AI detection software and authors: what every entrepreneur needs to know when an algorithm scuppers a multi-million contract

Imagine this: you’ve spent years working on a product, it’s generating massive interest in the market, and just as you’re about to sign on the dotted line, the biggest player pulls out. Not because of the quality, not because of the price — but because of an algorithm.

That’s what happened to 25-year-old thriller writer Jerry Falade. His debut novel sparked a bidding war between fourteen international publishers (Headliner.nl). The winning bid stood at $2.4 million, including film rights. Shortly afterwards, the deal fell through. An AI detection programme had labelled 97 per cent of the text as machine-generated. Lead publisher Minotaur Books (US) withdrew. Falade categorically denies using AI, but to no avail.

This case is more than just a writer’s anecdote. For entrepreneurs working with content, intellectual property or creative products, it raises an urgent question: what is the legal and commercial value of a text flagged by an AI detector, and how do you protect yourself?

The copyright risk of AI-generated content

Publishers and investors fear not only reputational damage should it emerge that a text was written by AI. Above all, they fear the loss of legal protection.

Copyright arises the moment a human being makes an original, creative choice. That is the core of the system, both in the Netherlands (Copyright Act) and in virtually all other countries. A text generated entirely by an AI model falls outside that protection: there is no human creator, so there is no rights holder. The result is that such a work immediately falls into the public domain and is free of rights.

For entrepreneurs in publishing, content marketing, e-learning or software development, this has concrete consequences:

• No exclusivity. Without copyright, a competitor can adopt or reproduce the work without a licence.
• No royalties or licence revenues. The economic exploitation of the work is largely lost.
• Risk of breach of contract. Where an agreement is based on copyright-protected content and that protection turns out to be lacking, this may constitute breach of contract.

In short: entrepreneurs who use AI to produce content bear an increased legal risk, even if they themselves are convinced of the human origin of the work.

How does an AI detector work — and why is the result unreliable?

AI detection software does not analyse intentions. It analyses statistics. Two metrics are central:

· Perplexity measures the predictability of word choices. Large language models systematically choose the most probable next words, resulting in text that feels fluent and smooth. Human writers are less consistent in their choices and therefore produce text that is statistically ‘rougher’.

· ‘Burstiness’ describes the variation in sentence length and rhythm. People write erratically: a brief observation, followed by a detailed explanation, then a powerful conclusion. AI models generally produce a more regular pattern.

If a text scores low on both metrics, the detector flags the text as likely to be AI-generated. This is emphatically not proof: it is a probability assessment. However, this nuance is lost as soon as a score is used as a factual given in a business or academic context.

The margin of error is high

Scientific research has repeatedly scrutinised the reliability of AI detectors, with worrying findings:

• Non-native speakers are disproportionately affected. A study by Stanford University (2023) showed that texts by non-native English-speaking writers are significantly more likely to be flagged as AI-generated. This raises anti-discrimination issues that are becoming increasingly relevant in legal proceedings (AI-Detectors Biased Against Non-Native English Writers | Stanford HAI).
• Classic literature scores highly on AI tests. Passages from works by Hemingway, Orwell and even the Bible are flagged by certain tools as machine-generated — which seriously undermines the validity of the underlying model.
• Clear, structured text is viewed with suspicion. Instructional and legal texts, written in clear Dutch, consistently score higher on AI indicators. Those who write clearly are penalised.

• The same text, completely different scores. Common tools such as GPTZero, Turnitin AI, Originality.ai and Copyleaks (used for English) regularly produce widely varying results for identical texts. They are scarcely comparable with one another. This therefore offers hope to anyone who has had their thesis rejected because of such AI text detection software. An overview from the US can be found at AI Cheating Lawsuits Tracker — Every Case, Who Won (2026) – GradPilot.
• An example of a case from the US (student v university): a high-profile American court case was Matter of Newby v. Adelphi University (January 2026).
• An autistic student (Newby) was accused of fraud by his university after the well-known detection tool Turnitin flagged his essay as 100% AI-generated. Newby categorically denied this.
• The investigation into the software: In this case, the reliability of the AI detection software itself was assessed by the judge. Scientific evidence was presented showing that the software has a huge margin of error and that, in particular, neurodivergent writers and non-native speakers (due to their specific, structured way of phrasing things) are wrongly flagged as AI.
• The ruling: the judge overturned the university’s penalty and described the decision to place blind trust in AI detection software as “without valid basis and devoid of any reason”.

