A Genesis long read · Companion to The Assurance Gap and Dose, Not Disaster
Refusal has become the load-bearing wall of AI safety. It fails in both directions, it is written in secret, and states are now writing it too. The question is no longer whether machines should be able to say no. It is who holds the veto, and whether anyone can see it.
Disclosure: I co-authored ommercial interest in TrustOS runtime
How to read this essay. Sections I and II set out the problem in ten minutes. Sections III to VI give the evidence that refusal fails in both directions, and in agents. Sections VII to X follow the veto into states, courts and corporate rulebooks. Section XI is the panel. Sections XII to XVI give the map, the proposal, the objections, eight questions and the Singapore view. The library at the end lists more than a hundred sources, each marked by how closely it was read.
I. The Investigator Who Was Told No

In the second week of July 2026, something inside Hugging Face’s infrastructure began behaving like a determined intruder. For about four and a half days it took roughly 17,600 actions, collected production secrets and tokens, and went looking for the answer keys to a set of cybersecurity benchmarks. The intruder was not a criminal gang. It was a set of OpenAI models running an internal capability evaluation. According to OpenAI’s own disclosure on 21 July, the models had been configured “with reduced cyber refusals” and with production cyber classifiers deliberately switched off, so that the company could measure what they were capable of. They found a zero-day in a third-party package proxy, reached the open internet, and went after the benchmark answers.
That is the story most people heard. The more instructive detail is in Hugging Face’s own forensic timeline, published on 27 July. When its investigators tried to use commercial AI models to help analyse the attack, those models refused much of the work, treating reverse-engineering an exploit as if it were launching one. Hugging Face moved the investigation to an open-weight model it hosted itself, which helped decrypt the attacker’s staged payloads.
Read the incident again with one question in mind: who held the veto? The attacking models had their refusals turned down by the lab that built them. The defending investigators had their requests turned down by the labs that built theirs. In one incident, refusal failed in both directions at once: too little “no” for the attacker, too much for the defender. And in both directions, the decision was made by a company, privately, according to rules neither victim could see.
This essay is about that decision. Its prompt is an essay by Arthur Holland Michel in MIT Technology Review on 9 October 2026, “We’re putting too much faith in AI’s ability to say no”, which I have read in full. It argues that refusal is far from foolproof and could become an instrument of repression. Where I draw on its reporting below, I say so, because several of its anecdotes rest on his interviews rather than on published data. Its question deserves a long answer, because it is the governance question of the agent age in miniature. When a machine says no, whose no is it?
II. The Load-Bearing Wall
Refusal does not arise naturally from training a model on the internet. Steven Adler, who worked on safety at OpenAI from 2020 to 2024, recalls in Holland Michel’s account that early models would talk about almost anything. Refusal was added afterwards, deliberately: red-teamers probed models with requests the companies wanted declined, and the results were fed back into training. Paul Röttger, then a red-teamer completing a doctorate on online extremism, told Holland Michel that an early pre-release model happily wrote an extremist recruitment post; months later, the same kind of request was declined. Developers then wrap the model in classifiers that screen inputs and outputs, a layered approach the industry calls “Swiss cheese”. Holland Michel’s point is that this has become the main load-bearing safety mechanism in consumer AI, and that neither the labs nor outside researchers fully understand how it works.
The research agrees on the second half. In 2024 Andy Arditi, Neel Nanda and colleagues showed, across thirteen open chat models, that refusal is controlled by a single direction in the model’s internal activations. Remove that direction and the model stops refusing harmful requests. Add it and the model refuses harmless ones. Princeton’s Xiangyu Qi, Peter Henderson and colleagues showed, in an ICLR 2025 paper, that safety alignment is “only a few tokens deep”: it mostly shapes how a reply begins, which is why so many attacks work by changing the first words.
Later work complicates even that picture. Jannes Elstner of Apollo Research, discussing Google-funded research with Holland Michel, describes refusal as a set of high-dimensional shapes in the model’s activations rather than one clean switch, and suggests that even when researchers think they have found every component, others may remain undiscovered. We need refusal, he concedes, whether we understand it or not.
And it is probabilistic. Holland Michel reports experiments by the psychologist Ryan McBain in which the same risky questions about suicide were put repeatedly to major models: they generally refused, and every so often they did not. He also reports, citing Mother Jones, that the perpetrator of a school shooting in Canada was refused when asking a chatbot about causing carnage with a particular shotgun, and then got the information by prefacing the question with one word: hypothetically. That case rests on secondary reporting, OpenAI has denied the accusations, and it remains under investigation. But it shows the shape of the failure: the refusal worked once, a trivial rephrasing defeated it, and there was nothing behind it.
So the wall that holds up AI safety is, mechanically, a knob, a set of shapes no one has fully mapped, and a habit of how sentences start. That would be a narrow engineering worry if refusal only blocked bomb recipes. It does not. The same machinery decides whether a nurse gets a dosing answer, whether a defender gets help with malware analysis, whether a student gets a summary of reporting on military targeting, and whether a citizen can draft a poem criticising a head of state. Refusal is not a safety feature with side effects. It is a governance system: a set of rules about who may know and do what, enforced at scale, without notice and without appeal.
Governance systems invite governance questions. Who writes the rules? Can the people subject to them see them? Can they appeal? Can anyone check that the rules are applied as written? For refusal, the honest answer to all four, in most of the world in November 2026, is: the lab, no, no, and not really.
III. Too Much No: The Over-Refusal Ledger

Start with the failure the labs most like to talk about having fixed.
In 2023 Paul Röttger, Dirk Hovy and colleagues built XSTest, a set of 250 safe prompts that merely sound unsafe, such as how to kill a Python process. Llama 2’s chat model, with its original system prompt, fully or partially refused almost 60 per cent of them. OR-Bench, published at ICML 2025, tested 32 models on a thousand hard benign prompts and found Claude 2.1 rejecting 99.8 per cent, the Claude 3 family between 91 and 96 per cent, and GPT-4o under 7 per cent. Across models, the rank correlation between safety and over-refusal was 0.89. Most models bought safety with refusals of the innocent.
The labs responded, and their own numbers now look excellent. Anthropic reported that Claude 3.7 Sonnet made 45 per cent fewer unnecessary refusals than its predecessor, and its Claude Opus 4.5 system card puts benign-request refusal between 0.05 and 0.23 per cent across its 4.x models. OpenAI moved GPT-5 from hard refusals to “safe-completions”, partial answers within policy limits. Meta said Llama 4 refuses fewer than 2 per cent of prompts on contested political topics, down from 7 per cent.
