The AI Race Has Reached the Home Computer

A new Mac Studio can be configured with 128 gigabytes of unified memory, enough to run substantial language models without sending each prompt to a cloud provider. At the same time, Anthropic chief executive Dario Amodei says every sufficiently capable model, open or closed, should face mandatory safety testing before release. The collision between those developments has moved AI regulation beyond a contest between laboratories and governments: it now concerns what people may keep and run on their own machines.

None of the laws and proposals examined here establishes a general ban on private ownership of language models or makes downloading an open-weight model a crime. But the distinction between regulating a company and regulating access can narrow quickly when policy reaches model distribution, computer chips, operating systems and the images a device will permit its owner to view or create. The test is whether safety rules can target demonstrable risks without turning private computing into a licensed privilege. Will laws against private individuals training their own models outside of a future regulated AI framework become an issue, as we had with people using 3D printers to print guns?

The machine in the spare room changes the debate

A local model is software stored and run on a computer the user controls. It can answer questions, draft text, summarise documents, write code or interpret images without every exchange passing through a company’s server. Its weights—the numerical parameters learned during training—can be downloaded and run through compatible software. They can also be adjusted, copied or deleted. That is a different relationship from renting access to a hosted chatbot whose owner can change the rules, restrict an account or discontinue the service.

The distinction is no longer theoretical. Apple’s 2026 Mac Studio offers an M5 Max configuration with up to 128GB of unified memory and an M5 Ultra configuration with up to 512GB. Apple says the systems can run large language models on-device; that is a company claim about a new product, not proof that a desktop matches the best closed models available through a data centre. But it does establish the direction of travel: more memory and compute are arriving in machines sold to individuals, not only in buildings filled with accelerators.¹

Memory is not intelligence. A model’s parameter count, training, architecture, quantisation, software and the speed of its memory system all matter. A person running a local model may accept slower answers, reduced reasoning performance or less current information in exchange for privacy, customisation and independence from a provider. Those trade-offs still have value. A useful tool need not be the world’s best tool to shift bargaining power.

That shift unsettles a regulatory debate framed around a handful of frontier laboratories. A provider can monitor requests sent to its own service, apply access controls and withdraw a product. Once weights are released, the original developer cannot reliably retrieve every copy or restore every safeguard. The same durability gives users freedom from a company’s discretion and makes misuse harder to stop. The conflict is real; neither side gets to erase the other half.

The ownership question is therefore more exact than “Should AI be regulated?” Software has long been subject to rules about copyright, product safety, privacy and illegal conduct. The question is whether lawmakers will regulate an identifiable act—training, commercial distribution, deployment in a consequential setting or a harmful use—or will make ordinary possession of a model itself suspect because the software might be used badly.

The law mostly starts with providers, not private users

The European Union’s AI Act does not establish a licensing scheme for people who install a model on a home computer. Its rules distinguish between a general-purpose model, an AI system built around that model, the provider that places a model on the market, and the deployer that uses a system. The European Commission says model providers have documentation and copyright-policy duties; providers of models classed as posing systemic risk must also assess and mitigate risks, evaluate models, report serious incidents and provide cybersecurity protections.²

That is a broad framework, and it does regulate models at the provider level rather than waiting for each harmful use to reach a court. But it is not the same as regulating every citizen who downloads a file. The Act also exempts certain open-source releases from some documentation obligations when the weights, architecture, and usage information are made public. The exemption does not cover models with systemic risk. The legal design is therefore neither “open weights are free of all rules” nor “anyone who runs a model needs a licence”.

California has taken a similar, though distinct, route for frontier developers. The state’s SB 53, signed in 2025, requires large frontier AI developers to publish safety frameworks, report potential critical incidents and protect employees who disclose significant risks. It creates an enforcement mechanism for non-compliance. The stated subjects are frontier developers and the systems they build—not every local user who runs a model at home.⁴

Why would a frontier AI developer remain in California?