The situation in the Netherlands: no court cases (yet), but real risks

For Dutch-language texts, most popular detectors have been trained primarily on English-language material. Tools such as Copyleaks and Scribbr explicitly support multilingualism and therefore generally perform better with Dutch, but even they are far from flawless.

Unlike in the US, the UK and Australia — where students have already been penalised on the basis of AI detection scores from tools such as Turnitin and GPTZero — there are as yet no published court rulings in this area in the Netherlands. There are several reasons for this.

Firstly, publishers and clients prefer to resolve such disputes through contractual negotiation or tacit rejection (in the case of educational institutions, by means of newly written master’s theses or final dissertations, or by providing sound counter-evidence). Legal proceedings generate publicity that is undesirable for all parties involved. This is presumably also the case at Dutch educational institutions.

Secondly, the evidential position is structurally problematic. An author or content creator who wishes to demonstrate that they wrote the text themselves must, in effect, prove that the software has made a mistake. However, the software is too unreliable for this purpose; and an instrument that is too unreliable cannot provide conclusive evidence.

Thirdly, there is no specific legal basis. There is no provision in Dutch law that recognises an AI detection score as a legal basis for the loss of copyright or contractual liability. As long as this basis is lacking, litigation on this point is legally complex and financially unattractive.

However, the fact that no legal proceedings have been brought does not mean that the risk is imaginary. In the academic world, dissertations and master’s theses are regularly rejected or assessed on the basis of AI detection. This has far-reaching consequences for students who claim to have written the work themselves.

What does this mean for businesses in practice?

For businesses involved in content production, intellectual property or creative services, the following considerations are relevant:

1. Document the creative process. Keep a version history, notes, research material and correspondence. This serves as indirect evidence of human involvement, should an AI detection score ever be called into question.

2. Set out clear contractual agreements regarding the use of AI. Specify the extent to which AI tools have been used, for which parts of the work and in what capacity. Transparency up front prevents disputes later on.

3. Be critical of AI detection as a contractual condition. If a client or publisher makes the acceptance of work contingent on an AI detection score, legal advice is advisable. Such a clause may be unreasonable, particularly given the proven margins of error in the software used.

4. Keep a close eye on copyright protection. Anyone using AI as a writing tool must guard against blurring the line between AI as a tool and AI as the author. The former is perfectly acceptable; the latter undermines the legal protection of the final product.

Q&A: frequently asked questions about AI detection and copyright

Does a text written partly with AI still enjoy copyright protection?

Yes, provided there is sufficient human creative input. AI as a tool — for structure, word suggestions or research — does not in itself affect copyright protection. However, if the AI generates the vast majority of the content without substantial human editing, the protection lapses. The line between the two is not legally defined and is assessed on a case-by-case basis.

Can a publisher or client terminate a contract on the basis of an AI detection score?

In principle, this is not possible solely on the basis of that score. An AI detection score is not legal proof of AI use; it is a statistical estimate with a proven margin of error. Termination on that basis is contestable, particularly if the author can demonstrate, with justification, that the software is unreliable or produces a false positive.

What are the risks if I use AI for content production in my business?

The main risks are: loss of copyright protection for content generated entirely by AI, contractual liability if an agreement requires human authorship, and reputational damage with clients or partners. Ensure transparency and set out the use of AI in the contract.

Is an AI detection score admissible as evidence in a court case?

In the Netherlands, this is still uncharted legal territory. Given the scientifically documented unreliability of these tools — including high error rates for non-native speakers — it stands to reason that judges will not accept such a score as stand-alone evidence. At best, it may serve as an indication within a broader body of evidence.

What can I do if my work is wrongly classified as AI-generated?

Gather documentation of the writing process: draft versions, notes, emails, source research. Have the text analysed by several AI detectors and record the varying scores. Engage a solicitor if a business relationship or contract is at stake. The inconsistency between different tools is, in itself, a strong argument against the reliability of the result.

Is there legislation in the Netherlands regulating the use of AI in writing?

No, there is currently no specific legislation on this matter. The existing Copyright Act provides the framework for copyright protection, but contains no provisions regarding AI detection or the use of AI as grounds for the loss of rights. The European AI Act introduces transparency obligations relating to AI-generated content, but is primarily aimed at system providers, not individual users.

Contact

Do you have questions about this article, or would you like advice regarding your situation? Please contact us by email or phone. You may also wish to read my article on AI applied to (human-written) text and AI-generated images from the US case “Zarya of the Dawn” (Copyright Office, 2023).