sign IP assignments with co-founders and contractors that also cover prompts, agent configuration files (such as CLAUDE.md) and projects on platforms;
• move repositories, keys and platform accounts from personal email addresses to corporate ones;
• run the code through a licence scanner that matches code fragments as well as packages;
• commission a security review focused on database access and keys;
• prepare the disclosure: which tools were used, where the AI-generated code sits in the product and who reviewed it.