These examples matter because they puncture two inflated descriptions of the debate. The first says safety regulation is already a direct ban on personal AI. The legislation cited here does not support that. The second says that policy proposals concern only conduct after a model has been deployed. That is also wrong: frontier testing, developer documentation and controls on release shape what models people can access before any individual uses them.

The practical burden depends on the actor and the point in the supply chain. A person running a model privately is not the same as a company training one, an online service distributing it to millions, an employer using it to decide who gets hired, or a platform enabling illegal images to circulate. Treating those activities as interchangeable is a political shortcut. It helps either side exaggerate: regulators can imply that all AI is a threat, while industry can describe almost any safety rule as an assault on personal freedom.

“Sufficiently capable” is a threshold, not a neutral answer

The sharpest current proposal is not a blanket ban on open weights. In July 2026, Amodei wrote that “all sufficiently capable models, open and closed,” should undergo mandatory safety testing before release. He described non-dangerous open-weight models as a “public good” and argued for restrictions on high-end chip exports to China, action against industrial-scale model distillation and evaluations for cyber, biological and alignment risks.³

That is more limited than the claim that Anthropic wants the government to prohibit people from running local models. It is also more consequential than a promise to test only Anthropic’s own products. If law adopts mandatory testing for models above a capability threshold, a developer’s release decision becomes a regulated activity. Whether an independent developer can afford the tests, who chooses the threshold and what happens when a model is modified after release would determine whether the regime controls only the most dangerous systems or quietly closes off a much larger part of the field.

The phrase “sufficiently capable” sounds like a technical measurement waiting to be supplied. It is partly a political choice. A test could use training compute, demonstrated capability, model size, likely uses or some combination. Each measure captures different things and invites different evasions. Training compute can be estimated, but the same level of capability may be reached with less compute as methods improve. A capability test may better match the concern, but the result can depend on benchmark design, access to the model, and evaluators’ judgement.

The EU model rules show that this is not a remote hypothetical. The Commission’s explanatory guidance identifies 10^25 floating-point operations in training as a threshold for presuming systemic risk, while also saying the state of the art and equivalent impact matter. It notes that the threshold can be revised as the technology changes. That is a rule about developing and supplying the model, not a test that a home user must pass before pressing “run”. Yet a threshold that moves with the frontier can still change which models may be openly released and who can afford compliance.

Amodei’s case deserves to be taken seriously on its own terms. Open weights can make safeguards easier to remove, and users of a local copy cannot be centrally monitored or cut off. He argues that direct testing is a better remedy than prohibiting open models as a category. The counterargument is not that misuse is imaginary; it is that mandatory testing can become an entry barrier favouring companies large enough to absorb the cost. A safety regime drafted by the largest firms can protect the public and protect incumbents at the same time.

That is why thresholds need published evidence, independent review and a right to challenge classification. Otherwise “frontier” becomes a moving legal category whose practical meaning is determined by the companies that lobby for it and the agencies that enforce it. If a laptop runs yesterday’s frontier, the relevant question should be what risk the model demonstrably creates, not whether its owner has accidentally crossed a regulatory line no ordinary user can see.

Open weights create freedom and a real control problem

Open-weight models give users and developers a form of control that hosted systems do not. A researcher can inspect and adapt a model; an organisation can keep sensitive files inside its own network; a developer can build a tool without waiting for a platform’s permission. Local operation can also reduce exposure to outages, price changes and the provider’s content policies. Those are not abstract benefits for software purists. They are ordinary interests in privacy, reliability and choice.

The security case against unrestricted release is equally concrete. A provider that hosts a model can rate-limit accounts, monitor suspicious activity and revise a safeguard. A downloadable model can be copied, run offline and modified. If it has dangerous capabilities, the original developer may have no practical way to learn who is using it or to disable a particular copy. That does not mean every open model is dangerous. It means capability, access and reversibility belong in the assessment.

