An artist still has the voice notes. The humming, the half-written second verse, the version they were finally happy with. What they do not have is any record of how their song reached an AI model, or whether anyone was ever entitled to put it there.

Joshua Moraira · Munyeri Advocates · 29 October 2026 · 29 min read

How adequately do the copyright laws of Kenya, Uganda and Tanzania protect creators against unauthorised use of their work in AI training and the outputs that follow? This article examines that question against recent authority from the United States, Germany and the WTO, and against Kenya’s own Copyright Tribunal.

§1  A song, a link, and a question only the records can answer

An artist has a song sitting in their phone. Several versions of it. The first is mostly humming. In the next, they have found the words, although the second verse keeps changing. Eventually, they record a version they are happy with and put it online.

Months later, a friend sends them a link. Have you heard this? It is a song made with an AI music generator. When the chorus comes in, the artist reaches for their own recording. Another voice is singing their words, to their melody.

They still have the voice notes and drafts, so proving the song is theirs is the easy part. Establishing how it reached the new track is harder. Was it copied during training, or supplied by someone requesting the track? Did either use require permission? The records capable of answering those questions are held by people the artist has never dealt with, in a jurisdiction they may never have visited. And had the friend never sent that link, the artist might never have thought to ask.

That gap, between a right that plainly exists on paper and a creator’s practical ability to use it, is the real subject of this piece. The underlying legal question is simple to state: how adequately do the copyright laws of Kenya, Uganda and Tanzania protect creators against unauthorised use of their work in AI training and the outputs that follow? The argument here is that existing reproduction rights already reach much of the copying involved. Their practical value, however, depends on whether creators can discover that copying, test the exceptions an AI developer relies on, and obtain an effective remedy. Reform should close those gaps while preserving genuine room for research and local AI development, rather than rewriting the underlying right.

§2  What fair use decided, and did not decide: Bartz and Warhol

Much of the global conversation about AI and copyright runs through a single American statute. Section 107 of the United States Copyright Act requires courts weighing fair use to consider four factors together: the purpose and character of the copying, including whether it is commercial or nonprofit educational; the nature of the work copied; the amount and significance of what was taken; and the effect on the work’s potential market or value. Research is an illustrative purpose under that section, but simply calling something research has never been enough to establish the defence.

In Campbell v Acuff-Rose Music Inc (1994), the group 2 Live Crew released a commercial parody of Oh, Pretty Woman after the copyright owner refused permission to use it. The Court of Appeals treated the commercial nature of the parody as presumptively unfair. The Supreme Court reversed, holding that all four factors had to be assessed together, and that parody could justify recognisable borrowing where it was being used to criticise the original. Fair use remained an affirmative defence to be proved, not a presumption in the user’s favour, and the case went back down because there was not yet enough evidence about the market for non-parody rap adaptations to decide the point on summary judgment.

Andy Warhol Foundation for the Visual Arts Inc v Goldsmith (2023) sharpened the focus further, onto the particular use being challenged rather than the work as a whole. Warhol had used Lynn Goldsmith’s photograph of Prince to create a series of portraits. In 2016, the Foundation licensed one of them, Orange Prince, to illustrate a magazine piece about the musician, the same kind of use Goldsmith’s own photograph had originally been licensed for. The Supreme Court held that the first fair-use factor favoured Goldsmith, because that particular commercial licence served substantially the same illustrative purpose as her photograph. Giving the work new meaning was not, on its own, enough to settle the inquiry. Importantly, the decision concerned that one licensing use; it left the creation, display and sale of the original Prince Series outside what it decided.

That same discipline, asking what the particular use actually does rather than reasoning from the technology in the abstract, shaped Bartz v Anthropic PBC, the case now cited everywhere AI training meets copyright. A group of authors sued after discovering that Anthropic had assembled a library combining pirated downloads with purchased books, and had used material from that library to train the models behind Claude. On 23 June 2025, the District Court for the Northern District of California refused to treat this as one question and instead separated it into three.

Training copies were found to be fair use. The court treated training a model to generate new text as highly transformative, and treated full-book copying as reasonably, though not strictly, necessary to do that. No infringing public outputs had been alleged in connection with those copies, and factors one, three and four favoured Anthropic on the court’s analysis, even though the nature of the books themselves weighed the other way. Scanning the books Anthropic had purchased qualified too: Anthropic destroyed the physical originals and kept only searchable digital replacements for its own internal use, without distributing them externally.

