What is extended collective licensing for AI training, and could it actually work?

Extended collective licensing (ECL) has become one of the most debated proposals for resolving the standoff between generative AI developers and copyright holders. The core idea is simple on paper: a collective management organization (CMO) negotiates licenses on behalf of large groups of rightsholders, and an 'extended' mechanism allows the license to cover creators who are not members of that organization. As of August 2026, jurisdictions including the Nordic countries, the United Kingdom, and several EU member states are actively weighing whether ECL can be adapted to AI training at scale — and the honest answer is that it is feasible in some sectors and deeply flawed in others.

What Extended Collective Licensing Actually Is

Also worth reading: How Do AI Trademark Review Services Actually Work in 2026? · How Do You Build an AI Trademark Clearance Checklist That Actually Reduces Risk? · How Do You Actually Evaluate AI Counterfeit Detection Tools in 2026?

Extended collective licensing originated in Nordic copyright law in the 1960s as a practical solution to mass uses of works where individual licensing was impossible. Under a classic ECL scheme, a CMO representing a substantial share of rightsholders in a given field negotiates a license with users. The 'extension' element then applies the terms of that license not only to members but also to non-member rightsholders in the same category, both domestic and sometimes foreign, provided they have no reasonable ability to opt out or their interests are adequately represented by the negotiating CMO.

The mechanism rests on three pillars: representative consent (the CMO must genuinely represent most rightsholders), legal presumption (non-members are presumed covered unless they opt out), and remuneration distribution (non-members who later claim are paid from collected funds). Countries such as Finland, Sweden, Norway, Denmark, Iceland, Hungary, Latvia, and others have operated ECL schemes for decades in areas like broadcasting, photocopying, and library lending. The UK's Copyright, Designs and Patents Act sections 95A–95D created a statutory framework for authorizing ECL schemes, though uptake has been limited to date.

The appeal for AI is obvious: model training requires ingesting millions to trillions of copyrighted works, and negotiating individually with every author, publisher, photographer, and record label is logistically impossible. Wolters Kluwer's two-part analysis published in 2026, 'Collective Licensing for Gen AI Training: Feasible or Flawed?', frames exactly this question and concludes that the answer differs sharply depending on jurisdiction, sector, and how the extension mechanism is designed.

Why AI Training Breaks Traditional Licensing Models

Generative AI training presents problems that conventional collective licensing never had to solve. First, scale: frontier models are trained on datasets containing billions of documents, images, audio files, and code repositories. No existing CMO controls rights across all of those categories simultaneously, and many works have fragmented ownership — a single news article may involve a freelance writer, a photographer, an illustrator, and a publisher, each holding different rights.

Second, the opt-out problem. The EU AI Act's GPAI Code of Practice and the DSM Directive's text-and-data-mining provisions assume rightsholders can signal machine-readable reservations ('opt-outs') against scraping. Critics argue this shifts the burden onto individual creators, who must register preferences across countless crawlers. ECL flips this: instead of opting out, non-members are automatically included unless they take affirmative action to exclude themselves. That inversion is precisely what makes ECL attractive to AI companies and alarming to some creator groups, who fear it converts silence into consent.

Third, provenance and transparency. Distributing license revenue requires knowing which works were used. Most web-scale training corpora lack reliable attribution metadata, so even a perfectly negotiated ECL license cannot easily answer the question 'how much does this novelist get paid?' This is arguably the single largest technical obstacle, and it explains why several proposed schemes focus on sector-specific corpora (news archives, image libraries, music catalogs) where metadata quality is higher.

Fourth, the policy backdrop is contested. In the United States, the Trump administration's position — reported by Pinsent Masons — is that training AI on copyrighted content is not inherently infringing, which reduces immediate pressure for a licensing regime there. Meanwhile Anthropic's reported commitment tying a $21.6 billion Australian investment to copyright certainty shows how much commercial weight now rides on these regulatory decisions, with Canberra weighing reform options that reportedly include extended licensing models.