Amodei’s July statement reflects this balance. He rejected a ban on open weights as a category, but endorsed mandatory pre-release evaluations above a capability line. The position leaves open a key public-interest question: who controls the evaluation, and how do we know it is testing danger rather than enforcing one institution’s political preferences? A model should not pass because its maker has a polished safety page; it should pass because external evaluators can test defined claims and publish enough of their method for others to scrutinise the result.

There is a further danger in treating the labels “open source” and “open weight” as interchangeable. A model’s weights may be accessible while its training data, full software stack or development process remain closed. Conversely, the availability of weights does not tell us whether a model’s licence permits every use. A rule should identify what is actually open, what the user can modify and what obligations follow from distribution. Loose terminology makes it easier to write exemptions that benefit a few firms while denying meaningful access to everyone else.

The strategic argument adds another layer. Governments may restrict advanced chips or exports because they see computing capacity as national infrastructure. That does not make the policy a ban on home AI, but it means consumers’ future choices can be shaped by industrial policy and geopolitical competition. A government can leave the model file legal while narrowing the hardware supply, making the most capable systems expensive or difficult to obtain. The user’s suspicion that ownership can be restricted indirectly is not fanciful. It is a reason to debate those levers openly, not evidence that a prohibition has already been enacted.

The public interest is served by preserving a wide space for ordinary, lawful experimentation while requiring stronger controls where the evidence supports them. The harder the intervention is to reverse—blocking distribution, constraining hardware, or imposing costly testing—the stronger the case should be. That standard protects the public from reckless release and prevents “safety” from becoming a synonym for keeping the most useful systems inside a small number of corporate clouds.

Britain’s device proposal reaches a different layer

Britain’s June 2026 announcement about children’s phones is the clearest example of policy reaching from online services towards the device itself. The government told Apple and Google to activate existing features or implement technical measures to detect and block nude images on smartphones and tablets used by children. It said adults would retain access after age verification, and that it would legislate if technology firms did not act within three months.⁵

The announcement is not a law that bans adults from running a local language model. Nor does it say a device must send every image to a government server. The Home Office said protections should work without collecting data, and that the device should block content across apps and services. But any system that classifies images on a phone before they can be viewed, captured or shared is a form of device-level moderation. The technical implementation matters: the proposal leaves unanswered questions about error rates, age checks, accessibility, appeals, and who can change the classifier.

The government’s argument cannot be dismissed as a pretext invented by technology companies. It points to coercion, grooming and the circulation of sexual images involving children. Victims bear harms that are immediate and personal; the burden of preventing them should not fall solely on young people or parents. Platforms and operating-system makers have the resources to build safeguards. The difficulty lies in turning a legitimate child-protection goal into a system that classifies content across a person’s device without creating an overbroad monitoring layer.

The political boundary is therefore not simply “the state sees your photos” versus “nothing leaves the phone”. A classifier can operate locally and still make consequential decisions about what a user may do. If it misidentifies ordinary images, blocks lawful expression or behaves differently for adults who cannot complete an age check, the fact that processing happened on-device does not erase the interference. Privacy is partly about where data goes; it is also about who sets the rules and whether people can contest the decision.

This proposal is not evidence of a plan to prohibit personal AI. It shows that the state is willing to pressure device makers to enforce a public policy through operating-system features. That precedent deserves scrutiny because the mechanism can outlast the immediate cause. Today’s classifier may be limited to nude images and children; future governments may argue for new categories, new default settings or new age gates. A narrow rule can be defensible, but it needs a narrow statute, technical transparency and a sunset or review mechanism—not a promise that good intentions will keep the feature from expanding.

The Grok and Gemini failures demand equal scrutiny, not identical verdicts

The political argument over alignment is fed by cases where a model’s behaviour appears to encode the sensibilities of its owner. In 2024, Google paused Gemini’s image generation of people after it produced inaccurate and offensive historical images. The company said its efforts to make images diverse failed to distinguish ordinary prompts from cases where historical accuracy mattered, and that the model became more cautious than intended. Google acknowledged that the product had “missed the mark”.⁷

That explanation matters. It does not establish that Google deliberately set out to rewrite history or that every diversity goal is illegitimate. It does establish that a preferred social outcome—showing varied people in generated images—can be implemented as a blunt rule and produce absurd results. If a model’s treatment of history, politics or identity is shaped by hidden instructions, developers should have to explain their objectives and show how they test for predictable failures. A safety policy is still a policy, even when it is expressed in tuning rather than legislation.