The pirated copies sitting in Anthropic’s permanent library were a different matter entirely. The possibility that they might later be used for training could not retroactively justify how they had been acquired and kept. The court denied Anthropic summary judgment on that question and sent it to trial, since Anthropic bore the burden on its own motion and any factual ambiguity had to be resolved in the authors’ favour. It is worth being precise about what that denial meant: it did not itself grant judgment to the authors, only that Anthropic had not yet proved it was entitled to win outright. The practical lesson travels well beyond the United States. A developer defending its practices must identify precisely which copies and which uses its justification actually covers, and a claimant must identify precisely which act is being challenged, because acquisition, training, continued retention and the outputs a model later produces can each raise a distinct legal question with a distinct answer.

The training conclusion in Bartz also deserves a second look, because of a tension sitting inside its own reasoning. The court assumed that a licensing market for AI training could plausibly emerge, yet excluded that very use from the market copyright was held to entitle the authors to control. Authors Guild v Google Inc (2015) offers a useful point of comparison on how that boundary has been drawn before. Google had scanned library books without permission from their copyright owners, to power search and limited snippet display. Authors sued, arguing the service substituted for their books and displaced licensing opportunities they could otherwise have sold. The Second Circuit upheld fair use: full-text copying was what made the search function possible, while the restrictions built into the display limited how much of each book’s actual expression anyone could access through it. The licensing arrangements the authors pointed to offered considerably greater access than Google’s limited service did, so they did not establish that a licence was owed for this narrower use, and the court went further still, rejecting any exclusive right to supply mere information about a work of that kind.

Applying that reasoning to generative AI training means examining, system by system, what the model actually enables, what protected expression a user can draw back out of it, and what safeguards limit substitution for the original. Two copies both sitting on a server somewhere does not make two uses equivalent, and an owner’s willingness to charge for a licence does not, by itself, establish a right to control the use in the first place, or licensing could simply absorb the very space an exception is meant to preserve. At the same time, excluding a claimed training market from an author’s protected interests requires a reasoned account of the market evidence, not an assumption. Declaring a use fair and then dismissing the market question because fair uses need no licence risks letting the fourth factor depend on the very conclusion it is supposed to help test. For East Africa, Bartz offers a useful method for separating uses from one another. It does not, by itself, answer whether any particular East African exception would reach the same result, and whatever justification a developer offers must still satisfy the applicable national law, including that law’s own international obligations under the three-step test.

§3  The three-step test’s continuing grip

East African copyright law is not written against the backdrop of the American fair-use factors. It answers to a different and, in some ways, stricter international framework. Article 9(2) of the Berne Convention permits reproduction exceptions only within three cumulative limits: they must be confined to certain defined special cases, they must not conflict with the normal exploitation of the work, and they must not unreasonably prejudice the author’s legitimate interests. Article 13 of the TRIPS Agreement restates these conditions in broader terms, extending them to exceptions affecting exclusive copyright rights generally. Article 10 of the WIPO Copyright Treaty applies the same three-step framework to exceptions under that treaty and under Berne, with agreed statements confirming that existing exceptions may be adapted, and new ones devised, to meet the digital environment. There is room in the framework to respond to AI. It is not unlimited room.

The clearest guidance on how those three conditions actually operate comes not from a copyright case at all, but from a trade dispute. In United States, Section 110(5) of the US Copyright Act (2000), the European Communities challenged American exemptions that let certain business establishments play broadcast music without a licence. The WTO panel held that the three Berne conditions are cumulative, meaning each one must be satisfied independently; failing any single condition is fatal to the exception, however well it performs on the others. On the facts, the panel upheld a narrower exemption tied to ordinary household-type equipment, but found a broader business exemption inconsistent with the TRIPS obligations the United States had accepted.

The panel’s reasoning also leaves genuine room for a measured market inquiry, which matters a great deal for AI. Normal exploitation, on its interpretation, includes plausible future markets of significant economic importance, not merely markets a rightholder currently serves. Commercial benefit to the user of an exception does not, by itself, establish interference with that exploitation. And the third condition tolerates some degree of prejudice to the author; the real question is when prejudice crosses into unreasonableness. These are different questions from the single, integrated weighing exercise fair use performs. Bartz‘s separation of training from permanent pirated-library retention is a useful diagnostic for identifying what an exception would actually need to permit, but a transformative purpose on its own cannot establish compliance with all three Berne conditions at once. Whether a comparable East African exception for AI training would survive that test would require its own distinct assessment, condition by condition.