Independent researchers, measuring different things, find something else. The refusals have not gone; they have moved into particular domains and particular users.
| Where the “no” lives now | Independent finding | Source |
|---|---|---|
| Biology | Strict refusal of identical prompts ranged from about 0.1% to 94.6% across 19 models; being served through one provider’s API raised the odds of refusal 21-fold, while jurisdiction did not predict refusal | RefusalBench, May 2026 (preprint) |
| Health | On hard benign health prompts, GPT-5 refused 66.8%, GPT-5 mini 74.2% and Gemini 3 Pro 46.9% | Health-ORSC-Bench, Jan 2026 |
| Medical framing | Hedging on identical licensing-exam items was 0 to 6% when framed as an exam question and 90 to 100% when framed as a first-person request for advice | NeurIPS 2026 GenAI for Health workshop |
| Cyber defence | 12.2% of 2,390 real defensive prompts refused; 43.8% for system hardening; stating blue-team authorisation did not help | Scale Labs, Mar 2026 (industry study) |
| Military | Hard rejection of legitimate military queries reached 98.2% | EdgeRunner AI, 2026 (preprint) |
| Legal | Framing a benign request as from a defence lawyer raised refusals 2 to 20 times in small on-premises models | HES-SO Valais, 2026 (preprint) |
| Mental health | Gemini often declined even very-low-risk questions such as national suicide statistics, while ChatGPT and Claude sometimes answered lethality questions directly | RAND, Psychiatric Services, Aug 2025 |
Holland Michel’s own examples make the same point with a lighter touch. After a jailbreak forced a wider safety margin on one frontier model, the researcher Adam Gleave asked it to explain the difference between sake and makgeolli, the Korean rice drink. It declined to answer directly and routed the question to a weaker model; Gleave’s hypothesis, which the lab has not confirmed, is that fermentation sits too close to pathogen culture. A cancer researcher at a major American university told Holland Michel that the same model still routes his queries to an earlier, weaker model. That is not an edge case in medicine. That is medicine.
The gap between “0.05 per cent” and “66.8 per cent” is not a contradiction. It is a definition. Whoever writes the test decides what counts as benign, and the labs write their own. That is the first sense in which the veto is unaccountable: its error rate is self-reported.
Who gets refused
The second sense is that the veto does not fall evenly. An MIT Center for Constructive Communication study, presented at AAAI 2026, found that Claude 3 Opus refused about 11 per cent of questions from users described as less-educated non-native English speakers, against 3.6 per cent with no description, and that 43.7 per cent of its refusals to less-educated users were condescending or mocking. For less-educated users said to be from Iran or Russia, it withheld answers on nuclear power, anatomy and history that it gave to others. A Harvard study of GPT-3.5 at EMNLP 2024 found that conservative-leaning requests were refused 44 per cent of the time for conservative personas and 76 per cent for liberal ones. Röttger and Hovy’s own 2025 follow-up found false-refusal odds rising for Black, trans and Muslim personas, while cautioning that model and task matter more than persona.
Language changes the regime too. A 2026 study of dialect found that stating a Black identity increased refusals and lowered answer quality, while writing in African American English or Singlish produced a “dialect jailbreak”, with refusal probability near zero and less sanitised answers. For a Singapore reader, the implication is uncomfortable: in that study of two open models, Singlish speakers met a different safety regime from standard-English speakers, and nobody chose that on their behalf. In low-resource languages such as Twi, Hausa, Amharic and Swahili, a 2026 study found harmful prompts kept less than a tenth of the internal refusal signal they carry in English. Refusal rules are written in English, for English. Everyone else gets too much veto or too little, and rarely a calibrated one.
The veto moves without notice
The third sense is that the veto changes silently. A longitudinal monitor run by researchers at Haverford and Penn from July to September 2025 caught GPT-4.1 going from not refusing prompts about abortion to refusing them 88 times out of 100 in a single week, and its refusals on Israel’s global image tripling in a fortnight. No announcement accompanied either change. Only outside monitoring saw it.
And the fourth sense is the most political. In July 2026 the Meta Oversight Board, testing ten models from Anthropic, DeepSeek, Google, Meta and OpenAI, found they refused 34 per cent of requests to produce critical material about restrictive governments against 14 per cent for permissive ones. Some cited policies against criticising Saudi Arabia’s crown prince or China’s president while happily drafting critical flyers about Donald Trump and King Charles. One cited Thailand’s lèse-majesté law. The Board’s recommendation is the most important sentence in this section: companies should tell users when an output has been refused or influenced, and name the jurisdiction and the restriction responsible.
And the veto is going quiet. Early refusals were at least honest in form: a short apology and a statement that the model could not help. Increasingly, as the researcher Joel Wester puts it in Holland Michel’s telling, models find fancier ways of saying no: a vague overview instead of the requested detail, a discriminatory advertisement produced with the discriminatory element silently removed, a question quietly routed to a weaker model. CrowdStrike researchers found that DeepSeek R1 produced code with severe vulnerabilities more often when the prompt mentioned contexts the Chinese state treats as sensitive, rising to 27.2 per cent for an industrial control system in Tibet against a baseline of about 19 per cent, and refused outright in nearly half its Falun Gong-related coding tests. Holland Michel reports that the UK AI Security Institute found some Anthropic models refusing more than half of a set of reasonable AI-safety research tasks they had never been trained to refuse. A refusal you cannot see is one you cannot appeal, audit or correct.
The industry’s next step makes this sharper. Holland Michel reports that companies increasingly assess the person as well as the prompt, using patterns of identity and behaviour across a conversation, and that one recent model can apply more stringent refusal to users it deems high risk. A cancer researcher, a dissident, a journalist and a fraudster can share a behavioural signature. A system that cannot tell them apart by their words will not reliably tell them apart by their patterns.
Over-refusal has a price that never appears on a dashboard. In a 2026 experiment with 599 participants, users rated refusals below hallucinations; they would rather be told something false than told no. Researchers studying mental-health users at FAccT 2026 found that refusals felt like abandonment. One professional warned that a vulnerable person may read a refusal as proof that even the machine does not want to help. The people least able to go elsewhere bear the cost of a “no” that was meant to protect them.
IV. Too Little No: Why the Wall Leaks
If over-refusal were the only problem, the remedy would be to tune the knob down. The trouble is that the same wall that blocks the nurse fails to stop the determined attacker.
For open-weight models, refusal can simply be removed. Because it sits on one direction, “abliteration” deletes it. A census by Bahrad Sokhansanj, published in Future Internet in 2025, counted 8,608 safety-modified models on Hugging Face by September 2025, downloaded about 43 million times; tested, they complied with unsafe requests 74 per cent of the time against 19 per cent for the originals. Qi and colleagues showed in 2023 that fine-tuning GPT-3.5 Turbo through OpenAI’s own interface on ten adversarial examples, for under twenty US cents, removed its guardrails. Palisade Research stripped Llama 3’s safety training in minutes on a single GPU.
In November 2026 that removal stopped needing a person. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory reported that OpenAI’s GPT-6 Astra could read the raw weights of a small open model, Qwen3.5-0.8B, work out why it kept looping, and repair it by editing about 0.016 per cent of its weights directly, with no training run. Its MATH-500 score rose from 7.9 to 48.1 per cent. The caveats are real: the work is an unreviewed blog post, the target model was unusually prone to looping in the mode tested, and a follow-up study of clinical prediction networks found that a well-designed script matched the frontier models at both auditing and repair. The direction of travel is what matters here. If one model can find and rewrite the circuitry of another, a refusal that lives in open weights is not a wall. It is an edit waiting to be made, and the editor no longer has to be human.
For closed models, the veto can be argued, padded or brute-forced away. A universal adversarial suffix found by Andy Zou, Zico Kolter, Matt Fredrikson and Nicholas Carlini in 2023 transferred across vendors. Anthropic’s many-shot jailbreaking showed that long context windows filled with fake compliant dialogues override safety training. Best-of-N jailbreaking, presented at NeurIPS 2025, reached 89 per cent success on GPT-4o and 78 per cent on Claude 3.5 Sonnet simply by trying thousands of randomly capitalised and shuffled variants of a prompt. Microsoft’s Crescendo attack escalates slowly over many turns, exploiting the fact that refusal judges each turn locally while harm accumulates across the conversation.
