Two Ratios, One Gap
On July 21, 2026, a US federal judge granted final approval to Anthropic's $1.5 billion settlement of Bartz v. Anthropic, the copyright class action brought by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson. The court found that training Claude on books was fair use, but that Anthropic's acquisition of millions of titles through shadow libraries LibGen and PiLiMi was not — and priced that piracy at roughly $3,000 per work across a class of some 500,000 works, four times the statutory minimum for willful infringement (Silicon Republic). It is, by that outlet's account, the largest copyright settlement in US history — and it exists only because a class was certified in July 2025 and every affected author was swept in automatically unless they opted out.
Taiwan's authors have no such mechanism. That gap is not a matter of Taiwan's law being more permissive toward AI trainers — if anything, the opposite is true. It is a matter of remedy architecture.
A Stricter Starting Line, No Finish Line
Taiwan's Intellectual Property Office (TIPO) addressed AI training data in a June 16, 2023 interpretive circular (Shou-Zhi-11252800520): feeding copyrighted works into a model constitutes "reproduction," and absent a license or one of the fair-use carve-outs in Articles 44–65 of the Copyright Act, that reproduction infringes (Lee, Tsai & Partners). Unlike the Bartz court, TIPO has never granted AI training a fair-use safe harbor of its own — it treats each case as requiring individual judicial determination, but the default presumption leans toward infringement, not fair use. On paper, a Taiwanese author's legal claim against an AI trainer starts from firmer ground than a US author's did before Anthropic settled.
What Taiwan's authors lack is any efficient way to convert that claim into compensation at scale. Taiwan's Code of Civil Procedure permits group litigation under Article 44-2, but only on an opt-in basis: each claimant must file their own pleading naming the facts, evidence and relief sought after a court-published notice, rather than being bound automatically like a US Rule 23 class (Laws & Regulations Database, Code of Civil Procedure). The Consumer Protection Act lets a qualified advocacy group sue on behalf of 20 or more injured consumers who assign it their claims, with punitive damages up to five times actual loss for willful misconduct — but that mechanism, under Article 50, is built for trader-consumer disputes, not authors against AI labs (Laws & Regulations Database, Consumer Protection Act). Copyright's own mediation channel under Article 82 exists and is almost never invoked.
The cost of the alternative — individual litigation — is visible in the Lawsnote case. In June 2025 the New Taipei District Court found that legal-tech startup Lawsnote had unlawfully scraped a rival database, awarding NT$104.5 million in civil damages and jail terms of two to four years for its founders; the case is now on appeal to the Intellectual Property and Commercial Court (whitehsu.blog). That result took roughly two years, criminal-level enforcement resources, and benefited exactly one corporate plaintiff. It is not a template 500,000 individual authors — or even 500 — could realistically use.
Legislating Around the Question, Not Into It
Taiwan's newest AI law does not fill this hole, because it was not built to. The Legislative Yuan passed the AI Basic Act on December 23, 2025, and it entered into force January 14, 2026, tasking the Ministry of Digital Affairs (MODA) with a risk-classification framework and requiring an annual national AI strategy committee (MODA). MODA's parallel response to training-data friction is the Taiwan Sovereign AI Training Corpus License, a voluntary licensing framework built with TIPO to let government-held data flow to AI trainers on agreed terms (MODA). It is a sensible tool for public-sector data. It does nothing for the novelist or journalist whose commercially published book was scraped by a private lab with no license offered at all.
The Case For an Aggregation Mechanism — and Against Importing the US One
There is a real case for closing this gap. Individual authors cannot afford to litigate against well-capitalized AI developers one claim at a time, and a regime that presumes infringement but offers no practical way to enforce it functions, for most rightsholders, like no regime at all. Deterrence requires that infringement actually cost something in aggregate, not just in theory.
But Taiwan should not simply transplant the US opt-out class action. That model produces enormous, unpredictable liability events — useful for punishing proven bad-faith piracy like Anthropic's shadow-library sourcing, but a poor fit for a smaller market actively trying to build a sovereign AI sector around initiatives like MODA's training corpus. The more proportionate fix sits inside tools Taiwan already has: extend the Copyright Collective Management Organization Act's existing collective-licensing infrastructure — already used for music and broadcast rights — to AI training, so publishers and authors can license collectively and split proceeds, rather than each having to prove reproduction and sue individually. Pair that with a clarified statutory-damages floor specific to AI-scale reproduction, high enough to make a single author's claim worth bringing without requiring class-action machinery Taiwan's civil procedure was never designed to carry. That gets closer to what the Bartz settlement actually delivered — a working payout mechanism — without importing the litigation overhang that a settlement of that size signals to any AI developer eyeing the Taiwanese market.