The treaty commitments actually binding each country also differ in ways reform proposals need to track carefully. All three, Kenya, Uganda and Tanzania, are Berne parties. Uganda is a party to the WIPO Copyright Treaty; Kenya has signed it without ratifying; Tanzania is not listed as a party at all. Uganda and Tanzania also benefit from the TRIPS transition period available to least-developed countries, deferring their obligations under Article 13 until 1 July 2034, or an earlier loss of least-developed country status. That deferral does not suspend either country’s separate Berne commitments, or the operation of its own domestic copyright statute in the meantime; it only postpones the specific TRIPS enforcement timeline. For a proposed East African research exception, the three-step inquiry should ask concretely who benefits from it, what copying it actually permits, and how long copies may be retained under it. Market evidence and safeguards need consideration against each applicable condition separately, and payment can form part of a sound design without curing a failure under a different condition; paying a rightholder does not, on its own, make an otherwise prejudicial exception reasonable. These remain legislative proposals requiring their own justification. Whether existing national exceptions already cover particular AI training activities, without any new legislation at all, is the next and more immediate question.

§4  What East African law already says, and what Aryeh decided

Kenya, Uganda and Mainland Tanzania already protect the reproduction of copyrighted works as a matter of ordinary domestic law. Copying a protected song or book into a training dataset therefore engages an existing right well before any AI-generated output ever appears. The question in each case is simply whether an applicable statutory exception permits that particular copying.

In Kenya, section 26(3) of the Copyright Act (Cap 130) confines the exclusive reproduction right to the extent set out in the Second Schedule, which permits fair dealing for scientific research, private use and a short list of other specified purposes. Training described as research would still need to fall genuinely within that provision and satisfy fairness on its own facts; the label alone proves nothing. Uganda’s approach is structurally similar but adds an open fairness inquiry: section 14 of the Copyright and Neighbouring Rights Act (Cap 222) combines a list of permitted circumstances, including qualified institutional reproduction, with four fairness factors that read like close cousins of the American section 107 test. That resemblance should not be mistaken for an equally open-ended exception; the listed circumstances still have to be met first. Parliament passed a further amendment to the Ugandan Act in March 2026 introducing a new section 14A, restricting reprographic copying to a small fixed proportion of a work within a given period, and permitting licensing schemes for that activity; how it will apply to AI training specifically will depend on the copying actually involved, and on the commencement date once it is finalised.

Mainland Tanzania’s Copyright and Neighbouring Rights Act (Cap 218, Revised Edition 2023) takes a narrower institutional approach. Section 12(f) permits qualified reproduction and electronic storage by public libraries, non-commercial documentation centres, scientific institutions and educational establishments, limited to the needs of each institution’s regular activities and subject to the same normal-exploitation and legitimate-interest safeguards the three-step test itself requires. Section 15 separately permits temporary copies that are strictly incidental to an authorised transmission or to making a work perceptible, conditioned on automatic deletion and no further use. Neither provision was written with a permanent commercial training library in mind, and neither automatically justifies one.

Kenya’s Copyright Act also demands closer attention on the other side of the transaction, when assessing whether an AI-generated output is itself capable of protection. Section 2, paragraph (g), attributes authorship of a computer-generated literary, dramatic, musical or artistic work to the person by whom the arrangements necessary for its creation were undertaken. Section 22(3) separately requires sufficient effort to give the work original character, together with fixation in material form, before any work, however it was made, becomes eligible for copyright at all. The relationship between those two provisions, identifying the arranger under section 2(g) while still having to clear the originality threshold in section 22(3), is precisely the question Kenya’s Copyright Tribunal was recently asked to confront.

In Aryeh Movement Limited v Akoth & another, Tribunal Case E001 of 2025 [2025] KECOT 6 (KLR), a company that had registered a set of literary works found its registration challenged by a contributor who argued the works had been created partly through AI assistance and that she had not been properly credited as a joint author. The Kenya Copyright Board had attempted to resolve the dispute itself, directing the parties to agree in writing on how to split the registered interests or face cancellation. The Tribunal set that directive aside, holding that the Board had exceeded its statutory mandate: deciding contested questions of authorship and ownership was not a function the Board’s registration powers gave it, and registration could not, on its own, conclusively settle a dispute of that kind. On the substantive question, the Tribunal held that section 22(3) presupposes genuine human involvement, so a work generated autonomously by AI does not meet that threshold without more. Aspects of a work produced solely by AI are not eligible for protection unless the party claiming authorship can show sufficient human effort and originality beyond simply entering a prompt. Because neither side had actually produced the disputed works for inspection, the Tribunal left their eligibility to be determined on the evidence in a future hearing, rather than deciding it in the abstract.