How an ECL Scheme for AI Would Work in Practice

A functioning ECL-for-AI system would follow a recognizable sequence. A government designates or authorizes a CMO as competent to negotiate for a defined class of works — say, published journalistic text or stock photography. That CMO negotiates terms with AI developers: permitted uses (training, fine-tuning, retrieval-augmented generation), scope (which corpora, which time periods), compensation formulas (per-work rates, per-token royalties, or lump sums), and audit obligations. Once agreed, the license extends automatically to non-member rightsholders in the same category within the jurisdiction.

Non-members would typically retain an opt-out window — often 30 days to six months after public notice of the scheme — during which they can exclude specific works or their entire catalog. Revenue collected would flow into a distribution pool. Because per-work attribution is imperfect, several proposals use sampling and statistical estimation: auditors sample the training corpus, identify the share attributable to each category or registered work, and distribute accordingly. Works registered with the CMO would earn directly; unregistered works might earn through category-level pro-rata pools.

Turkey's proposed amendments to Law No. 5846 on Intellectual and Artistic Works illustrate a national legislature attempting to modernize collective management for digital-era uses, and similar drafting exercises are underway elsewhere. The practical steps for a rightsholder today are: register works with relevant CMOs, embed machine-readable rights reservations in websites and feeds, document ownership chains for works with multiple contributors, and monitor consultation periods when regulators open comment on proposed schemes. For AI developers, the steps are: inventory training data sources, engage early with CMOs before litigation forces worse terms, build attribution logging into pipelines, and budget for retroactive licensing of already-trained models.

Comparing the Main Licensing Alternatives

ECL is one of at least five competing approaches, and understanding the trade-offs clarifies why no single option dominates.

Direct licensing already exists at the top of the market: TikTok signed licensing deals with Sony Music in November 2020 and Warner Music Group in December 2020, demonstrating that bespoke agreements are viable for high-value catalogs. But those deals cover thousands of tracks, not billions of web pages. Statutory compulsory licenses — where law itself sets the rate — avoid negotiation entirely but risk setting prices too low to sustain creation, and they remove market discovery entirely. Pure opt-out regimes under TDM exceptions preserve autonomy but, as noted, push enforcement costs onto individuals. Open licensing sidesteps everything but cannot supply the volume or diversity of contemporary material that competitive models require.

ECL sits in the middle: it offers coverage breadth approaching compulsory licensing while retaining negotiated pricing and limited creator exit rights. That middle position is also its weakness — it satisfies neither maximalist rightsholders nor free-market technologists fully.

Where ECL Works Well — and Where It Fails

ECL is most plausible where three conditions hold: a single dominant CMO already exists, works are registered and identifiable, and usage is measurable. Music is the strongest candidate; performance rights organizations already track plays with granular data. Newspaper archives are another: publishers hold clear rights, back-catalogs are digitized, and several national press agencies function as de facto collective negotiators. Stock photography follows a similar pattern.

It fails or strains badly in four areas. Web-scraped general corpora mix copyrighted works, public domain material, user-generated content, and personal data, making category definition nearly impossible. Visual art by living individual artists is highly fragmented, with weak collective representation in many countries. Code hosted on platforms raises questions about whether platform terms even permit collective licensing on behalf of uploaders. And cross-border training runs collide with territoriality: an ECL license granted in Finland says nothing about rights held by French authors whose works appear in the same corpus, since extension mechanisms generally do not cross borders except via reciprocal mandates.

There is also a governance critique. CMOs have historically faced criticism over administrative cost ratios, distribution accuracy, and board capture by large members. Extending their mandate to AI without reforming distribution mechanics risks channeling AI royalties disproportionately to major publishers while individual writers receive negligible amounts. Any credible scheme needs capped administrative deductions, transparent sampling methodologies, and independent audit rights.

Common Mistakes in Evaluating ECL Proposals

Several recurring errors distort the debate. The first is assuming ECL equals consent. Extension presumes representation; a rightsholder who never joined a CMO and never opted out may still be swept in. Treating automatic inclusion as voluntary participation misstates the legal mechanics and fuels legitimate objections from authors' alliances.

