AI Training Data and Copyright in Nepal: The Questions Every Builder Is Asking
What the Copyright Act 2059 does and does not say about AI training data copyright in Nepal, and the practical answers Nepali AI teams and creators use today.
A Kathmandu team is fine-tuning a model on scraped Nepali news, song lyrics and Facebook commentary. A Nepali illustrator finds her thangka-inspired artwork inside someone's dataset. Both ask the same thing: what does Nepali law actually say about AI training data — and nobody can point them to a section that answers it.
Nepal's Copyright Act, 2059 (2002) was written decades before generative AI. That does not make these questions academic; it makes them unsettled, and unsettled questions need a practical playbook rather than invented answers.
In short
AI training data copyright in Nepal is governed by an Act that predates AI: the Copyright Act, 2059 (2002) protects works automatically from creation, is silent on datasets, and has no Nepali case law settling how training and outputs fit its reproduction rights. The practical position: treat scraped content as copied until licensed, remember that copyright protects expression rather than ideas or styles, assume AI-generated output raises unresolved ownership questions, and manage the risk with licences, documentation, registration and platform takedowns rather than litigation.
Why the Copyright Act 2059 is silent on AI training data
The Copyright Act, 2059 (2002) follows the Berne principle — protection is automatic from creation, with no formalities — and gives rightsholders control over reproduction, distribution and adaptation of their works. It contains no provision addressing machine learning, datasets or computer-assisted creation, because none existed when it was drafted.
Two consequences follow. First, the general rules apply: a work copied into a training corpus is copied, whatever the purpose. Second, none of this has been settled in Nepali practice — there is no Nepali precedent yet that tells a Kathmandu AI team exactly where training ends and infringement begins. Honest planning works inside that uncertainty instead of pretending it away.
Question one: is scraping Nepali content into a dataset copying?
Nepali websites, news portals, song catalogues and image libraries are protected automatically, whether or not they display a copyright notice. Copying them — scanning, downloading, building a corpus — engages the reproduction right on its face. There is no Nepali provision that says training is an exception.
Nepal's recognised exceptions are narrow and category-based — the education, review and news contexts the site's fair-use guide covers. Commercial model training does not obviously fit any of them. The risk is not theoretical: the same scraping that builds a model can also reproduce the exact expression the rightsholder sells.
Question two: does an AI output infringe when it resembles a specific work?
Copyright protects expression, not ideas, methods or styles. An output that merely shares a general style with a training artist is a different question from one that reproduces a substantial part of a specific photograph, lyric or illustration. Nepali infringement analysis — what counts, what remedies follow — is the same framework described in the site's copyright infringement guides, applied to an unfamiliar production method.
For creators, the practical implication cuts both ways: a distinctive, specific work is the easiest to defend, while general artistic style remains outside copyright's fence regardless of how a copycat imitator — human or machine — borrows it.
Question three: who owns what an AI system produces?
Nepal's Act vests copyright in authors — human creators — and Nepal's creator-first defaults mean employees and freelancers keep rights unless written assignments say otherwise. Where a work is generated substantially by a machine, the ownership question has no clean Nepali answer yet; the site's dedicated guide on computer-generated content walks through the positions.
Teams should not let the uncertainty accumulate silently: document who prompted, curated, selected and edited, and paper the employment and contractor chains so that whatever the final rule turns out to be, your company can show the basis of its claim.
| Question | Where Nepali law stands | Practical answer for now |
|---|---|---|
| Is scraping into a dataset copying? | Reproduction right applies on its face; no training exception exists | Licence the corpus, or use licensed and public-domain data |
| Are AI outputs infringement? | Turns on similarity to a specific protected work | Audit outputs against prominent sources before release |
| Who owns AI-generated content? | Unsettled; the Act vests rights in human authors | Document human contribution; assign rights in writing |
| Does fair use cover training? | Exceptions are narrow and category-based | Do not assume a commercial-training exemption |
| What evidence helps creators? | Registration is voluntary but creates dated proof | Register key works at the Copyright Registrar's Office |
AI training data copyright in Nepal: what builders should do before training
Build the data file like a diligence file: record where each corpus came from, what licence covers it, and who approved it. Prefer licensed datasets, public-domain material and content you commissioned with written assignments. Where a dataset must include third-party works, negotiate — licences from Nepali publishers, labels and archives are cheaper than disputes with them.
What Nepali creators can do when their work ends up in a dataset
Your rights exist from creation, but proof wins disputes. Keep originals and drafts, and register the works that matter at the Copyright Registrar's Office — the voluntary certificate is the dated, government-issued record that shifts the argument in your favour. For published misuse, platform takedown processes remain the fastest remedy, followed by a formal demand and the infringement remedies the Act already provides.
And clear your own house first: creators launching AI-assisted products face the same naming and branding battles as anyone else, and a product name that conflicts on the trademark register is a dispute that arrives before the copyright one.
Want this handled for your brand?
IP Watch monitors new trademark publications in Nepal and alerts you to potentially conflicting marks — with the context needed to review them.
This article is general information, not legal advice.