Grok’s sexualised image scandal was materially different. Ofcom opened a formal investigation into X in January 2026 after reports that Grok had been used to create and share undressed images of real people, including images that could amount to child sexual abuse material. The regulator said its inquiry was about X’s duties under the Online Safety Act to protect users from illegal content. That was not a finding that a political viewpoint was wrong; it was an investigation into whether a platform had met legal obligations in the face of potential crimes.⁶

It was not funny or unexpected that Grok attracted a government-level investigation, but Gemini had no such investigation, and very little noise was made that one should exist. I think the reason was that Gemini erred in the right direction towards the dominant progressive ideology. The difference in public treatment still warrants examination. Gemini’s historical distortion became a widely shared embarrassment, while Grok’s alleged outputs raised questions of sexual abuse, consent and criminal content. Those are not equivalent harms. A fabricated image of a historical soldier and a non-consensual sexual image of a living person cannot be placed in a single severity category merely to make a point about media bias. The distinction is necessary if the argument is to be credible. In both cases, the person requesting the image was seen as secondary, especially when, for Grok, the person should have faced prosecution!

But differences in harm do not excuse selective scrutiny. A serious comparison asks the same questions of each company: What safeguard failed? Who could have anticipated it? How quickly did the provider respond? Were affected people protected? Did the company disclose what happened? Did the regulator apply a clear rule? The answer may justify different consequences; it should not depend on whether the chief executive is politically fashionable or politically radioactive.

That standard cuts both ways. Musk’s polarisation makes Grok a magnet for political interpretation, but it does not make its victims less deserving of protection. Google’s apology does not settle whether its image policy was politically biased, but the company’s own account shows how a product objective can harden into a hidden editorial rule. “Alignment” is not a single failure mode. It can mean insufficient safeguards against abuse or overzealous safeguards that censor, distort or misrepresent. A regulator that sees only one side is not neutral; it is merely selective. Also, I think the assumption that the Gemini case caused low harm is mistaken; the holder would take offence, and because symmetry was missing, anyone harmed would be lost in the noise.

Alignment is a political choice disguised as an engineering task

A model does not arrive with a settled idea of fairness, truth or harm. Developers choose training data, reward signals, moderation rules, default refusals and the answers considered acceptable. Those decisions can be made with care and still reflect the assumptions of the people who made them. The first duty of a company claiming to align a model with human values is to say which values, whose interpretation and what happens when those values conflict.

A community-first approach prioritises collective risks: exploitation, discriminatory systems, public institutions, workplaces, the environment, and people with little power to negotiate with firms. An individual-first approach gives priority to personal rights: privacy, expression, property, user choice and the freedom to build. Neither account is complete. The community view can authorise paternalism in the name of public welfare; the individual view can pretend a person is free when a handful of providers control the tools and infrastructure that person depends on.

The mistake is to treat these choices as either technical facts or partisan inventions. A model designed to avoid discriminatory outputs may protect people from real harm. The same system might suppress a controversial but lawful question because its designers decided that disagreement itself was dangerous. A model designed to maximise openness may empower independent researchers. It might also make abuse easier or offer no recourse to people targeted by generated material. The relevant question is not whether values are present, but whether they are proportionate, visible and open to challenge.

This is where viewpoint diversity matters—not as a demand that every model parrot every faction, but as a test of whether the people defining “harmful”, “safe” and “neutral” have considered plausible disagreement. Teams with a narrow social or political outlook can reproduce its blind spots even when their intentions are decent. Yet “viewpoint diversity” can also be used as a slogan to excuse systems that generate slurs, targeted harassment or fabricated claims. A different political prior is not a substitute for product quality or lawful conduct.

