Two competing answers exist right now for a question almost nobody has actually settled. Donna Medeiros, VP of AI and Data Advisory at Data Society, watches this fight play out at both the global and national level, and it comes down to who gets final say over what AI owes the public.
“There’s two schools of thought going on out there right now with how AI technology advances, and also the use of things like indelible marking and what is copyright and so on,” Medeiros says. “The two schools of thought are that government and public sector AI should be public infrastructure, and that it’s up to the end users and governments to regulate those types of matters, on how it is public infrastructure and of use. And then there’s the frontier technology companies, such as Anthropic, OpenAI, Google, Microsoft, that are currently advancing the technology and considered the frontier leaders. But should they be deciding what is in the public good? What becomes copyright? What is single-use, in other words not what is public infrastructure?”
Governments Are Already Acting Like AI Is Infrastructure
The first camp isn’t hypothetical. It’s already showing up in policy. Harvard’s Ash Center points to a growing “public AI” movement that treats cloud infrastructure, data, and model development the way earlier generations treated the interstate highway system or the electrical grid: as layers the public sector builds so the public benefits, even when private companies help construct them, according to the Allen Lab for Democracy Renovation’s policy primer. The comparison isn’t accidental. DARPA built the backbone of what became the internet. NASA’s research fed decades of private aerospace innovation. The argument is that AI deserves the same public foundation instead of starting out as someone else’s product.
That shift is already underway in practice, not just in theory. India’s IndiaAI Mission funds public compute access for domestic AI development, and the government has discussed taking a minority stake in Sarvam AI, one of the country’s leading AI firms. The UK now treats data centers as critical national infrastructure. The EU is mobilizing capital for what it calls AI factories. In the US, OpenAI has reportedly discussed giving the federal government a 5 percent equity stake, according to reporting covered by the Centre for International Governance Innovation. None of this is full nationalization. It’s governments deciding they can’t just be regulators standing outside infrastructure someone else owns outright.
The Copyright Question Nobody Can Answer Unilaterally
Medeiros’s second question, what becomes copyright, already has a partial answer, and it didn’t come from any frontier lab. The US Copyright Office’s January 2025 report concluded that purely AI-generated material gets no copyright protection at all. Human authorship still has to show up somewhere. Prompts alone don’t count. The office wrote that prompts “do not provide sufficient human control” over a generated work under current technology, meaning a person typing a detailed request into a model isn’t automatically the legal author of what comes out.
That’s a meaningful check on the idea that whoever builds the model gets to define what its output is worth or who owns it. But it’s a narrow one. It answers who owns an individual output. It says nothing about whether a company’s frontier model should have been trained on copyrighted material in the first place, and that fight is still being litigated case by case rather than settled by policy.
Indelible Marking Is Being Written Right Now, and Not Only by Regulators
Medeiros’s mention of indelible marking points to a deadline that already passed this year. Article 50 of the EU AI Act required generative AI providers to mark their outputs in a machine-readable, detectable format starting August 2, 2026, with systems already on the market getting until December 2, 2026 to comply, according to the EU AI Act’s official transparency guidance. But the actual technical standard, whether that ends up being watermarking, embedded metadata, or a provenance system like C2PA, is still being finalized through the Code of Practice on AI-generated content, expected to be complete by June 2026.
A Code of Practice gets built with heavy input from the companies it will regulate. The same frontier labs Medeiros names as advancing the technology are also sitting at the table where the rules for labeling that technology’s output get written. That’s not necessarily corruption. It’s often the only way regulators get the technical expertise to write a workable standard. But it does mean the line between a government setting the rule and an industry writing the rule a government signs is blurrier than the two-camps framing suggests.
Open Future, a European digital rights organization, argues that concentrations of power in AI let a narrow group of companies set terms for a broad range of public activity, and it wants foundation models regulated as core platform services with mandatory third-party data access instead.
Governments are buying equity stakes and building public compute. The same frontier labs are sitting in the room where the EU’s AI marking standard gets drafted. Nobody has declared a winner between Medeiros’s two schools of thought, and both sides are moving before anyone does.
Frequently Asked Questions
Donna Medeiros describes two competing schools of thought, one holding that government and public sector AI should be public infrastructure regulated by governments and end users, the other leaving those calls to the frontier companies advancing the technology. Nobody has declared a winner, and both sides are moving before anyone does.
Yes, and it is already showing up in policy rather than theory. India’s IndiaAI Mission funds public compute access and the government has discussed taking a minority stake in Sarvam AI, the UK now treats data centers as critical national infrastructure, and the EU is mobilizing capital for what it calls AI factories.
The US Copyright Office concluded in its January 2025 report that purely AI-generated material gets no copyright protection, because human authorship still has to show up somewhere. Prompts alone do not provide sufficient human control over a generated work, so a person typing a detailed request into a model is not automatically the legal author of what comes out.
Article 50 of the EU AI Act requires generative AI providers to mark their outputs in a machine-readable, detectable format starting August 2, 2026, with systems already on the market getting until December 2, 2026 to comply. The technical standard, whether that ends up being watermarking, embedded metadata, or a provenance system like C2PA, is being settled through a Code of Practice on AI-generated content.
Not cleanly, because the same frontier labs advancing the technology are also sitting at the table where the rules for labeling that technology’s output get written. That is often the only way regulators get the technical expertise to produce a workable standard, but it blurs the line between a government setting the rule and an industry writing the rule a government signs.
