Europe Should Force AI Firms to Pay for Training Data

A recent study showed that online traffic to news sites will drop by a third in 2025. This problem should be of concern to everyone, not just journalists and media executives, because democratic societies cannot function without quality information. In an era of polarization, fragmentation and democratic rollback, news outlets providing quality journalism are needed more than ever.

But to produce it, publishers must be remunerated – a perennial problem in the age of the Internet. For years, search engines like Google and social networks like Facebook have made huge profits from news content while paying little to the media.

The same was true for artificial intelligence companies, which copied the content needed to train their large language models without compensating its producers or getting their consent. OpenAI has since entered into agreements with several major publishers (including News Corp, Axel Springer, and Le Monde) to ensure that ChatGPT has access to the most up-to-date information. But many others have stayed away.

In a new working paper, we argue that AI companies should automatically pay for the content they use. The current approach, which allows virtually unlimited use of news articles without compensation, leads to the disappearance of original content. However, prohibiting generative artificial intelligence from using all creative works is virtually impossible and not in anyone’s interest. The most sustainable policy is to require payment to publishers and creators – known as “statutory licensing” – for the use of their works.

How to address the “power imbalance”

Governments should step in to help broker these agreements because the current process of demanding compensation, in which groups of creators or individual publishers must sue AI companies, is slow, costly and unfair.

When relatively powerless authors and news publishers face powerful technology companies, the playing field is hardly a level playing field and settlement amounts are often small.

In September 2025, Anthropic agreed to pay $1.5 billion to settle a copyright infringement class action lawsuit, or about $3,000 per book – a shockingly low amount when you consider that authors can spend decades researching and writing a single book.

Some countries have recognized the power imbalance between large tech companies and content creators, and the negative impact this has on the perceived value of creative output. In 2021, Australia adopted the News Media Negotiation Code, which resulted in search engine and social media giants paying media outlets for content posted on their platforms.

A similar approach should be applied to artificial intelligence companies. In the US, by contrast, copyright law contains a broad definition of “fair use”, which artificial intelligence companies invoke to defend their content scraping.

Europe has shown a willingness to address this issue, with policymakers debating whether and how to update the European Union’s copyright directive, which currently includes an exception for “text and data mining.” In late January, the European Parliament’s Legal Affairs Committee adopted draft proposals, which the Parliament is expected to vote on in March, to ensure fair compensation for copyright owners.

The proposals were formulated in a report commissioned by the Legal Affairs Committee and published in June 2025, which highlights the ambiguities and shortcomings of applying the current copyright framework to AI training.

A new system of relationships is needed

The report calls for a new system that imposes remuneration obligations on suppliers of general-purpose AI models and creates a licensing market that restores the bargaining power of rights holders.

First steps could include facilitating collective licensing agreements and enforcing the duty to remunerate even before broader reviews or reforms are undertaken.

The report also recommends introducing transparency obligations for AI companies, a special regime for news media, and a central registry for organizations that want to opt out of data collection.

Many hope the report will also set the agenda for the European Commission’s upcoming 2019 review of the Copyright Directive, which could lead to new mandatory legislation later this year or in 2027.

Of course, questions remain about whether all rights holders in Europe will be automatically included in the compensation system and what the opt-out mechanism will look like. We support some mandatory payments to publishers and content creators, while recognizing that simplified rules that do not allow for as strict a distinction on quality as we would like may have to be applied. This could look like a fixed, predetermined scale of fees, similar to those used for pharmaceutical licensing and music royalties.

It is important that laws ensure fair collective bargaining, which should include media of all types and sizes. The current system, in which only the largest news organizations can enter into agreements with AI companies, is not conducive to an open information ecosystem and media pluralism.

Preserving journalism in the AI era will likely require a system that includes prior authorization as a standard for educational use and a fixed pay scale. However, pressure from big tech companies to derail any legislation requiring remuneration for training materials is mounting. Europe must act now before these companies become too powerful to be regulated.

Anya Shiffrin

Senior Lecturer and Co-Director of the Technology Policy and Innovation Program at Columbia University’s School of International and Public Affairs.