Vibe Coding for Web Developers: Who Owns AI-Written Code and Who Is Liable

Vibe coding has cut the path from brief to working website down to hours. It has also opened questions a standard website development contract never asks: who owns the code that AI wrote? What exactly are you promising the client when you grant an exclusive license to something that is not a copyrighted work at all? And who pays for the damage when a quickly generated solution turns into a security hole?

This article is written for firms and freelancers who build websites to order and who use AI in production. We cover how Czech copyright law treats AI output, what risks open source brings, what liability for defects and security really involves, and which obligations under the AI Act apply from August 2, 2026. At Preegal, we work on IT and AI law every day, so this comes from what we handle with clients in practice.

Vibe coding is a way of building software in which a person describes in natural language what the application should do, and AI generates the actual code. The developer judges the output mainly by whether it works, rather than reading it line by line. Andrej Karpathy coined the term in February 2025, and it has since become a routine working method - from prototypes and landing pages to small plugins, forms, and internal tools.

For firms that build websites to order, the appeal is obvious. Work that used to take two days is finished in two hours, and a tender proposal comes together overnight. But the law does not ask how fast you managed it. It asks what you delivered to the client, what rights you granted, and who bears the consequences when the deliverable fails.

There is a difference between development where AI writes the code but a person reads it, understands it, and edits it, and vibe coding in its pure form, where a person merely tests the output and ships it. The degree of human creative contribution decides both of the things a professional vendor should care about - the copyright in the work, and your ability to defend the quality of what you delivered.

The short answer: no one. Code generated by AI without a human creative contribution is not a copyrighted work, and nobody holds copyright in it.

Under Section 5(1) of the Czech Copyright Act (Act No. 121/2000 Coll.), the author is the natural person who created the work - so a machine, a model, or a company cannot be an author. Under Section 2(2), a computer program is protected as a literary work if it is original in the sense that it is the author's own intellectual creation. The Court of Justice of the EU applies the same originality standard. Where there is no human creative choice behind the output, there is nothing to protect. US courts reached the same conclusion in the Thaler case.

That does not mean the whole project is left unprotected. The following will typically be protected:

• the solution architecture and the data model, where they result from your creative choices,
• portions of code a person wrote or substantially reworked,
• the creative selection and arrangement of components into a working whole,
• graphics, UI design, copy, photographs, and other elements of the site,
• the database as a whole, under the sui generis database right (Section 88 et seq. of the Copyright Act).

The prompt itself, however long and however carefully crafted, creates no protection in the result. Copyright protects a specific expression, not a brief and not an idea.

A common licensing clause in website development contracts reads roughly like this: “The contractor grants the client an exclusive and unlimited license to all outputs of the work.” For AI-written code, that sentence cannot be performed. A license is permission to exercise a right - and no right exists in unprotected code. You are promising something you cannot deliver, and opening the door to liability for legal defects in the work.

The practical consequence for the client is worse still. Unprotected code can be copied and used by anyone, including their competitors. If the client paid for a “unique custom solution,” they should know this up front - not when a competitor launches the same functionality.

What to put in the contract:

• separate the deliverable into parts protected by copyright and parts without protection, and tie the license to the former,
• add a statement that AI tools were used in creating the work, and define the extent of that use,
• for the unprotected parts, offer commitments you can actually keep instead of an unperformable license: confidentiality, an undertaking not to supply the same solution to the client's competitors, and delivery of the complete source code and documentation,
• clarify who holds access to accounts, repositories, domains, and API keys once the engagement ends,
• consider an indemnity clause covering third-party claims arising from infringement of rights in the output.

Apply the same logic inside your own company. For employees, rights are handled by the employee work regime under Section 58 of the Copyright Act; with external developers working as sole traders, however, you must expressly agree a license or an assignment of the economic rights. Otherwise you have nothing to pass on to the client.

Large language models were trained partly on public repositories. They will occasionally generate a passage identical or very similar to existing code that is protected by copyright and distributed under a specific open-source license. The model will not hand you the license header or a warning.

Responsibility for complying with license terms rests with whoever distributes the code - the vendor, and after handover the client as well. With permissive licenses (MIT, BSD, Apache 2.0), this usually means attribution and including the license text. With copyleft licenses, especially GPL and, for web applications, AGPL, a breach can trigger an obligation to make the source code of the entire derivative work available. For a proprietary solution delivered to a client, that is a serious problem, commercially and legally.

The minimum that should run on every project: automated software composition analysis (SCA), an inventory of the libraries used and their licenses (SBOM), an internal list of prohibited licenses, and a check that the package the AI suggested actually exists and is maintained. Models sometimes invent library names, and attackers exploit this by publishing that non-existent package for real - with malicious content.

You are liable to the client for the deliverable regardless of what you built it with. A contract for work under Section 2586 et seq. of the Czech Civil Code requires the contractor to perform properly; liability for defects and for damage is strict, and in the eyes of the law AI is a tool, not a legal person you can shift responsibility onto.

Practice also shows that applications built this way share a typical set of weaknesses: missing server-side authorization checks, credentials and API keys written straight into the code or the frontend, unsanitized inputs, broken validation, needlessly exposed endpoints, and outdated dependencies. AI writes code that works for the expected scenario; it will not invent the hostile one on its own. If the site also ships AI features such as a chatbot or content summarization, the risk of indirect prompt injection is added.