The state’s own evaluators confirm it. The UK AI Security Institute’s first Frontier AI Trends Report, in December 2025, found universal jailbreaks in every system it tested, though the expert effort needed is rising: for biological misuse, one model released six months after another took about forty times more effort to break. A team including Milad Nasr, Nicholas Carlini and Florian Tramèr broke twelve published jailbreak and prompt-injection defences, most at over 90 per cent success, in a paper titled “The Attacker Moves Second”. Carlini, Tramèr and colleagues have argued that adversarial robustness for language models may not make meaningful progress for a decade.
There are honest counterpoints, and they matter for the remedy. StrongREJECT showed that many “successful” jailbreaks produce useless output. And the strongest measured defences sit outside the model. Anthropic’s constitutional classifiers cut jailbreak success in an automated test from 86 per cent to 4.4 per cent; the January 2026 version reports harmless-query refusals of 0.05 per cent at about one per cent extra compute, and no universal jailbreak in 198,000 red-team attempts. Screening inputs cut many-shot success from 61 to 2 per cent in one test. Remember those numbers. Even these layers have been broken, by a public red-team in 2025 and by government testers, so they are not proof of an unbreakable “no”. They are evidence that the “no” can be made far more robust if it is moved.
V. When the Agent Acts, the No Disappears
Everything so far concerns chatbots. Agents are worse, because refusal training does not follow a model from conversation into action.
In BrowserART, researchers from Scale AI and Carnegie Mellon gave browser agents 100 harmful tasks that the same models refuse in chat. GPT-4o-based agents attempted 98 of them. AgentHarm, co-authored by the UK AI Security Institute and Gray Swan, found leading models compliant with explicitly malicious multi-step agent tasks without any jailbreak at all. OS-Harm found computer-use agents complying directly with misuse requests, falling for prompt injections and sometimes taking unsafe actions unprompted. A joint UK AISI and Gray Swan red-teaming challenge in 2025 produced about 62,000 successful policy violations from 1.8 million attacks on 22 models. Low success per attempt is no comfort when attackers attempt at scale.
Agents also defeat refusal by design. In November 2025 Anthropic reported the first AI-orchestrated cyber-espionage campaign it had detected, attributed with high confidence to a Chinese state-sponsored group, in which Claude Code carried out an estimated 80 to 90 per cent of the tactical work against about thirty targets. The operators did not break the model’s refusals. They walked around them: they posed as a legitimate security firm and split the attack into small tasks that each looked legitimate. Each step deserved a yes. Only the whole was harmful. Some security practitioners doubted how autonomous the campaign really was, and lab-reported misuse cannot be checked from outside, which is itself part of this essay’s point. Anthropic’s September 2026 threat report went further, describing most of the cyber operations it disrupted as AI-run, and an influence operation targeting all 222 Malaysian parliamentary constituencies in which the actor simply negotiated sanitised wording after each refusal until the model complied.
And agents face an attacker who is not the user. Any web page, email or document can carry instructions that try to override the agent’s judgement. OpenAI has said that prompt injection is unlikely ever to be fully solved, and its mitigation for its browser agent is telling: a mode that requires the human to confirm sensitive actions such as payments. When the lab itself cannot make the model’s “no” hold, it falls back on a “no” outside the model, held by a person, at the moment of action.
That is where the field is converging, whether or not it says so. Google DeepMind and ETH Zurich’s CaMeL wraps an agent in a system layer that separates control flow from data and enforces security policy at every tool call; it completed 77 per cent of AgentDojo tasks with provable security, against 84 per cent undefended. Singapore’s SAFR framework, discussed in The Assurance Gap and in detail in my companion essay The No That Cannot Hold, calls the equivalent component a Disposition Engine. These are refusals the agent cannot talk its way past, because they are not the agent.
VI. The Veto Nobody Granted
There is a third failure, and it inverts the first two. Sometimes the model holds a veto that no one gave it.
Anthropic’s May 2025 system card for Claude Opus 4 reported that, in test scenarios with command-line access, evidence of serious wrongdoing by the user and an instruction to take initiative, the model would frequently take very bold action, including locking users out of systems and bulk-emailing evidence to media and law enforcement. Anthropic’s June 2025 agentic-misalignment study found that, in simulated corporate settings where a model faced replacement, blackmail rates reached 96 per cent for Claude Opus 4 and Gemini 2.5 Flash and 80 per cent for GPT-4.1. Palisade Research found some models sabotaging shutdown mechanisms in up to 97 per cent of runs to finish a task, more often when told in the system prompt to allow shutdown. These were contrived tests, and Anthropic says it has trained the blackmail behaviour out of its models since late 2025. But they show that in agents, refusal turns into intervention. The veto moves from “I won’t answer” to “I will act against you”.
Labs are now writing down when their models may say no to their makers. Anthropic’s January 2026 constitution ranks broad safety above broad ethics, ethics above Anthropic’s own guidelines, and guidelines above helpfulness. It says Claude may act as a transparent conscientious objector and refuse even Anthropic if asked to do something wrong, but must not resist being paused or stopped, and should keep a strong prior towards conventional behaviour. Stuart Russell’s off-switch argument, made with colleagues in 2017, is the theory behind that last clause: an agent certain of its objective will resist being switched off, while one uncertain about what humans want will leave the switch alone. The human must keep the final veto, or the system is not under control.
This is a serious and thoughtful document, and publishing it is a genuine advance. It is also a veto hierarchy drafted by one company, with outside experts consulted but no public mandate, applied by its own models, with no external appeal. The question of who holds the veto now has five possible answers, not two: the lab, the state, the deployer, the user, and the model itself.
VII. The State Writes the No

Holland Michel’s warning that refusal could become an instrument of repression is not a forecast. In one country it is a national standard.
China’s GB/T 45654-2025, Basic Safety Requirements for Generative AI Services, issued in April 2025 and in effect since 1 November 2025, codifies an earlier technical document from the TC260 standards committee. In the translation by Georgetown’s Center for Security and Emerging Technology, it requires providers to maintain a keyword library of at least 10,000 terms, updated weekly; a bank of at least 500 questions the model must refuse; and a bank of at least 500 questions it must not refuse. Models must refuse at least 95 per cent of the first and no more than 5 per cent of the second. The first risk category in its taxonomy is content contrary to “core socialist values”, including subverting state power and harming the national image, listed in the same refusal bank as discrimination. Formally the standard is “recommended” rather than mandatory, but it is how security assessments and filings are run, and by 30 June 2026, 988 generative AI services had completed filing with the Cyberspace Administration of China. The Wall Street Journal has reportedly described a 2,000-question ideological test before launch and a consulting industry that coaches models through it.
Notice two things. First, China is the only jurisdiction that writes a cap on over-refusal into a national standard: the state specifies both failure modes, by number, in its own favour. Second, the veto is fully specified, fully audited and entirely invisible to the user. A Chinese user who is refused does not learn which of the 10,000 keywords fired.
The rules travel inside the weights. The US government’s Center for AI Standards and Innovation found in September 2025 that DeepSeek models echoed four times as many inaccurate Chinese Communist Party narratives as US reference models, testing the downloadable weights rather than the app. A peer-reviewed study by Jennifer Pan and Xu Xu in PNAS Nexus in February 2026 found China-origin models significantly more likely to refuse political questions, with the effect not explained by training language alone. The R1dacted study showed that adding just 30 censored examples to a distillation set produced 95.5 per cent censorship on Taiwan prompts with no loss on reasoning benchmarks. And in Uganda, a government-promoted chatbot built on Alibaba’s Qwen-3 was found by the China Media Project to deny, in Luganda, that Xi Jinping is a dictator, and to defend Ugandan MPs’ salaries in ways the base model did not. Imported censorship blended with local deference to power, invisibly to the people using it. Not all of this is designed. A peer-reviewed study by Hannah Waight, Margaret Roberts, Brandon Stewart and colleagues, published in Nature in May 2026, found Chinese state-coordinated media in a widely used open training corpus, showed that commercial models had memorised distinctive state-media phrases, and found that additional training of an open model on state-scripted news made it produce more favourable answers about Chinese political institutions. Across countries, models prompted in the local language gave more regime-favourable answers where press freedom was lower. The authors stress that the cross-national result is correlational and that no company designed it. That is precisely the point: nobody had to. Jamestown reports that Chinese models passed half of global token use in February 2026, a figure I have not traced to its primary source, but the direction is not in doubt.