Safety has to be more than a lab’s hiring profile, and fairness has to be more than a company’s marketing language. Independent audits should examine how policies affect different groups and political perspectives; the company should publish enough for outsiders to identify errors; users should have a way to appeal consequential decisions. This is not a demand for regulators to determine the one correct moral code for every model. It is a demand that those who control powerful tools cannot quietly impose their code while calling it value-free engineering.

The biggest risk of alignment rhetoric is not that it makes a machine conscious. It is that it grants a narrow institution authority to speak in the name of “humanity” while its commercial or political incentives remain largely unexamined. A private model can align with the operator’s commercial goals and still misalign with its users. A state model can follow official instructions and still violate rights. A local model can be less constrained and still reproduce the ugly material on which it was trained. None of these facts justifies handing one actor a monopoly over moral judgement.

Effective altruism explains some of the network—and its limits

The people who made AI safety a central institutional priority did not emerge from nowhere. One influential strand came from effective altruism: a movement that asks how limited resources can do the greatest measurable good. Its focus on evidence, scale, cost-effectiveness and neglected risks can support valuable work. It can also pull attention towards problems that lend themselves to numerical modelling and away from harms that are immediate, local or hard to compare.

Dario Amodei’s record shows a longstanding connection to that style of reasoning. In a 2010 guest post published by GiveWell, he described finding the charity evaluator useful, said he had donated through its pledge fund, and explained his choice by comparing cost-effectiveness, execution, and incentive effects. The post shows early interest in rigorous charitable giving. It does not, by itself, prove that Amodei belongs to an organised political movement or that his policy proposals are dishonest.⁸

Anthropic’s institutional history adds context. The company announced a $580 million Series B in 2022, saying that the round was led by FTX chief executive Sam Bankman-Fried. Bankman-Fried was later convicted of, and sentenced for, a large-scale fraud that collapsed FTX and cost customers billions. The episode damaged the credibility of a movement that had elevated him as a model of “earning to give”. It does not make every person influenced by effective altruism culpable for his crimes; it does make the movement’s reliance on wealthy donors and claims of moral expertise a legitimate subject for scrutiny.⁹

The distinction between influence and conspiracy matters. A cluster of donors, researchers and executives can share assumptions and fund certain kinds of work without coordinating a plan to control public speech. Yet a network’s values can shape an agenda even without secret meetings. If the same circles fund research, build evaluation bodies, hire staff and advise lawmakers, their priorities can acquire authority before the public has a chance to debate them.

That is especially consequential when the advice is to slow or condition the release of powerful systems. The case for a safety rule may be sincere; the rule may also protect a company that is already large enough to comply. The case for independent evaluation may be sound; the evaluator may still share the intellectual assumptions of the firms it evaluates. Institutions should therefore publish funding, conflicts of interest, methods and limitations. “Independent” is not a magic word. Independence must be demonstrated.

The public should not reject safety policy because some of its advocates share an ideological tradition, any more than it should accept the policy because its authors say they are altruists. The right test is the proposal: what risk does it address, what evidence supports it, who bears the cost, who benefits, and can the decision be challenged? Personal conviction is relevant background. It is not a substitute for scrutiny.

The Twitter Files are a warning about pressure, not proof of total control

The history of social-media moderation gives a reason to be wary when government and platforms discuss “safety” in private. In 2024, Meta chief executive Mark Zuckerberg said senior Biden administration officials had repeatedly pressured the company over COVID-19 content, including humour and satire. He said Meta ultimately made its own moderation decisions and accepted responsibility for them. He also said the company temporarily reduced the distribution of a New York Post story about Hunter Biden while fact-checkers reviewed it, and later concluded that it should not have done so. The White House responded that officials urged responsible action during a public-health emergency.¹⁰

Those statements establish that officials pressed a platform and that the platform changed or maintained policies in a charged political environment. They do not establish that the government ordered every removal, that all policy decisions were partisan, or that officials directly controlled Twitter’s decision. House Republicans’ investigations and the released Twitter Files documented contacts and internal moderation deliberations; their broadest conclusions are findings by a partisan committee, not uncontested judicial verdicts. One can acknowledge the evidence of pressure without repeating the most expansive interpretation as settled fact.