Aryeh therefore stands for two connected propositions worth separating clearly. The first is institutional: a copyright regulator exercising registration or administrative powers cannot use those powers to decide what is, at bottom, a judicial question of authorship. The second is substantive: identifying who made the arrangements under section 2(g) does not, by itself, answer whether the resulting work clears the originality bar in section 22(3). Prompts, selection among outputs, editorial choices and subsequent human revision may all bear on that second inquiry, and the more of that work a human contributor can document, the stronger the claim. Section 31 of the Act then addresses ownership of commissioned works and works created during employment, subject to the parties’ own agreement, but a contract allocating copyright between two parties cannot, by itself, make material eligible for protection that the statute does not otherwise recognise.

For the artist in this article’s opening, the practical lesson from Aryeh runs in both directions at once. Good records of their own creative process, the voice notes, the drafting history, the choices they made along the way, would help substantiate a claim over their own contribution if it were ever challenged. But none of that answers the separate question of whether whoever built the AI track had any authority to copy the original song in the first place, even if the resulting AI-generated track turned out to attract no copyright of its own. Eligibility of an AI-assisted output and infringement of the human work that fed it remain two different inquiries, decided on different evidence, and East African creators and developers alike need to keep that distinction straight.

§5  Who gets paid, and when

For the artist, a headline settlement announcement leaves a stubbornly practical question hanging: who has actually been paid, and when?

In Bartz v Anthropic PBC itself, the court approved a $1.5 billion settlement on 20 July 2026, the largest reported copyright recovery in United States history. It was a negotiated compromise rather than a damages award establishing that AI training was unlawful, and its release covered specified past copying and uses, including training, of the books named on the settlement’s own Works List. It expressly excluded claims over AI-generated outputs and granted no licence for any future use. Money had genuinely reached the settlement fund: class counsel’s report of 2 September 2026 recorded $1.05 billion in principal paid into escrow, plus accruing interest. But funding a settlement and distributing it are two separate events, and as at 22 October 2026 the fund administrator still expected the first payments to individual claimants only from 15 November 2026, subject to each claimant’s allocation and payment-information requirements being in order. The records available confirm that the fund was financed. They do not yet confirm that any individual author has been paid.

Judgments require the same caution as settlements, if not more. In GEMA v OpenAI, the Munich I Regional Court recognised damages liability after finding that protected song lyrics had been reproduced both within OpenAI’s models and in the outputs they generated. Its later decision in GEMA v Suno went further, concerning musical works reproduced through training, retention within the model itself, and the outputs produced from it, and imposing a substantial penalty for each future breach. Both judgments were non-final when first announced, and GEMA has since reported that OpenAI is appealing. The published accounts of either case do not establish a quantified payment that has actually reached a copyright owner; a finding of liability and a cheque are not the same thing, and the gap between them can run for years through an appellate process.

Music-industry settlements add a further distinction worth keeping separate again: compensation for past conduct is not the same thing as a licence for future use. Universal Music Group’s October 2025 announcement with Udio expressly identified a compensatory settlement alongside new recording and publishing licences going forward, but disclosed neither the settlement amount nor any evidence that payment had actually reached individual creators rather than the corporate rightsholder.

That last point raises a question distinct from any of the others: who, within a chain of rightsholders, is actually entitled to a share once money does arrive? In June 2026, the American Federation of Musicians sued Universal and Warner, alleging breaches of their collective bargaining agreement. It alleged, on information and belief, that the companies had received substantial compensation from their AI settlements while failing to pass any of it on to the musicians who had actually performed on the recordings concerned. Those allegations are not findings that money was received or improperly withheld; they are, at the time of writing, contested claims. They nevertheless identify a question any East African framework should answer in advance rather than leave to litigation later: what must a corporate rightsholder pass on to the people who performed the underlying work, and on what terms? The answer depends on the particular rights involved and the contracts governing them, not on the existence of a settlement alone. An effective East African compensation framework should make that path traceable from the outset: where payment is legally due, the allocation rules, permitted deductions and distribution dates should be clear in advance, with accounts available for verification, and amounts merely agreed, amounts actually deposited, and amounts actually distributed should be reported as three separate figures, not folded into one headline number.