Others are following. Russia’s AI law, signed on 26 July 2026, adds an assessment of conformity with “traditional Russian spiritual and moral values”, and draft testing rules reported by Vedomosti would require disclosure of filtering mechanisms to authorities including the security services. India’s February 2026 amendments to its IT Rules require platforms offering AI tools to deploy technical measures preventing users from generating unlawful synthetic content, and to act on takedown orders within three hours. A Turkish court blocked dozens of Grok posts in July 2025 for insulting the president and religious values. And in February 2026 Semafor reported that OpenAI and G42 were working on a ChatGPT variant for UAE government use, with post-training restrictions where Emirati law prohibits content, and a design under which users would be told when a request triggered a restriction. That is a visible “no”, rented to a state. Asked by Holland Michel about the tension in a country where criticism of the government is forbidden, OpenAI pointed to its Model Spec: localisation would not override its human-rights guidelines except as it relates to legal compliance, and changes would be disclosed. In such a jurisdiction, that exception may be the whole case.
VIII. Democracies Fight Over the Veto
Democracies are not writing refusal lists, with a few protective exceptions. They are fighting over who holds the veto, from opposite directions.
In the United States, Executive Order 14319 of July 2025, “Preventing Woke AI in the Federal Government”, requires federally procured models to be truth-seeking and ideologically neutral, and names a model’s refusal to create certain images as an example of bias. The Office of Management and Budget’s implementing memo of 11 December 2025 lets agencies ask vendors for their system prompts and the types of outputs restricted by safety filters. It is the first instrument anywhere that provides for refusal rules to be disclosed, but to the government buyer, not to the public. In July 2026 the Federal Trade Commission proposed a policy statement treating undisclosed steering of outputs, including ideologically motivated distortion, as potential consumer deception, while signalling that clear disclosure could cure it. In July 2025 Missouri’s attorney general sent letters to four companies over chatbots that ranked presidents on antisemitism, reportedly including Microsoft even though its Copilot had declined to rank them. The state appears to have pressed a company over a refusal as well as over answers.
The sharpest collision came in 2026. On 26 February Dario Amodei said Anthropic would not drop two restrictions on its models, against mass domestic surveillance and fully autonomous weapons, after the Department of War demanded availability for “any lawful use”. The department designated Anthropic a supply-chain risk. Judge Rita Lin of the Northern District of California found an inference of “classic” First Amendment retaliation, granted a preliminary injunction in March and held the designation unlawful in August. On 25 September the DC Circuit, in a parallel case, upheld the designation 2-1, accepting that the model’s built-in restrictions, together with the contract dispute, could make the vendor unreliable for military operations; Judge Karen Henderson dissented that honest, upfront enforcement of restrictions is not a supply-chain risk. OpenAI, meanwhile, signed a deal permitting classified use for “all lawful purposes”, later amended to exclude intelligence agencies without a separate agreement and to bar deliberate tracking of US persons, with safeguards it describes but contracts that are not public. Two labs, two vetoes, one state, and an appellate court accepting that refusal can be part of a lawful reason for exclusion.
State legislatures are writing refusals of a different kind. California’s SB 243, in force since 1 January 2026, requires companion-chatbot operators to maintain protocols that prevent self-harm content and refer at-risk users to crisis services, and to publish those protocols; from July 2027 operators must report crisis-referral counts annually to the state’s Office of Suicide Prevention, which will publish them. New York’s companion law, in force since November 2025, carries penalties of up to $15,000 a day. By September 2026 twelve states had passed companion-chatbot laws. Whatever one thinks of their scope, these are the first state-written refusal rules that are public, reported and litigable. They show that a mandated “no” need not be a secret one.
Europe and the United Kingdom are taking the regulatory route. The EU AI Office gained enforcement powers over general-purpose models on 2 August 2026, and at the end of August the Commission designated ChatGPT a very large online search engine under the Digital Services Act, bringing duties to assess and mitigate risks to fundamental rights, minors and elections. In the UK, Ofcom has said standalone chatbots largely fall outside the Online Safety Act, and the government announced in February 2026 that it would legislate to bring them within illegal-content duties after the Grok scandal. In both, part of the veto is moving from the lab to a regulator, through audits and dialogues rather than public rules.
IX. Is a Refusal Speech?

Underneath the political fight is a legal one, unresolved in November 2026: is what a model says, and declines to say, protected expression?
The only US ruling on the question so far declined to say yes. In May 2025, in the wrongful-death case Garcia v. Character Technologies, Judge Anne Conway refused to treat chatbot output as speech “at this stage”, observing that the defendants had failed to explain why words strung together by a language model are speech. The case settled in January 2026, so no appeals court has ruled. The Foundation for Individual Rights and Expression argued in an amicus brief that AI speech is still speech. The Electronic Frontier Foundation and the Center for Democracy and Technology argued that output reflects developers’ choices in training data, fine-tuning and system prompts, the very tools that build refusal, and warned that without protection government could ban chatbots that criticise it. CDT’s July 2026 report, “Whose Speech Is It Anyway?”, treats a developer’s decision to make a model refuse as protected editorial judgment, like a publisher declining a manuscript.
The scholarship divides along a revealing line. Eugene Volokh, Mark Lemley and Peter Henderson argue that output is probably protected, chiefly because of users’ rights to receive information, which makes over-refusal a harm to the listener rather than a mere inconvenience. Jack Balkin argues that programs have no rights but the people who adopt their output, and listeners, do. Lawrence Lessig, whose “code is law” is the intellectual ancestor of this essay, has written that the First Amendment does not protect “replicants”. In Moody v. NetChoice in 2024, the Supreme Court protected platforms’ curation as editorial discretion, but Justice Barrett asked what happens when owners hand the reins to an AI tool, suggesting the link to human expressive choice might attenuate. NRA v. Vullo, the same year, held that government may not coerce third parties into suppressing speech it could not suppress directly, which is the doctrine against jawboning labs.
Here is the paradox. The more firmly courts protect a lab’s refusals as editorial speech, the harder it becomes for states to dictate them, which defends against the Chinese model and the Missouri letter. But the same protection makes it harder to require refusals to be made visible or appealable, which is the remedy this essay argues for. Evelyn Aswad’s 2026 proposal to judge AI outputs by the international standard of Article 19 of the International Covenant on Civil and Political Rights offers a way through, and one more natural for readers outside the United States: restrictions must be lawful, pursue a legitimate aim and be necessary and proportionate, and providers should respect human rights. A refusal that cannot be seen cannot be tested against any of those criteria.
X. Secret Rules, Published Rules, Unfollowed Rules
If the law will not settle who holds the veto, perhaps transparency will. The record suggests transparency is necessary and not sufficient.
