On 30 October 2026, the Electronic Frontier Foundation's Corynne McSherry and Tori Noble published "Rewriting Copyright Law for Generative AI Would Hurt Everyone", tied to an EFF amicus brief in In re OpenAI. It argues against narrowing fair use for AI training and against a licence-for-every-work regime. The dispute is American, but Switzerland has been working through the same question in its own parliament. Its answer is more instructive than the headlines suggest.
The strongest case for the other side
Rightsholders have a serious argument. The text of Motion 24.4596 says Swiss media content is used by international AI services to develop language models without permission, and that retrieval-based systems pull individual articles, rephrase them and serve them as answers, bypassing paywalls. If a model's value is built on journalism and the model then displaces the outlet, a market failure is plausible. The National Council accepted that concern in September 2025, warning that human-made content may stop being worth producing.
That is a real harm and deserves a real policy response. The question is whether mandatory prior consent is the right one.
What the Swiss motion actually says now
Petra Gössi filed Motion 24.4596 ("Besserer Schutz des geistigen Eigentums vor KI-Missbrauch") on 20 December 2024. As submitted, it required the Federal Council to amend the Copyright Act (URG) so that rightsholder consent is necessary whenever journalistic content or other original works are read, processed and offered by generative AI. It also asked that the exceptions in Art. 19(3) and Arts. 24a, 24d and 28 URG be clarified so AI services cannot invoke them, and that Swiss law and courts apply to services offered in Switzerland. The Council of States adopted it on 20 March 2025 without a counterproposal, and the Federal Council recommended adoption, according to the Council of States committee report.
The National Council then rewrote it. On 16 September 2025 it adopted an amended text by 121 to 66 votes with 8 abstentions. The three concrete demands were struck. What remains asks the Federal Council to create the conditions for comprehensive protection of journalistic and other copyright-covered works when used by AI providers, while ensuring that Switzerland's position in AI research, development and commercialisation is not weakened or disadvantaged internationally.
The reasoning is the interesting part. According to the National Council committee report, the committee heard business, research, media and culture representatives and was told that copyright protects form, not information, and that AI systems are mainly interested in the information. It said an approach built on automatic participation with an opt-out looks most promising, with reservations about the role of collecting societies, and that the original wording left the Federal Council too little room. The committee backed the amended text 18 to 6 with 1 abstention. The National Council's own summary, in the later report, says the original formulation would create perverse incentives and in practice amount to a ban, with consequences for research, innovation and international connectivity. The Council of States committee then recommended the amended version unanimously, 13 to 0, noting that the Federal Council can look to solutions abroad to avoid a "Swiss finish."
One correction to the framing in circulation: opt-in consent is no longer the motion's text. I could not confirm from primary sources whether the Council of States has completed its final vote on the amended wording, so check the parliamentary record for business 24.4596. A Schellenberg Wittmer summary says that once the motion passes, the Federal Council drafts a detailed proposal and the Swiss Federal Institute of Intellectual Property is already doing preparatory work, though it gives no dates.
Why the comparison matters
The EFF makes three points. First, courts have already found fair use in some training cases because training serves a different purpose. Second, a permission or lawful-acquisition requirement would cast doubt on the ordinary copying that browsers, email and streaming perform, and would hit watchdog researchers whose scraping methods resemble those of AI developers. Third, as the post quotes Professor Lawrence Lessig, if all training requires a licence, only "a handful of tech giants" could afford it.
Switzerland has no fair-use doctrine; it works through specific exceptions, such as Art. 24d URG. So the instruments differ, but the economics probably do not. This is our own analysis, not something the EFF post or the Swiss reports state: a consent-first rule in a small market would raise costs mostly for domestic labs, start-ups and universities, while the largest foreign providers can sign deals or simply serve Switzerland regardless. The National Council's insistence on alignment with other states and the EU reflects that dynamic.
A proportionate path
The opt-out route the committees favour is not a free pass. Done well, it gives publishers a machine-readable way to say no, preserves research access, and leaves room for collective licensing where rightsholders want payment. Its weakness is enforcement: an opt-out is only as good as crawler compliance and transparency about what was ingested. That, not a blanket licensing mandate, is where a Swiss bill should concentrate: standardised signals, disclosure of training sources and a clear route to litigate in Switzerland.
The EFF is right that rewriting copyright for generative AI affects far more than the parties in a courtroom. The Swiss parliament reached a compatible conclusion by a different route: after hearings, it moved from consent-for-everything toward opt-out, flexible remedies and international alignment. A consultation draft from the Federal Council, if and when it arrives, should be judged on whether it keeps to that restraint.