When such a flaw leads to a data breach, further layers of liability follow:

• the GDPR - the client is usually the controller and the vendor the processor; without a data processing agreement, both sides have a problem,
• Act No. 264/2025 Coll., the Czech Cybersecurity Act, effective November 1, 2025: if you supply a regulated entity, security requirements reach you contractually through the supply chain,
• reputational damage, which often costs more than the compensation itself.

The answer is not to stop using AI, but to have a process: mandatory human code review for everything that goes to production, automated security testing, a liability cap in the contract, and professional indemnity insurance sized to your actual volume of work.

Ask AI for “a stronger email capture” or “higher cart conversion” and you will get exactly what the model saw on thousands of websites - including practices that are prohibited in the EU. A pre-ticked consent box. A “Reject” button two levels deeper than “Accept all.” A countdown that resets when the page reloads. Unsubscribe buried in account settings. The model does not distinguish between what lifts the numbers and what is lawful. You have to.

This is already enforceable today, under several sets of rules at once:

• misleading and aggressive commercial practices under Act No. 634/1992 Coll., the Czech Consumer Protection Act - these apply to every online store regardless of size, and are enforced by the Czech Trade Inspection Authority,
• Article 25 of the DSA - providers of online platforms must not design their interfaces so as to deceive or manipulate users,
• consent under the GDPR and the cookie rules - consent must be freely given and refusing must be as easy as accepting; unequal buttons in a cookie banner are a classic finding by the Czech Data Protection Authority,
• the rules on price reductions and reviews in force since January 2023 - the lowest price of the preceding thirty days, a ban on fake reviews, and a ban on costs revealed only at the final step of the order.

The forthcoming Digital Fairness Act is meant to tackle dark patterns comprehensively, but the Commission plans to table it only at the end of 2026, and it still has a long road ahead before it applies. The argument that “there is no rule on this yet” therefore does not hold today. For the wider picture, see our glossary entry on AI and consumer law.

Watch how the consequences split. The fine goes to the client, because it is the client who acts as the site operator toward consumers and regulators. It comes back to you as a warranty claim - you delivered something that cannot be operated lawfully, and that is a defect. So the contract should also address who is responsible for the interface complying with consumer law, and your process should include a review of the cookie banner, consents, and checkout before the site goes live. And when the client asks for a dark pattern themselves, get their instruction and your warning in writing.

The Artificial Intelligence Act (Regulation (EU) 2024/1689) does not regulate vibe coding as such. It does regulate what leaves your company, and how you work with AI inside it.

Already in force. Since February 2, 2025, anyone who deploys or uses AI systems must ensure a sufficient level of AI literacy among their people (Article 4). For a firm where half the team prompts daily, that means documented training and internal rules, not just a good feeling. That is exactly what our courses and training are built for, in particular the one-day AI and Law training we run together with Skrivanek - its certificate serves as evidence of AI literacy training.

From August 2, 2026. The transparency obligations under Article 50 apply. For the sites you build, this means in practice that a chatbot or voice assistant must tell the user they are communicating with AI, and artificially generated or manipulated content must be marked in machine-readable form. The single exception concerns the machine-readable marking under Article 50(2) for generative systems already on the market before August 2, 2026 - those have until December 2, 2026. The duty to disclose that the user is interacting with AI has no grace period, and systems deployed from August 2, 2026 must meet the marking requirement immediately. Breaches carry fines of up to EUR 15 million or 3% of worldwide annual turnover. The scope of these obligations is set out in the European Commission guidelines adopted on July 20, 2026; for the wider context see our article AI Compliance.

The other half of the risk sits not in the output but in the input. When you paste a client's source code, internal documentation, or customer database into an AI tool, you may breach a confidentiality clause and the protection of trade secrets under Section 504 of the Czech Civil Code. If the input contains personal data, you need a legal basis to pass it to the tool provider, and usually a data processing agreement as well.

Set up three things that cost nothing and solve a great deal: corporate accounts for AI tools with training on inputs switched off, a written internal policy on which tools are permitted and what data must never go into them, and one sentence in the client contract by which the client approves the use of AI and you define its scope.

• Revise the licensing clauses in your standard contract for work and separate protected from unprotected parts of the deliverable.
• Add a statement on the use of AI to the contract, together with the client's consent to it.
• Agree licenses with external developers in writing; for employees, verify the employee work regime.
• Make human code review mandatory for everything that goes to production.
• Run automated license and dependency checks (SCA, SBOM) and prohibit copyleft licenses in proprietary projects.
• Verify that packages suggested by AI actually exist and are maintained.
• Use corporate accounts for AI tools with training on inputs switched off.
• Write an internal policy on permitted tools and prohibited data, train the team, and document the training (Article 4 of the AI Act).
• Review the cookie banner, consents, and checkout for dark patterns before launch.
• Label chatbots and generated content on client sites by August 2, 2026.
• Check the liability cap in your contract and the scope of your professional indemnity insurance.

Vibe coding is not a legal problem in itself. The problem arises where delivery speed outruns documentation, the contract, and review. A vendor who has those three in order can use AI to the full - and turn it into a competitive advantage, because they can explain to the client exactly what they are buying.

Using AI in development and want your contracts and processes in order? Get in touch.
We review contracts for work and licensing terms for website vendors, set up internal rules for working with AI tools, train teams, and prepare AI literacy documentation under the AI Act. Before you sign another engagement with a clause granting an exclusive license to all outputs, talk to us about what you are actually delivering to your clients.