Policy

Simon Willison rounded up the recent wave of open letters about AI development, which have been piling up from researchers, companies, and advocates. When this many people feel the need to write formal letters, it usually means the informal conversations are not going well.

Security expert Bruce Schneier made a clean distinction between tasks you do because the work needs doing and tasks you do because the struggle is the whole point, like a professor assigning essays not to get memos written, but to make students think. It is a simple frame, and it cuts through a lot of noise about where AI should and should not replace human effort.

The EU AI Act is now in force, and this survey maps out what companies actually need to do to comply, translated into the kind of concrete requirements engineers can test. Wherever you work, if your organization uses AI in any customer-facing way, the compliance question is no longer hypothetical.

Nvidia, Microsoft, and Meta sent a joint letter to Washington arguing that open-weight AI models, the kind anyone can download and run, should not be heavily regulated. How this debate lands will determine whether the AI tools your company uses five years from now are built by three big players or by a much wider field.

A group of startup founders is publicly asking the U.S. government to keep Chinese open-weight AI models accessible, arguing that blocking them would hurt American developers who rely on those models for real products. If you build anything with open-source AI tools, the outcome of this debate will directly shape which models you are allowed to use.

Codeberg, a popular open-source code hosting platform, is updating its terms of service to push back against AI companies bulk-scraping the work that volunteer developers share there for free. This is the slow, unglamorous policy fight that will actually determine who controls the training data that future models are built on.

A federal judge approved a $1.5 billion settlement between Anthropic and authors whose books were used without permission to train Claude. This is the largest copyright settlement in AI so far, and it signals that the era of training on whatever was lying around the internet is getting expensive in a very concrete, legal way.

China's open-weight models, meaning models anyone can download and run, are becoming the default building blocks for developers worldwide. If the tools your company uses get built on top of them, then the choice of which AI to trust was already made for you, upstream.

Power companies in some states can use eminent domain to run transmission lines to data centers, which means your neighbor's field, or yours, could legally become infrastructure for someone's AI workload. The energy demand is real, and it is landing somewhere specific.

New York City is moving toward requiring landlords and realtors to disclose when AI-generated images are used in property listings. If this passes, the next time you are apartment hunting and a living room looks suspiciously perfect, you will have a legal right to know it was made up.

The State of Open Source AI report landed on Hacker News and pulled 449 upvotes and over 300 comments, which means a lot of technically informed people found it worth their time. Open models are increasingly the thing businesses use when they cannot send their data to someone else's server, so this snapshot of where that ecosystem stands is genuinely useful context.

Linus Torvalds, the person who created the Linux kernel, publicly put his foot down in favor of allowing AI-assisted contributions to Linux. When the person who sets the tone for one of the most important open-source projects in the world shifts his stance, the ripple effects on how software gets built are real.

A well-regarded policy brief from the Siegel Endowment made the case that governments, companies, and nonprofits should be putting real money into free, open-source AI, and it picked up serious traction online. If you have ever worried about a handful of private companies controlling the AI tools everyone depends on, this is the argument that something can actually be done about it.

Hacker News users are asking for a simple label on AI-generated articles, the same way nutrition labels tell you what is in your food. It sounds minor, but it gets at a real question: do you want to know what you are reading, and who decides?

A new non-profit called Current AI launched an open map showing where open-source AI is strong and where it still has big gaps compared to closed commercial models. If you or your organization care about not being locked into one vendor, this is a useful honest picture of how far the open alternatives actually go today.

The U.S. Department of Commerce lifted export controls on two Anthropic models, Claude Fable 5 and Mythos 5, meaning those models can now be accessed in countries that were previously blocked. This is a small bureaucratic update with a concrete effect: teams in restricted regions can now build with those tools legally.

The Godot game engine, which is free and built by volunteers, announced it will no longer accept code contributions that were written by AI, because maintainers say they can't trust that the contributor actually understands what they submitted. If you work on any open-source project, this is a real policy conversation coming your way soon.

A new study looks at how calling AI systems things like 'thinks,' 'understands,' or 'feels' shapes what ordinary people expect from them, and not always in accurate directions. The words journalists and marketers choose about AI are quietly setting expectations that the technology then has to live up to, or disappoint.