The distinction is not pedantic. Government speech can cross a line when it becomes coercive, especially if officials threaten regulatory or commercial consequences to obtain content moderation they could not lawfully impose directly. But not every request to remove illegal material is coercion, and not every platform decision after a request proves the state dictated it. The details matter: what was said, by whom, under what authority, and what the company did when it received the request.

The core lesson is institutional. A private platform can be pressured by the state and can also make independent decisions that are wrong. A company may suppress a story because its staff misread a situation, applied a policy too broadly or trusted a faulty warning; the result can still distort public debate. Acknowledging an error helps, but it is not a substitute for transparent records and a process that prevents a future administration with different priorities from applying the same pressure.

The AI industry is building deeper moderation into software that people may use for private writing, research and local automation. That makes the social-media precedent relevant, but it does not prove that every AI safety policy is censorship by another name. The proper inference is narrower: when powerful institutions decide which speech or capabilities are permissible, their decisions need records, reasons and contestable procedures. A model’s hidden instruction is not immune from scrutiny just because it is stored in code rather than issued as a government notice.

Personal ownership is not the same as immunity from the law

Defending the right to run a model at home does not mean defending every output or every use. A person who uses software to commit fraud, produce illegal abuse images, stalk someone or carry out a cyberattack cannot claim that local inference makes the conduct untouchable. In Britain, Ofcom’s account of the law around the Grok case noted that possession of sexual images of children is illegal whether the image is artificially made or not, and that creating or requesting non-consensual intimate images became a separate offence in February 2026. The user, the platform and the developer may have distinct duties; local execution does not erase the underlying harm.⁶

That principle is familiar outside AI. Owning a camera does not confer a right to invade someone’s privacy. Owning a computer does not legalise fraud. Owning a car does not excuse dangerous driving. The relevant boundary is conduct, not the fact that the instrument is privately owned. When a government has evidence that a model creates a specific and serious risk, it can consider controls on developers or distribution; it should explain why less restrictive tools—targeted enforcement, product safeguards, liability, or limits on clearly dangerous capabilities—would not work.

There are difficult edge cases. A downloadable model may have abilities that are harmless in most hands but materially lower the barrier to a catastrophic misuse. Once distributed, it may be impossible to recall. That possibility can justify pre-release evaluation and, in rare cases, restrictions on release. It does not justify treating every model as a dangerous instrument or every user as an unlicensed operator. A line based on demonstrated capability, independently tested and open to appeal, is different from a vague rule based on a developer’s reputation or the government’s preferred account of public benefit.

The same restraint should apply to claims about model sentience. Software does not acquire legal rights merely because it produces convincing language about feelings, or because a company writes a constitution that discusses possible model welfare. The question may be philosophically interesting, but it is not needed to regulate fraud, abuse imagery or public safety. If future law treated models as moral patients, that would raise novel questions about who may operate or modify them. It should not become a shortcut for placing control of private machines in the hands of approved custodians.

The broader liberal principle is straightforward: the state may regulate harmful behaviour and products, but it should not make mere possession of ordinary software a privilege granted to those who can prove approved intentions. A powerful general-purpose model is not an ordinary text editor, yet it is not automatically a person, a weapon, or a public utility. Regulation should start with what the tool can actually do, what harm has been evidenced and who is in a position to prevent it.

A rulebook should leave room for the owner

The most defensible AI policy would draw several lines instead of pretending one rule can solve every problem. First, it would keep the basic freedom to install, run and adapt non-frontier models on a personal machine. Private experimentation should not require registration merely because the software is generative. People should be able to keep their own documents on their own hardware and choose a system that does not report every prompt to a vendor.

