WIPO’s New AI Copyright Study Examines Training, Live Answers and Data Reuse

AI developers can keep using a training collection for years after their first model is finished, a new study from the World Intellectual Property Organization explains. Published on October 23, it examines how retaining files, updating models and supplying documents for live answers raise separate questions about permission and payment. WIPO study

WIPO commissioned the study for its copyright committee. The authors compare approaches without recommending one, and their views are not agreed policy. Delegates meet in Geneva from November 30 to December 4. WIPO study Committee dates

Published terms and technical rules make the distinctions concrete. Some address a company’s own staff, some cover models offered to customers, and others separately describe training and using documents to answer a question. Comparing those documents reveals why the phrase “licensed for AI” leaves too much unanswered when a product changes. License portfolio RSL specification

Training a model and answering a customer require different descriptions

Training changes a model using examples. Retrieval supplies documents when the model prepares an answer, often so it can use current information or quote a source. Really Simple Licensing, a machine-readable licensing format, separates training, AI inputs and retrieval indexes. Its search category excludes AI-generated summaries. RSL specification

Copyright Clearance Center, which licenses published material, described separate corporate and higher-education licenses for internal AI uses in March. Its portfolio also included an AI Systems Training License for external systems and planned pay-per-use rights beginning with summarization. Coverage needs checking against the relevant terms. License portfolio

The training product’s own page adds another condition: works must be lawfully acquired. It offers rights to use them as training inputs and directs prospective customers to request pricing. It does not present a universal price or say that buying this license supplies every document a developer needs. Access to material and permission to train on it therefore need separate confirmation in a purchase. Training license

For a buyer, this comparison changes which project description belongs in a negotiation. A staff research assistant that reads subscribed journals and a paid service answering customers from the same journals have different audiences and delivery arrangements. Those are proposed uses to specify, not evidence that either product is automatically covered by a particular license. License portfolio

A quote also needs to identify what triggers a charge: receiving files, starting a training run, serving an answer or renewing access. A buyer can request separate descriptions of permission fees and any content delivery service, then compare offers for the same intended use. CCC’s public page requests a quote; it does not establish a standard fee or the terms of an individual deal. Training license

Letting a crawler enter does not describe every later use

Cloudflare’s Content Signals Policy separates ordinary search, AI training and supplying content to AI answers. Its example permits search while rejecting training. A blank AI-input signal expresses no preference through that particular field. Website owners evaluating discovery and licensing can consequently make different choices about these activities rather than treating every bot visit as the same request. Content Signals Policy

Really Simple Licensing goes further by providing fields for payments and reporting alongside permissions and restrictions. Its pay-per-crawl example pairs training permission with a payment framework. The amount charged for an access request alone cannot tell a buyer which later uses the associated terms permit. RSL specification

Technical enforcement still has a separate job. The Internet Engineering Task Force’s Robots Exclusion Protocol says crawler instructions do not provide access authorization and cannot replace security controls. A private document library therefore needs a way to restrict entry in addition to whatever public instructions describe acceptable use. A publisher also needs records of what was delivered if its bill depends on measured access. Robots Exclusion Protocol

The protocol also specifies what crawlers should do when instructions cannot be fetched. A missing robots file can leave crawling permissible. Without usable cached instructions, a server or network failure requires an initial assumption of no access. Publishers need to test the response actually returned to bots, alongside the words in the file. Otherwise a publishing error can change how conforming crawlers proceed. Robots Exclusion Protocol

Permission depends on what the supplier can grant

WIPO’s authors identify another difficulty: publishers and repositories may hold older grants of rights whose coverage of AI training is uncertain. A developer’s agreement with an intermediary must be considered alongside the rights that intermediary received from contributors. WIPO study

Creative Commons Attribution 4.0 illustrates the boundary in an actual public license. It grants only rights the licensor has authority to provide. It excludes patent and trademark rights, treats privacy and similar personal rights separately, and generally disclaims warranties of ownership and non-infringement. Those provisions affect what protection a recipient receives even though the license allows broad reuse. CC BY terms

Creative Commons also explains that its licenses work with new technologies when copyright permission is required and their conditions are met. An old publication date does not by itself make a license unusable for AI. Conversely, where an exception permits a use, the copyright license does not control that use. The applicable law and exact license both matter. Creative Commons guidance

Taken together, these documents suggest a more useful comparison than a single yes-or-no claim about AI permission. A buyer can identify the files being supplied, the intended product and audience, which rights the supplier controls, and any rights or uses left outside the agreement. If the supplier offers contractual protection against claims, its wording and financial limits need their own review. CC BY terms

Countries are still deciding how these permissions should work

European Union law already distinguishes two text-and-data-mining exceptions. Its 2019 copyright directive provides one for qualifying research organizations and cultural heritage institutions doing scientific research and another subject to appropriately expressed rights reservations. Lawful access matters. Applying those rules to a particular AI system requires examining its activity and relevant national law; merely calling an operation research supplies an incomplete description. EU copyright directive

Singapore is asking related questions about lawful access, safeguards for rights owners and responsibility for infringing outputs. Its current consultation page extends the response deadline to November 5, replacing October 22. That process remains a consultation. The extension gives participants more time to address specific proposals before officials decide their response. Singapore consultation

For companies preparing a transaction now, a useful next step is to describe what continues after delivery: future training runs, access to updated documents, retention of files and any distribution to another organization. Parties can then discuss the permissions and payment for each intended activity against actual terms. The next scheduled international discussion is WIPO’s Geneva committee session starting November 30. Committee dates