A Brown University professor flagged what appears to be widespread AI-written submissions on a single exam, enough to make it a public story. If you work in education, or manage people whose work you review, this is the moment where 'assume good faith' gets complicated.

When one company's models get blocked by export rules, competitors fill that shelf space fast. If you are building a product that depends on a specific AI provider, this is a good week to check how locked in you actually are.

Someone compared learning to prompt AI well to learning to be a manager, pointing out that just because your team does what you ask does not mean you automatically know how to lead. The analogy is sharper than it sounds: getting good results from AI, like getting good results from people, is a skill you build, not a switch you flip.

Anthropic is accusing Alibaba of systematically extracting Claude's capabilities, essentially probing it with carefully designed prompts to replicate what it knows. If you have ever wondered whether 'it is just a chatbot' means nobody would bother stealing it, this lawsuit answers that.

A new paper asks whether LLMs can genuinely help with democratic deliberation, public consultation, policy feedback, that kind of thing, or whether they mostly flatten nuance and favor whoever writes the best prompt. If your city or employer starts using AI to gather public input, this question matters more than it sounds.

The Netherlands built its own large language model, trained on Dutch text and governed by Dutch institutions, instead of renting that capability from a US company. If your government or employer handles sensitive data, this is the model for what 'keeping AI local' actually looks like in practice.

Researchers found signs of undisclosed AI-generated text in parliamentary documents from the UK and Sweden, places where the words carry real legislative weight. The question of where AI writing ends up, without a label, is going to keep getting louder.

Anthropic sent staff to Washington to repair a relationship with the White House after a reported falling-out, suggesting that even the companies most focused on safety research spend real energy navigating political proximity. How close AI labs sit to government shapes which rules eventually land on all of us.

A city government announced its own local AI model as a point of civic pride, and the community quickly found it was largely a repackaged version of an existing open model. If a public institution is making decisions about your services based on a tool it claims to own but does not understand, that is worth knowing.

When the CEO of one of the biggest investors in an AI company talks to government officials, the model you use at work can quietly change overnight, with no announcement and no explanation. That is not a tech story, it is a supply-chain story.

A post titled 'Open source AI must win' picked up nearly a thousand upvotes on Hacker News, which is a lot of agreement for a tech crowd that argues about everything. The Fable 5 situation gives that argument a sharper edge: if the model is open and already downloaded, no government directive can flip the off switch.

The US government, citing national security, ordered Anthropic to cut off access to its Fable 5 and Mythos 5 models. If you use any product built on those models, it stopped working, not because of a bug or a policy update, but because of a government directive you never saw coming. This is the first time most people are seeing just how quickly access to an AI tool can vanish.

A new benchmark called Polar tests AI models for political bias across multiple countries and languages, not just American English political categories. As these models get used in news tools and civic applications around the world, knowing whether they lean one way in one country and another way elsewhere is a genuinely important question that nobody has had a clean way to measure until now.

A German court ruled that when Google's AI Overview gives you a wrong answer, that is Google's own statement and Google is liable for it. This is the first major ruling to treat an AI-generated summary like a published claim, which means every company surfacing AI answers to users just got a new legal risk to price in.

The Ladybird browser project announced it will no longer accept public pull requests, partly because a large commit used to signal genuine effort, and now it does not. This is a small policy change in one open-source project, but it points at a real question that every collaborative software team will face: how do you maintain trust in contributions when effort is no longer a reliable signal of good faith.

Someone asked Hacker News why the tech community there seems so skeptical of AI, and the thread got 175 upvotes and a genuinely interesting conversation. The short answer seems to be that experienced engineers are the ones closest to the real failure modes, and proximity tends to replace enthusiasm with caution.

Charity Majors wrote a piece arguing that AI optimists and AI skeptics are essentially running different clocks, one betting urgency wins, the other betting friction wins. Worth reading on a Saturday morning not because either side is right, but because understanding the shape of that disagreement helps you make better decisions about where to place your own bets at work.

Meta is allowing employees to opt out of workplace tracking for up to 30 minutes at a time. The fact that opting out requires an active choice, and only lasts half an hour, tells you quite a bit about how these systems are designed by default.

Senator Bernie Sanders published an op-ed arguing the public should own half of the major AI companies, since public research dollars helped build the technology. You do not have to agree with the policy to recognize the underlying question, who benefits from AI and who paid to build it, is one that is going to get louder, not quieter.