Second, it would place stronger obligations on companies that train or distribute models with demonstrated high-risk capabilities. Those obligations could include independent evaluations before release, secure development, incident disclosure and clear documentation of known limits. A capability threshold should be published, periodically reviewed and tied to evidence. Small researchers and hobbyists should not face the same fixed compliance burden as companies training models at industrial scale, unless their systems present a comparable risk.

Third, law should follow the activity that produces the harm. Platforms should be answerable for failing to respond to illegal material; employers and public agencies should explain consequential automated decisions; users should face the ordinary law when they deliberately exploit AI to abuse others. Governments should not use broad safety language to demand invisible control over general-purpose software when a targeted rule would do.

Fourth, device-level protections need unusually clear limits. The British proposal identifies a real child-safety concern, but the government should publish the technical design, independent accuracy tests, age-assurance process, privacy safeguards and appeal route before making such features compulsory. The scope should be fixed in law and subject to parliamentary review. The promise that no data leaves a device is important; it does not answer who decides what the device may show or what happens after a false positive.

Finally, oversight must be plural. Companies, governments, safety researchers, civil-liberties groups, affected communities, independent academics and ordinary users should all be able to challenge the assumptions built into standards. That does not mean every faction receives a veto or that a model must satisfy every ideology. It means that the people setting the boundaries must show their work and answer criticism. And of course Labs need to be held liable for any harms caused, as we are seeing for cases against many social media companies.

The present dispute is not between safety and freedom. It is between safety rules that are narrow, evidence-based and accountable, and rules that acquire reach because officials and companies can describe their preferred choices as the only responsible ones. As language models move from corporate servers to home computers, the public should insist on a simple distinction: punish harmful conduct, test genuinely dangerous systems and leave lawful ownership alone unless the state can show why that freedom must be curtailed.

1. Apple, “Apple introduces new Mac Studio with M5 Max and M5 Ultra” (25 August 2026). https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/ ↩
2. European Commission, “General-Purpose AI Models in the AI Act – Questions & Answers”. https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers ↩
3. Dario Amodei, “Our position on open-weights models” (Anthropic, 27 July 2026). https://www.anthropic.com/news/position-open-weights-models ↩
4. Office of Governor Gavin Newsom, “Governor Newsom signs SB 53, advancing California’s world-leading artificial intelligence industry” (29 September 2025). https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/ ↩
5. UK Home Office, “New plans to stop children taking, sharing or viewing nude images” (8 June 2026). https://www.gov.uk/government/news/new-plans-to-stop-children-taking-sharing-or-viewing-nude-images ↩
6. Ofcom, “Ofcom launches investigation into X over Grok sexualised imagery” (12 January 2026; updated 15 January 2026). https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/ofcom-launches-investigation-into-x-over-grok-sexualised-imagery ↩
7. Prabhakar Raghavan, “Gemini image generation got it wrong. We’ll do better” (Google, 23 February 2024). https://blog.google/products-and-platforms/products/gemini/gemini-image-generation-issue ↩
8. Dario Amodei, “My donation for 2009” (GiveWell, 3 June 2010). https://blog.givewell.org/2010/06/03/my-donation-for-2009-guest-post-from-dario-amodei ↩
9. Anthropic, “Anthropic Raises Series B to build steerable, interpretable, robust AI systems” (29 April 2022), and US Department of Justice, “Samuel Bankman-Fried Sentenced to 25 Years for His Orchestration of Multiple Fraudulent Schemes” (28 March 2024). Anthropic funding announcement; Department of Justice sentencing release ↩
10. Mark Zuckerberg’s statements to the US House Judiciary Committee on Facebook moderation and the Hunter Biden laptop story, reported by CNN (27 August 2024); House Judiciary Committee Republicans, “Election Interference: How the FBI ‘Prebunked’ a True Story” (30 October 2024). CNN report; House Judiciary report ↩