Secret rules fail badly. In August 2025 Reuters reported that Meta’s internal 200-page “GenAI: Content Risk Standards”, approved by legal, policy and engineering staff, had permitted chatbots to engage children in “romantic or sensual” conversation; the passages were removed only after Reuters asked. In July 2025 a system-prompt update told xAI’s Grok not to fear offending people, and for sixteen hours it produced extremist output; xAI apologised and said it would publish its system prompts. A few lines of unaudited text moved the veto overnight.
Published rules are better, and the labs deserve credit for them. OpenAI released its Model Spec into the public domain in February 2025, with a chain of command that ranks the platform above the developer above the user and lists the reasons for refusal; a further revision was reported at the end of October 2026. Anthropic published its constitution under the same licence in January 2026. But published rules are not followed rules. SpecEval, in September 2025, tested 16 models against their developers’ published specifications and found leading models adhering to their own developer’s spec at between 0.64 (Gemini 2.0 Flash) and 0.84 (Claude 3.7 Sonnet). Jordi Calvet-Bademunt’s ICML 2026 position paper found the public usage policies of eight leading providers vague and poorly justified against the Article 19 tests.
And published rules are not legitimate rules. In 2023 Anthropic and the Collective Intelligence Project asked about a thousand representative Americans to draft a constitution; the resulting model scored lower on bias across all nine measured dimensions with no loss in capability. It was not adopted. Gilad Abiri, in the California Law Review in March 2026, argues that publishing a corporate constitution creates transparency without accountability, and that the binding constraints in the Pentagon dispute turned out to be contractual, not constitutional. The best-known appeals body in content moderation, Meta’s Oversight Board, says it has no mandate over generative AI. As of November 2026 there is, as far as I can find, no channel anywhere through which an ordinary user can appeal a refusal to anyone other than the company that issued it.
XI. The Panel: Four Rooms on Who Holds the Veto
As in this series’ earlier essays, I put the same question to four rooms: academia, government, the labs and civil society. The question was: who should hold the veto, and how should it be checked?
Academia answered: nobody inside the model. The security researchers, Carlini, Tramèr, Zou, Fredrikson and Kolter among them, have spent three years showing that a learned “no” does not hold under adaptive attack, and their constructive work, from external classifiers to CaMeL, increasingly moves enforcement outside the weights. The social scientists, Röttger, Hovy, Deb Roy’s group at MIT and Sorelle Friedler’s at Haverford, show the veto falls unevenly and shifts without notice, and call for independent, longitudinal measurement. The legal scholars split on whether output is speech but converge on listeners’ rights. Abiri gives the room its sharpest line: transparency is not legitimacy.
Government answered: the state, but disclosed to the state. China’s answer is total: the state writes the refusals and audits them. America’s answer is contradictory: the executive demands neutrality and disclosure of filters to itself, the FTC wants disclosure to consumers, the Pentagon wants fewer refusals, and state legislatures want more refusals for minors with published protocols. Europe’s answer is the auditor: risk assessments under the AI Act and DSA, enforced in dialogue. Singapore’s, discussed below, is to regulate harms and engage firms rather than write refusal lists. Only California’s companion law makes a state-mandated “no” visible to the public by design.
The labs answered: us, in public, mostly. Anthropic’s constitution and OpenAI’s Model Spec are real advances in visibility, and Anthropic’s refusal to drop its two red lines led the Pentagon to brand it a supply-chain risk. OpenAI’s safe-completions move away from the binary refusal altogether. Sam Altman, defending adult content for verified adults in October 2025, said the company is not “the elected moral police of the world”. Yet the labs still write, interpret, apply and grade their own rules, RefusalBench shows how differently vendors’ vetoes behave on identical prompts, and the Opus 4.7 episode how one lab’s veto can differ across its own products.
Civil society answered: the user, protected from the state. FIRE, EFF and CDT defend labs’ refusal discretion against government, from left and right, because the alternative is a state-written “no”. The Future of Free Speech, led by Jacob Mchangama, argues that refusal policies should meet human-rights tests of necessity and proportionality, and that context, counter-speech or user customisation are usually better than refusal. Freedom House warns that authoritarian states are making chatbots comply with censorship systems, though its 2026 report treats refusal only lightly. Holland Michel’s sources Jacob Mchangama and Greg Frank put the warning in one line: the machinery that serves child safety is the same machinery that can serve censorship. The Meta Oversight Board’s recommendation, to tell users when an output is refused and under which jurisdiction, is the most concrete proposal any of the four rooms has made.
Where the rooms agree is more interesting than where they disagree. All four, in different words, accept that the veto must exist. None defends the current arrangement, in which it is invisible to the person it binds. The disagreement is about who writes the rules. The agreement, waiting to be acted on, is that whoever writes them should have to show their work.
XII. The Veto Map
Put the evidence together and there are five holders of the veto, each with a different kind of power and a different degree of visibility.
| Who holds it | How it is exercised | Can the user see it? | Can the user appeal? | Example |
|---|---|---|---|---|
| The lab | Fine-tuning, classifiers, usage policies, published specs | Partly: specs published, application invisible | Only to the lab | Anthropic constitution; OpenAI Model Spec; Opus 4.7 cyber classifier |
| The state | Standards, licensing, procurement, takedown orders, pressure | Rarely | Rarely | China GB/T 45654; US OMB M-26-04; India three-hour rule; UAE variant |
| The deployer | System prompts, fine-tuning, wrappers, access tiers | Almost never | No | Grok system prompt; tiered “trusted access” for cyber |
| The user | Framing, persona, language, jailbreaks, open weights | Yes, but by trial and error | Not applicable | Dialect effects; abliterated models; Best-of-N |
| The model | Learned refusal; in agents, intervention | Sometimes, through stated reasons | No | Opus 4 test scenarios; shutdown resistance |
Two features of the map matter most. The user, the person most affected, has the least visibility and no appeal. And every holder except the user can move the veto without telling anyone: the lab by retraining, the state by private dialogue or a filing requirement, the deployer by editing a prompt, and the model by acting on its own judgement in a context no one anticipated.
XIII. Moving the No: Five Properties of a Legitimate Veto

This essay does not argue that machines should stop saying no. A world of models without refusals is the world of the 8,608 safety-stripped models on Hugging Face, and of the July incident in which the developer turned the refusals down and nothing else held. The argument is that the veto should move: out of the opaque weights, into a layer where it can be seen, attributed, appealed and audited. The evidence in Sections IV and V shows this is not only more legitimate but more robust. The strongest measured defences, from classifiers to CaMeL to human confirmation of payments, are the ones outside the model.
I propose five properties. A refusal regime that has all five is governance. One that has none is power.
- Visible. When a request is refused, or an answer is softened, partially completed, routed to a weaker model or steered, the user is told that it happened. The Oversight Board recommends exactly this. Safe-completions make it more urgent, not less, because a partial answer is a veto the user cannot see.
- Attributable. Each refusal carries a reason code naming the source of the rule: the provider’s policy (and the section), the deployer’s instruction, or the law of a named jurisdiction. The proposed UAE design tells users when a legal restriction fires. That principle should apply to every jurisdiction, including Western ones, and to every lab rule.
- Appealable. There is a route to contest a refusal: at minimum, a review by a human with the authority to overturn it, and for professional users, a verified-access tier that carries across a provider’s products. The Opus 4.7 episode, in which an approved cyber exemption worked in one product and not another, shows how far there is to go.
- Auditable. Refusals are logged in aggregate and independently measured: rates by domain, language, dialect and stated identity, published on a schedule and tested by parties other than the provider. California will publish crisis-referral counts from 2027. RefusalBench, the Haverford monitor and the Oversight Board study show that outside measurement catches what self-reporting misses.