The second mistake is ignoring the opt-out asymmetry. In practice, opting out of an ECL scheme requires knowing the scheme exists, meeting a deadline, and having registered your claims. Freelancers, hobbyist photographers, and non-English-speaking creators are systematically less likely to manage this, meaning ECL tends to redistribute money toward organized, professionalized rightsholders.

The third is conflating training licenses with output licenses. An ECL agreement covering ingestion of works says nothing about generated outputs that may reproduce protected expression. Regulators who bundle them carelessly create schemes that underprice one use or overreach on the other.

The fourth is treating the US position as settled globally. The current US administration's view that training on copyrighted content is not infringing reduces domestic pressure, but US firms exporting models face the EU AI Act, UK considerations, Japanese and Singaporean exceptions, and Australian deliberations tied to the Anthropic investment announcement. A global product must navigate all of these simultaneously, so a US-only analysis misleads planning decisions.

Finally, stakeholders underestimate timing. Schemes take years: designation, negotiation, legal extension clauses, distribution systems, and dispute procedures. Waiting until litigation outcomes force terms means accepting whatever framework courts or legislators impose under pressure.

When to Act and What It Costs

For rightsholders, the actionable window is now. Regulatory consultations in the UK, EU member states, Australia, and Turkey are open or imminent through 2026–2027, and positions taken in consultations shape statutory language. Registering works with CMOs costs little or nothing beyond standard membership fees (typically modest annual dues plus a percentage deduction from distributions, historically ranging roughly 10–30% depending on the organization). Embedding rights-reservation signals in robots.txt, meta tags, and feed headers is essentially free. Legal advice on opting out of a specific scheme, once drafted, runs from a few hundred to a few thousand dollars per engagement — trivial compared to potential lifetime royalty exposure.

For AI developers, costs are harder to pin down because no mature AI-ECL market exists yet. Analogies help: music sync and streaming licenses run into hundreds of millions annually for major platforms, and news licensing deals between AI labs and publishers have been reported in ranges from single-digit millions to over $100 million for multi-year access. A broad ECL covering national news archives might plausibly price in the tens of millions per jurisdiction per year, though any figure today is speculative. Developers should also budget for compliance infrastructure — corpus auditing, attribution logging, and reporting — which industry estimates suggest can add low single-digit percentages to total training expenditure.

Timing matters asymmetrically. Rightsholders who organize early negotiate from strength; those who wait receive whatever default the eventual scheme assigns. Developers who license early buy litigation insurance and regulatory goodwill; DeepSeek's rise — described as 'upending AI' with open-weight models whose training data is not openly licensed — shows that unlicensed competitors can undercut licensed ones on price, which is precisely the dynamic ECL designers must counteract or the whole framework collapses commercially.

The Verdict: Feasible in Pockets, Flawed as a Universal Fix

Extended collective licensing for AI training is neither a panacea nor a dead end. It is feasible where collective infrastructure, metadata, and measurability already exist — music, news archives, commercial photography — and it offers the only realistic path to compensating the long tail of creators whom individual deals will never reach. It is flawed wherever works are unregistered, ownership is fragmented across borders, or output reproduction remains unresolved, and it carries real risks of entrenching incumbent CMOs and redistributing value away from the least organized creators.

The likely trajectory through 2027–2030 is patchwork adoption: sector-specific ECL pilots in Europe and possibly Australia, continued direct deals at the premium end of the market, statutory exceptions persisting for research uses, and ongoing uncertainty in the United States pending litigation outcomes and any shift in federal posture. Stakeholders should treat ECL as one instrument in a portfolio rather than THE answer — registering works, signaling rights machine-readably, participating in consultations, and structuring direct deals where leverage permits. At AI Trademark Review, our assessment is that the jurisdictions that pair ECL with strong distribution transparency and genuine opt-out usability will produce models worth copying; those that extend licenses without fixing attribution will simply generate new grievances at industrial scale.