§6  Five gaps a workable framework must close

Reform in Kenya, Uganda and Mainland Tanzania should address five connected gaps, each requiring implementation through each country’s own laws and procedures, consistent with the treaty commitments that country has actually accepted.

None of these five gaps is best closed country by country in isolation. Regional cooperation could reduce duplicated compliance costs for developers and fragmented enforcement for creators across the three jurisdictions alike. Article 43 of the EAC Common Market Protocol already provides for cooperation among partner states on intellectual property administration, information exchange and enforcement. Common record formats, compatible rights information and coordinated training would give that existing commitment practical effect for AI specifically, though the specific duties this article proposes would still require their own legal implementation in each country’s law, not merely a reference to the Protocol.

§7  Back to the artist

The artist from the opening of this piece already has what the law has always asked creators to keep: the voice notes, the drafting history, proof that the song is theirs. What the current framework does not reliably give them is a way to find out what happened to that song after it left their phone, or an efficient path to a remedy if someone’s use of it turned out to require permission nobody sought. That is not, on the evidence surveyed here, primarily a gap in the substance of East African copyright law. Reproduction rights already reach the copying. Section 22(3) and the Tribunal’s reasoning in Aryeh already set a workable, human-centred line for what AI-assisted output can and cannot claim as its own. The gap sits downstream of all of that, in discovery, in testing the exceptions a developer invokes, and in turning a finding of infringement into money an actual creator receives on a known date, rather than a headline settlement figure sitting in an escrow account somewhere else in the world.

Closing that gap does not require East Africa to copy any single jurisdiction’s answer wholesale. It requires provenance records a creator can actually reach, exceptions built to survive the three-step test rather than assume their way past it, disclosure rules that do not force a claimant to prove the whole case before seeing any evidence of it, licensing that reaches the people who performed the work rather than stopping at the corporate rightsholder, and enforcement powers institutions are willing to exercise strictly within their own statutory limits. None of that is exotic. Each piece already has a foothold somewhere in Kenyan, Ugandan or Tanzanian law. What is missing is the connective work of putting them together before the next friend sends the next link.

Bottom line

East African copyright law is not empty-handed against unauthorised AI use. Reproduction rights already reach the copying behind AI training, and Kenya’s Copyright Tribunal has already drawn a sensible, human-centred line around what an AI-assisted output can claim for itself. What the region lacks is the machinery to make those rights count in practice: a way for a creator to discover that their work was used, a disciplined test for whatever exception a developer invokes, and a reliable path from a finding of infringement to an actual payment. Reform should be built around closing that gap, not around declaring AI training categorically lawful or categorically unlawful before the facts of any particular use have even been examined.

Sources and further reading: Aryeh Movement Limited v Akoth & another, Tribunal Case E001 of 2025 [2025] KECOT 6 (KLR); the Copyright Act (Cap 130) and Evidence Act (Cap 80) via Kenya Law; Bartz v Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.), order of 23 June 2025 and settlement approval of 20 July 2026; Campbell v Acuff-Rose Music Inc, 510 U.S. 569 (1994); Andy Warhol Foundation for the Visual Arts Inc v Goldsmith, 598 U.S. 508 (2023); Authors Guild v Google Inc, 804 F.3d 202 (2d Cir. 2015); United States, Section 110(5) of the US Copyright Act, WTO Panel Report WT/DS160/R (2000); GEMA v OpenAI and GEMA v Suno, Munich I Regional Court; the Berne Convention, Article 9(2); the TRIPS Agreement, Article 13; the WIPO Copyright Treaty, Article 10; Uganda’s Copyright and Neighbouring Rights Act (Cap 222); Tanzania’s Copyright and Neighbouring Rights Act (Cap 218, R.E. 2023); and the EAC Common Market Protocol, Article 43.

Joshua Moraira Joshua Moraira is an Intern at Munyeri Advocates and an LLB Finalist at the University of Embu School of Law, where his work focuses on labour, constitutional and intellectual property law.

This article is provided for general information only and does not constitute legal advice. It reflects the author’s own analysis and opinion on an evolving area of law across multiple jurisdictions, and readers with a specific copyright or AI licensing question should seek advice on their particular facts.