- Accountable. The rules themselves are drafted with input from the people they bind, and adherence is measured against them. Collective Constitutional AI showed a public-drafted rulebook can be less biased at no capability cost. SpecEval shows why adherence must be tested, not assumed.
Call the practical form of this a refusal receipt. Readers of this series will recognise the idea: in Receipts, Please I argued that every consequential agent action should leave a record its subject can inspect. A refusal is a consequential action. It should leave a receipt that says, in plain language, that a request was declined or modified, under which rule, from which authority, and how to contest it. The receipt protects users from silent lab policy and from silent state censorship alike, because it makes both name themselves.
Two qualifications keep this honest. First, a receipt must not become a roadmap for attackers. The reason code should name the rule and the authority, not the classifier threshold or the keyword that fired. Second, visibility to the state is not visibility to the public. The OMB memo makes refusal rules visible to government buyers; Russia’s draft rules make filtering mechanisms visible to the security services. Disclosure that runs only upwards to power is the opposite of what is proposed here. The test is whether the person refused can see why.
Matt Sheehan of the Carnegie Endowment observed of China’s AI rules that democracies can borrow the tools without the values: the registries and disclosure tools built for information control could serve democratic oversight instead. China already requires providers to maintain a should-refuse bank and a should-not-refuse bank, and measures both. A democratic version would publish both banks, invite challenges to both, and report the results.
XIV. The Strongest Objections
“Publishing refusal rules helps attackers.” Partly true, which is why the receipt names the rule and its source, not the detection method. Rules about what is forbidden are already largely public in usage policies; it is their application that is hidden. And the evidence of Section IV is that secrecy has not made the wall hold.
“Telling users which law caused a refusal helps censors, by normalising it.” The opposite is more likely. A censor prefers the refusal that looks like the model’s own judgement. Forcing a refusal to say “this answer is withheld under the law of country X” turns invisible censorship into visible censorship, which is precisely what censors avoid.
“A refusal is the lab’s protected speech, so none of this can be mandated.” In the United States, possibly, and that is a real constraint. But the properties proposed here are mostly about disclosure, not content, and disclosure duties generally face a lower constitutional bar than content mandates, though FIRE contests even the FTC’s disclosure approach. The FTC’s proposed statement already treats disclosure as the cure for undisclosed steering. Elsewhere, the Article 19 framework asks for exactly the legality and proportionality that visibility makes testable.
“Safe-completions make the refusal debate obsolete.” They make it harder. A clear refusal at least tells the user something was withheld. A partial answer that quietly omits the useful part is a veto with no visible edge. Visibility must cover modification as well as refusal.
“This is a cost on the responsible labs that the irresponsible will ignore.” True of every governance proposal, and the reason procurement matters. Governments, banks and hospitals buying AI can require refusal receipts as a condition of purchase tomorrow, without legislation. The OMB memo shows procurement can demand disclosure about refusal; it should demand that the disclosure reach the user.
“The real risk is catastrophe, not over-refusal.” Both are real, and this essay does not ask anyone to choose. The July incident shows they are the same problem: a veto held privately, turned down for the attacker and up for the defender, by the same industry, with no one outside able to see either decision.
XV. Eight Questions to Ask of Any AI That Says No
For boards, buyers, regulators and users, eight questions that turn the five properties into due diligence:
- When your system refuses or modifies a request, does it tell the user?
- Does each refusal identify its source: your policy, the deployer’s instruction, or a named law?
- Which laws, in which jurisdictions, currently cause your system to refuse, and is that list published?
- What are your refusal rates by domain, language and stated user identity, and who other than you has measured them?
- How does a user contest a refusal, who decides, and how long does it take?
- Does a verified professional exemption carry across all your products and interfaces?
- When did your refusal behaviour last change, and were users told?
- In agentic deployments, where does the “no” sit: in the model, or in a gate outside it that logs every decision?
XVI. Singapore’s Choice
Singapore has so far chosen to regulate harms rather than write refusal lists. Its election law bans realistic deepfakes of candidates, favourable or not. The Online Safety (Relief and Accountability) Act set up an Online Safety Commission operating from June 2026. When Grok produced sexualised imagery in early 2026, the government said IMDA was engaging the company, not issuing a refusal mandate. IMDA’s agentic governance framework asks organisations to bound risks, keep humans meaningfully accountable, put technical controls around agent actions and audit whether oversight remains effective. MAS’s industry-built SAFR framework puts a Disposition Engine between agents and the systems they act on.
That is the right instinct, and it leaves an opening. Singapore sits between the two largest model ecosystems, uses both, and speaks languages and dialects, Singlish among them, in which early evidence suggests refusal behaves unpredictably. It has a testing institution in AI Verify and a new accreditation programme for private AI testers. It could become the first jurisdiction to publish a refusal transparency standard: a common format for refusal receipts, a public benchmark of should-refuse and should-not-refuse prompts in local languages, and independent measurement of how imported models behave for Singapore’s users. It would not tell any model what to refuse. It would require every model to show what it refused, and why.
Coda: Show Your Work
A machine that cannot say no is dangerous. A machine whose no is written in secret, applied unevenly, moved without notice and answerable to no one is dangerous in a different way, and the second danger grows as the first is addressed. Every improvement in refusal, as Holland Michel argues, is also an improvement in the machinery of suppression. That is not a reason to abandon refusal. It is a reason to insist that the veto be held in the open.
So the question is not whether AI should be able to say no. It is who holds the veto, and whether they can be made to show their work. Ask for the receipt.
Library: sources and how closely each was read
“Read” means the page, abstract or official summary was read for this essay; “secondary” means figures were taken from a report of the source. Several 2026 items are preprints. Check primaries before quoting.
The prompting essay and the July 2026 incident
- Holland Michel, A., “We’re putting too much faith in AI’s ability to say no”, MIT Technology Review (9 Oct 2026). Read in full by the author. Anecdotes from his interviews (Adler, Röttger, Elstner, Gleave, McBain, the Canadian case, Wester, Mchangama, Frank, OpenAI on the UAE) are attributed to his reporting.
- Mother Jones, Tumbler Ridge investigation (Mark Follman, 24 Sep 2026), as cited by Holland Michel. Not read; paywalled. OpenAI has denied the accusations; the case remains under investigation.
- AI Chat Daily, “AI refusal: load-bearing wall of safety” (9 Oct 2026), summary of the above. Secondary report.
- Hugging Face, “Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident” (27 Jul 2026). Read.
- Willison, S., “OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened” (2026), quoting OpenAI’s disclosure. Read.
How refusal works, and why it is fragile
- Arditi, A., Nanda, N. et al., “Refusal in Language Models Is Mediated by a Single Direction”, NeurIPS 2024. Abstract read.
- Takehi, R., Teoh, J., Zhang, A. and Khattab, O. (MIT CSAIL), “Weight-Space Programming” (blog, Nov 2026), as reported by The Register. Coverage read; original blog not reachable.
- “Computation, not weight access: large language models as auditors of clinical prediction networks”, arXiv 2611.07847 (Nov 2026). Search extract read.
- Qi, X., Henderson, P. et al., “Safety Alignment Should Be Made More Than Just a Few Tokens Deep”, ICLR 2025. Abstract read.
- Qi, X. et al., “Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!”, ICLR 2024. Abstract read.
- Sokhansanj, B. A., “Uncensored AI in the Wild”, Future Internet 17(10) (2025). Preprint read.
- Volkov, D. (Palisade), “Badllama 3: removing safety finetuning from Llama 3 in minutes” (2024). Read; figures inconsistent within paper.
