Training AI on Copyrighted Works: Text and Data Mining Risk Under Kenyan Law

Training AI on copyrighted works raises a question that Kenyan law does not answer directly. The Copyright Act, as consolidated on Kenya Law, has no text and data mining exception and no provision that mentions machine learning or training data. A developer collecting text, images, music or code to train a model must therefore fit the copying within the ordinary exceptions or obtain a licence, and a rights holder must fit an objection within the ordinary infringement provisions. This article sets out what the Act says, what it leaves open, and what the pending Bills are reported to propose.

Why training engages copyright

Section 26(1) of the Copyright Act gives the owner of copyright in a literary, musical, artistic or audio-visual work the exclusive right to control the doing in Kenya of specified acts, including reproduction in any material form and the translation or adaptation of the work. Section 35(1) provides that copyright is infringed by a person who, without the owner’s licence, does or causes to be done an act that the copyright controls. Building a training dataset usually involves copying works into storage and preparing them for processing, so it engages the reproduction right, whether or not the finished model reproduces any work.

The words “in Kenya” in section 26(1) matter. Kenyan copyright controls acts done in Kenya. Where a model is trained abroad on works that originate in Kenya, the acts of copying take place outside Kenya, and the position is governed by the law of the place where they occur. A developer training in Kenya is within the Act, whatever the nationality of the works.

Training AI on copyrighted works: the exceptions that exist

Section 26(3) limits the exclusive rights to the extent provided in the Second Schedule. Part A of that Schedule provides that the exclusive rights do not include the right to control acts done by way of fair dealing for the purposes of scientific research, private use, criticism or review, or the reporting of current events; acts done by way of parody, pastiche or caricature; the right to quote; and certain incidental uses. Paragraph 2 makes these exceptions subject to acknowledgement of the author. Four features limit their usefulness for training:

• No exception names data mining or AI. Nothing in the Second Schedule refers to text and data mining, machine learning or training data.
• Fair dealing is undefined. We found no definition of fair dealing in the Act, and no Kenyan decision applying it to training. Whether non-commercial training for research falls within “scientific research” is untested, and commercial model training is harder to fit.
• Acknowledgement. The requirement to acknowledge the author is difficult to satisfy for a dataset of millions of works, and the Act does not say how it applies to a model.
• Software. Section 26A allows fair dealing with a computer program for the purposes of Part A, and allows a person in lawful possession of a program to make copies necessary for its use in accordance with its intended purpose, such as correcting errors, making a back-up copy or testing. That does not extend to using the code as training data.

What follows for developers

Absent an exception, the safe routes are licences from rights holders or collecting agencies, and training on material that is in the public domain or that the developer owns. A developer working in Kenya should record the source and terms of each dataset, keep evidence of licences and acknowledgements, and avoid copying from sources whose terms prohibit it. Personal data in the dataset raises a separate compliance layer; see our article on data protection for AI training.

What follows for rights holders

A rights holder whose work is copied in Kenya without a licence can rely on section 35(1), but must show an act controlled by the copyright, and evidence of what a developer copied is hard to obtain. Because the Act contains no statutory opt-out, contractual and technical controls, such as website terms and access restrictions, and collective licensing, carry more weight. Questions of who owns what a model produces are separate; see who owns what an AI tool generates and AI-generated content and copyright in Kenya.

What the pending Bills are reported to propose

The Kenya Copyright Board published the Copyright and Related Rights Bill, 2026 for public consultation in March 2026, as reported by law firms. A submission to the Board by KICTANet dated 31 March 2026 describes clause 34 of the Bill as a text and data mining exception limited to non-commercial use, and says the Bill does not mention artificial intelligence. Separately, the Artificial Intelligence Bill, 2026 received its first reading in the Senate on 2 April 2026, and one law-firm summary says it would require records of training datasets to be kept for at least five years. We have not seen the current text or status of either Bill, so these points are reports, not a statement of the law. Check the current position before relying on them.