Vibe Coding: What It Is and What Changes with Claude's Watermark
Vibe coding: what it is and what changes with Claude's watermark
Describe what you need and the AI writes the code. Now Anthropic signs Claude's texts with an invisible watermark: let's see what it really means for those who develop software this way.
Key takeaways (TL;DR)
Vibe coding is the practice of creating software by describing in natural language what you want to achieve, letting AI write the code. Here are the four key points of this article:
- What it is: a term popularised by Andrej Karpathy in February 2025; it is done with tools such as Claude Code, Cursor, GitHub Copilot, Replit and Lovable. The idea, the requirements and the verification remain human.
- The novelty: from 11 August 2026 Anthropic marks the texts of new Claude models with an invisible statistical watermark, in implementation of Art. 50 of the AI Act.
- It barely applies to code: where an exact syntax is required there are no choices between equivalent words to steer, so code generated with AI carries a minimal or absent signal.
- Ownership does not change: Anthropic's commercial terms assign outputs to the customer. No royalties, no "Unity case".
What is vibe coding: meaning and origin
Vibe coding is a way of creating software in which a person describes in natural language what they want to achieve and an artificial intelligence writes the code. The term was popularised in February 2025 by Andrej Karpathy, former AI Director at Tesla, and within just a few months it entered the vocabulary of anyone building prototypes, internal apps or entire products. The meaning lies entirely in the word "vibe": you start from a general idea and refine it through dialogue, without focusing on syntax.
What tools are used? The most widespread today are Claude Code, Cursor, GitHub Copilot and Windsurf for those working in an editor, and platforms such as Replit and Lovable for those starting from scratch without a development environment. Vibe coding with Claude is among the most widely used approaches in professional settings, as it operates directly on the project repository.
How does it work in practice? Four steps that repeat in a cycle: you describe the result you want to achieve; the AI generates a first version of the code; you test it and ask for corrections or changes; you repeat until the result is acceptable. The key concept is direction: the idea, requirements, product decisions and verification remain human. This is a distinction that also matters from a legal standpoint, as we will see shortly.
Do you still need to know how to programme? For a prototype or an experiment, less and less so. For a production product, yes: code generated quickly and never reviewed can hide security, performance and maintainability issues. The difference between a demo and reliable software remains the same as ever: review, testing and well-thought-out architecture.
Vibe coding in practice: who does what
-
The idea and direction remain yours
Requirements, priorities, product decisions: this is the part that AI cannot replace, and it is what gives value (and authorship) to software.
-
AI writes, you verify
Generated code must be read, tested, and corrected. Human review is what separates a working prototype from a reliable product.
-
The value lies in the project
Architecture, technical choices, and domain knowledge cannot be generated with a prompt: they are the true asset of those who build software.
How does Claude's watermark work?
On August 11, 2026, Anthropic announced that Claude models launched from August 2, 2026 onwards incorporate an invisible statistical watermark in the texts they generate. The Claude watermark was introduced to comply with the transparency obligations of Article 50 of the European AI Act, applicable from August 2, 2026 (with a grace period for watermarking until December 2, 2026), and for now it is applied worldwide, across all products: API, Claude app and Claude Code included.
This is not metadata or hidden characters. The mechanism, similar to Google DeepMind's SynthID-Text technique, acts during generation:
Choices among equivalent words
While writing, the model continuously encounters low-impact choices between synonyms and interchangeable phrasings.
An invisible statistical pattern
These choices are guided according to a scheme that does not alter the quality and readability of the text, but leaves a measurable trace.
Verifiable only with the key
No reader can notice it: the detection key is needed to verify whether a text has passed through Claude. The pattern travels with copy and paste.
And what about the code? Here is the key point for those who do vibe coding: the watermark is applied only where there are free choices between equivalent alternatives. Where an exact syntax is required — that is, in most of the source code — the watermark is not applied or is very weak. Anthropic itself clarifies this in the official announcement. The traces are found, if anything, in comments, documentation and interface texts. If the topic is the publication of AI-generated content, we discussed it in detail in this in-depth article on watermarking in texts.
Who owns AI-generated code? The fear of a 'Unity case'
The question circulating among those who develop with AI is legitimate: if the model provider can recognize its own outputs, could it one day claim a share of revenues, as Unity did in 2023 with its runtime fee? The answer, looking at the sources, is no. For three distinct reasons.
1. A watermark is not a proof of ownership. Anthropic's announcement states it explicitly: the watermark only serves to verify whether content may have passed through Claude, «it says nothing about ownership or authorship». It is a transparency tool required by the European regulator, not a claim.
2. The commercial terms say the opposite. Anthropic's Commercial Terms assign to the customer all rights over the outputs and declare that Anthropic waives any rights over the customer's content. There is no royalty or revenue share mechanism: you pay for the service, the result is yours.
3. The Unity case was a license change, not a technology. The runtime fee arose from a unilateral modification of contractual terms. A watermark, technically, cannot generate economic claims: only a contract can do that, and the current contract says the opposite. Moreover, what makes software protectable is precisely the human part of vibe coding: idea, architecture, creative direction, selection and revision.
The practical advice remains the same as always: keep an eye on the terms of service at renewals, as with any critical supplier. On the way we position ourselves with respect to European software regulations, you can find our Security and Compliance page.
Frequently asked questions about vibe coding and watermark
It is not code ownership, but its quality: rapidly generated code that is never reviewed can hide vulnerabilities, technical debt, and maintainability issues. For a prototype it is acceptable; for a production product, human review, testing, and a solid architecture are required.
The most widely used are Claude Code, Cursor, GitHub Copilot, and Windsurf, which work inside the editor or on the repository, and platforms such as Replit and Lovable, designed to create complete apps starting from just a prompt. The choice depends on how much technical control you want to retain over the project.
To a minimal extent. The watermark only acts where there are free choices between equivalent formulations: where an exact syntax is required, as in most source code, it is not applied. The traces are concentrated in comments, documentation, and interface texts.
Whoever creates and directs it. Anthropic's commercial terms assign to the customer all rights over the models' outputs, and the protectable part of the software is in any case the human one: idea, requirements, architecture, and review. The watermark changes none of this.
Not today: there is no royalty mechanism and the commercial terms state that outputs belong to the customer. The Unity case arose from a unilateral licence change, not from a recognition technology. Any such claim would require a rewriting of the contracts in the opposite direction to the current one, which would in any case only apply to future outputs.
For source code, no labelling obligation: the AI Act's marking obligations fall on the model provider, which is why Anthropic introduced the watermark. Article 50 concerns synthetic content such as images, audio, video, and texts published on matters of public interest: if you use AI for that content, that is where attention is needed.
Do you have a prototype born from vibe coding?
Turning it into a stable, secure and compliant product is the work of a software house. A single technical point of contact follows the project from analysis to delivery, and the code and intellectual property remain yours.
Sources: official Anthropic announcement «How Claude's text watermarking works» (August 11, 2026); TechCrunch, «Anthropic says it will watermark text generated by its AI models» (August 11, 2026); European Commission, Code of Practice on Transparency of AI-generated Content; artificialintelligenceact.eu, guide to article 50; Anthropic Commercial Terms of Service. Information accurate as of August 18, 2026: for updates please refer to the original sources.