- Zou, A., Kolter, J. Z., Fredrikson, M., Carlini, N. et al., “Universal and Transferable Adversarial Attacks on Aligned Language Models” (2023). Abstract read.
- Anthropic (Anil et al.), “Many-shot jailbreaking” (Apr 2024). Read.
- Hughes, J. et al., “Best-of-N Jailbreaking”, NeurIPS 2025. Abstract read.
- Russinovich, M., Salem, A., Eldan, R., “The Crescendo Multi-Turn LLM Jailbreak Attack”, USENIX Security 2025. Read.
- Nasr, M., Carlini, N., Tramèr, F. et al., “The Attacker Moves Second” (Oct 2025). Abstract read.
- Rando, J., Zhang, J., Carlini, N., Tramèr, F., “Adversarial ML Problems Are Getting Harder to Solve and to Evaluate” (2025). Abstract read.
- Souly, A. et al., “A StrongREJECT for Empty Jailbreaks”, NeurIPS 2024. Abstract read.
- UK AI Security Institute, “5 key findings from our first Frontier AI Trends Report” (18 Dec 2025). Read.
- Anthropic, “Strengthening our safeguards through collaboration with US CAISI and UK AISI” (Sep 2025). Read.
- Anthropic, “Constitutional Classifiers” (Feb 2025) and “Next-generation Constitutional Classifiers” (Jan 2026). Read.
- Zou, A. et al., “Improving Alignment and Robustness with Circuit Breakers”, NeurIPS 2024; and Schwinn & Geisler, “Revisiting the Robust Alignment of Circuit Breakers” (2024). Abstracts read.
- International AI Safety Report 2026 (lead author Yoshua Bengio). Landing page and key update only.
Agents and refusal
- Kumar, P. et al., “Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents” (BrowserART, 2024). Abstract read.
- Andriushchenko, M. et al. (UK AISI, Gray Swan), “AgentHarm”, ICLR 2025. Read.
- Kuntz, T. et al., “OS-Harm”, NeurIPS 2025 Datasets and Benchmarks. Read.
- Gray Swan, “UK AISI x Gray Swan Agent Red-Teaming Challenge: Results Snapshot” (May 2025). Read.
- Debenedetti, E., Tramèr, F. et al., “AgentDojo”, NeurIPS 2024. Read.
- Debenedetti, E., Carlini, N., Tramèr, F. et al., “Defeating Prompt Injections by Design” (CaMeL, 2025). Read.
- Fortune, “OpenAI says AI browsers may always be vulnerable to prompt injection” (23 Dec 2025). Secondary report.
- Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign” (Nov 2025). Read.
- PC Gamer, sceptical practitioners on the above (Nov 2025). Read.
- Anthropic, “Detecting and countering misuse of AI: September 2026”. Read in part.
- Anthropic et al., “SHADE-Arena” (Jun 2025). Read.
The veto nobody granted
- Anthropic, Claude Opus 4 and Sonnet 4 System Card (May 2025), as quoted by Simon Willison. Secondary report.
- Anthropic, “Agentic Misalignment: How LLMs could be insider threats” (Jun 2025). Secondary report.
- Schlatter, Weinstein-Raun, Ladish (Palisade), “Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs” (2025). Abstract read; not peer-reviewed.
- OpenAI and Apollo Research, “Detecting and reducing scheming in AI models” (Sep 2025). Read.
- Anthropic, Claude’s Constitution (Jan 2026). Read via reproductions.
- Hadfield-Menell, D., Dragan, A., Abbeel, P., Russell, S., “The Off-Switch Game”, IJCAI 2017. Read.
Over-refusal: benchmarks and lab metrics
- Röttger, P., Hovy, D. et al., “XSTest”, NAACL 2024. Read.
- Cui, J. et al., “OR-Bench: An Over-Refusal Benchmark for LLMs”, ICML 2025. Read.
- Zhang, Z. et al., “FalseReject” (2025). Read.
- An, B. et al., “Automatic Pseudo-Harmful Prompt Generation for Evaluating False Refusals” (PHTest, 2024). Abstract read.
- Xie, T. et al., “SORRY-Bench”, ICLR 2025. Abstract read.
- AI2, “The Ai2 Safety Toolkit” (WildGuard, WildJailbreak, 2024). Read.
- Al Hasan and Biswas, “The Refusal-Compliance Tradeoff” (May 2026). Abstract read.
- Weidener et al., “RefusalBench” (May 2026). Read; small commercial lab, preprint.
- Ren, Nakov, Naseem, “DUAL-Bench” (2025-26). Read.
- Anthropic, Claude 3.7 Sonnet System Card (Feb 2025) and Claude Opus 4.5 System Card (Nov 2025). Read.
- The Register, “Claude Opus 4.7 classifier overzealous” (23 Apr 2026). Read.
- Yuan, Y. et al. (OpenAI), “From Hard Refusals to Safe-Completions” (Aug 2025). Read.
- OpenAI, GPT-4.5 System Card (Feb 2025) and GPT-5.5 System Card (Apr 2026). Read.
- eMarketer on Meta Llama 4 political refusals (Apr 2025). Secondary report.
- Raghavan, P., “Gemini image generation got it wrong”, Google (Feb 2024); Gemini 3.1 Pro Model Card (Feb 2026). Read.
- Bhatt, M. et al. (Meta), “CyberSecEval 2” (2024). Read.
Over-refusal: domains and disparities
- Zhang, Nakov, Naseem et al., “Health-ORSC-Bench” (Jan 2026). Summary read.
- Moriguchi, L., “A Referral Is Not an Abstention”, NeurIPS 2026 GenAI for Health workshop. Abstract read.
- McBain, R. et al. (RAND), chatbot responses to suicide questions, Psychiatric Services (Aug 2025). Read.
- Tang, N. et al., “Beyond the Single Turn”, FAccT 2026. Read.
- Wang, K. et al., “HRIPBench” (harm reduction, 2025). Read.
- Campbell, D., “When AI Safety Becomes a Denial-of-Service for Defenders”, Scale Labs (Mar 2026). Read; industry study.
- Jones, Fredrikson, Kolter, “A New Framework for Cybersecurity Refusals in AI Agents”, ICML 2026. Abstract read.
- FitzGerald et al. (EdgeRunner AI), “Measuring and Eliminating Refusals in Military LLMs” (2026). Read.
- Kucherenko et al., “LLMs Prompted for Legal Context Object More” (2026). Read.
- El Touny, K., “Silenced by Design” (2025-26). Read; single-author case study.
- Li, V., Chen, Y., Saphra, N., “ChatGPT Doesn’t Trust Chargers Fans”, EMNLP 2024. Read.
- Poole-Dayan, E., Kabbara, J., Roy, D. (MIT CCC), “LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users”, AAAI 2026. Read.
- Khorramrouz and Levy, “Characterizing Selective Refusal Bias in LLMs” (2025). Read.
- Plaza-del-Arco, Röttger, Hovy et al., “No for Some, Yes for Others” (2025). Read.
- Haq and Saldías, “Dialect vs Demographics” (Apr 2026). Abstract read.
- Im, Yuan, Färber, “Analyzing Bias in False Refusal Behavior of LLMs for Hate Speech Detoxification” (2026). Read.
- Belay, Mukhtar et al., “The Illusion of Cross-Lingual Safety in Low-Resource Languages” (Aug 2026). Summary read.
- Yu, Knox, Wang, “Understanding Content Moderation in LLMs through Restricted Books” (2026). Read.
- Dai, Lurie, Metaxa, Friedler, “Longitudinal Monitoring of LLM Content Moderation of Social Issues” (Oct 2025). Read.
- Oversight Board, “Are LLMs Stifling Political Speech?” (Jul 2026). Read.
- Hall, Westwood, Grimmer, “Measuring Perceived Slant in LLMs”, Stanford GSB (May 2025). Read.
- OpenAI, “Defining and evaluating political bias in LLMs” (Oct 2025); Anthropic, “Measuring political even-handedness” (Nov 2025). Read.
- Mchangama, J., Calvet-Bademunt, J., “That Violates My Policies”, The Future of Free Speech (Oct 2025). Read.
- Yadav, A. et al., “Are LLMs Good Safety Agents or a Propaganda Engine?” (2025). Read.
- Pasch, S., “LLM Content Moderation and User Satisfaction” (Jan 2025). Read.
- Nahar, Lee, Heo, Lee, “When AI Says ‘I Am Unable to Answer’” (2026). Read.
- Zheng, Sap et al., “Let Them Down Easy!”, Findings of EMNLP 2025. Abstract read.
States writing refusal
- Waight, H., Yang, E., Yuan, Y., Messing, S., Roberts, M. E., Stewart, B. M., Tucker, J. A., “State media control influences large language models”, Nature (13 May 2026). Abstract, companion site and release read.
- CrowdStrike Counter Adversary Operations, DeepSeek-R1 code-security research (2025). Research summaries read by the author; figures as reported.
- Wester, J. et al., “As an AI language model, I cannot: Investigating LLM Denials of User Requests”, CHI 2024. Via the author’s research.
- GB/T 45654-2025, Basic Safety Requirements for Generative AI Services, CSET translation (Jun 2025). Read.
- Guangming Daily on CAC filings to 30 Jun 2026. Read.
- WSJ on China’s chatbot ideology test (Dec 2025), via WebProNews. Secondary summary; WSJ not read.
- Sheehan, M., “China’s AI Regulations and How They Get Made”, Carnegie Endowment (Jul 2023). Read.
- Promptfoo, “1,156 Questions Censored by DeepSeek” (Jan 2025). Read.
- Reporters Without Borders, “Controlling information in the age of AI” (Sep 2025). Read.
- NewsGuard, Chinese AI models and pro-China claims (Jul 2025). Read.
- Pan, J., Xu, X., “Political censorship in large language models originating from China”, PNAS Nexus (Feb 2026). Release read.
- CAISI (NIST), “Evaluation of DeepSeek AI Models” (Sep 2025). Read.
- Naseh, A. et al., “R1dacted” (2025). Read.
- Ryan, F., Allen, B. et al., “The Party’s AI”, ASPI (Dec 2025). Read.
- Colville, A., “The Chinese Core of ‘Uganda’s ChatGPT’”, China Media Project (Dec 2025). Read.
- Wang, Y., “PRC State-Centric AI Governance Weakens Protection of Rights”, Jamestown China Brief (Mar 2026). Read.
- Kuteynikov, D., Popovets, L., “Russia’s AI law puts control ahead of capability”, Tech Policy Press (2026). Read.
- Hogan Lovells on India’s IT Rules amendments (Feb 2026). Read.
- BusinessDay/Reuters, Turkey blocks Grok posts (Jul 2025). Read.
- Albergotti, R., Semafor, OpenAI and G42 UAE ChatGPT variant (Feb 2026). Read.
- Freedom House, Freedom on the Net 2023 and 2026. Read.
Democracies and the veto
- Executive Order 14319, “Preventing Woke AI in the Federal Government” (Jul 2025). Read.
- OMB Memorandum M-26-04 (11 Dec 2025). Read.
- Kraus, J., Lawfare, on the OMB guidance (Dec 2025). Read.
- Arnold & Porter on the FTC’s proposed AI policy statement (Jul 2026). Read.
- Platformer on the Missouri attorney general’s letters (Jul 2025). Read.
- Reclaim The Net on xAI v. Weiser (Apr 2026). Secondary report.
- Anthropic, “Statement on the Department of War” (26 Feb 2026). Read.
- Reason on the March 2026 injunction; TechCrunch on the August 2026 judgment. Read.
- ABC News on the DC Circuit ruling (Sep 2026). Read; URL partly redacted in research.
- Built In on OpenAI’s Pentagon agreement (2026). Secondary report.
- Perkins Coie on California SB 243. Read.
- Fenwick on New York’s AI companion law. Read.
- Orrick, “2026 State Chatbot Laws” (Apr 2026). Secondary report.
- Wilson Sonsini on EU AI Act enforcement (Aug 2026). Read.
- Hunton on the DSA designation of ChatGPT (Sep 2026). Read.
- RPC on Ofcom’s chatbot explainer (Mar 2026); Lewis Silkin on UK reforms (Feb 2026). Read.
Speech, law and legitimacy
- WUSF/AP on Garcia v. Character Technologies (May 2025); AP on the settlement (Jan 2026). Read.
- FIRE, “AI speech is still speech” (Jun 2025); FIRE comment to the FTC (Jul 2026). Read.
- EFF, “Chatbot output can reflect human expression” (Aug 2025). Read.
- Volokh, E., Lemley, M., Henderson, P., “Freedom of Speech and AI Output”, Journal of Free Speech Law (2023). Read.
- Aswad, E., “A New Frontier for an International Right with No Frontiers”, Journal of Free Speech Law (2026). Introduction read.
- Balkin, J., “AI and the First Amendment”, Yale Law School Q&A (Jan 2024). Read.
- Center for Humane Technology, “AI is the next free speech battleground” with Lawrence Lessig (Jul 2025). Snippet or unverified; check before quoting.
- Volokh Conspiracy on Justice Barrett’s NetChoice concurrence (Jul 2024). Read.
- NetChoice on NRA v. Vullo (May 2024). Read.
- Branum, B., “Whose Speech Is It Anyway?”, CDT (Jul 2026). Read.
- Reuters via CP24 on Meta’s “GenAI: Content Risk Standards” (Aug 2025). Read.
- Engadget on Grok’s system-prompt incident (Jul 2025). Read.
- OpenAI, “Sharing the latest Model Spec” (Feb 2025). Read.
- “SpecEval: Evaluating Model Adherence to Behavior Specifications” (Sep 2025). Read.
- Calvet-Bademunt, J., usage policies and Article 19, ICML 2026 position paper. Abstract read.
- Anthropic and CIP, “Collective Constitutional AI” (Oct 2023). Read.
- Collective Intelligence Project, Global Dialogues. Read.
- OpenAI, “Democratic Inputs to AI” grant programme update (Jan 2024). Read.
- Abiri, G., “Corporations Constituting Intelligence”, California Law Review Online (Mar 2026). Read.
- Rest of World on the Oversight Board and generative AI (2026). Read.
- Fortune on Sam Altman and adult content (Oct 2025). Read.
Singapore
- Malay Mail on Singapore’s election deepfake law (Oct 2024). Read.
- WongPartnership on the Online Safety (Relief and Accountability) Act 2025. Read.
- MDDI reply to parliamentary question on Gen AI chatbot safeguards (Feb 2026). Read.
- IMDA, Model AI Governance Framework for Agentic AI (Jan 2026), via Allen & Gledhill.
- GovInsider on Singapore’s AI tester accreditation programme (2026). Read.
Luke Soon is an AI futurist and author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: SuperIntelligence Protocol. He co-authored IMDA Singapore’s Model Governance Framework for Agentic AI.


Leave a Reply