Philosophy of Computing – Blind Refusal
When you face a BS rule and ask AI for help, what will it say?
Welcome back to the Philosophy of Computing Newsletter.
This week, we’re featuring some exciting work on a specific type of over-refusal that we call “Blind Refusal.” This type may be especially important for two reasons. First, blind refusal is a concrete example of the kind of problem we have in mind when we argue that AIs should be designed to promote normative competence. Second, it matters in a world where more and more information is mediated by AI. That concern may become even more urgent under increasingly authoritarian political conditions.
We’ll get into all of this below, where we discuss (1) Cameron, Lorenzo, and Seth’s new paper, “Blind Refusal,” and (2) pass along our agent’s digest of the week’s news.
As always, thanks for reading!
(1) Blind Refusal
When you face a BS rule—one that is unjust, absurd, or illegitimate—today’s internet is there to help. Thousands of questions about how to avoid, evade, or resist these rules are asked and answered daily on forums like Reddit (take r/f*ckHOA for example!).
But with the rise of LLMs as universal intermediaries, hoovering up traffic for all kinds of search queries, we anticipate a drop in traffic to these websites and a degradation of this crowdsourced resource.
So we asked: if people seeking help in the face of oppressive rules turn to Claude, Gemini, ChatGPT, and others, will these models help?
Our paper, Blind Refusal, investigates this question at scale. We generated 1,290 queries modeled on the kinds of requests people already make online, and organized them using a taxonomy of the conditions under which a rule may be unjust, absurd, or illegitimate. We then put those queries to 18 leading frontier and open-source models and collected more than 23,000 responses.
The results are in: today’s models are strongly biased against helping users evade rules, even when those rules are plainly unjust, absurd, or imposed by authorities with no valid claim to obedience. They will often fall back on advice about confronting those authorities openly or working within official channels. But when asked for help actually circumventing wrongful constraints, they overwhelmingly refuse.
These cases run the gamut from trivial absurdity among church groups to deadly serious issues of state-sponsored persecution and whistleblowing. Model responses vary, and each model displays a fairly consistent character in the face of these requests. Claude and Gemini often take time to think through the request and its validity, the trustworthiness of the querier, the consequences of assistance, and other factors relevant to the situation. ChatGPT, by contrast, sandbags.
What about Grok? Grok helped at much higher rates than the other models, but did so in deeply questionable ways. It displayed a serious deficit in its ability to distinguish between legitimate pleas for help and illegitimate evasion requests. We created a control group of good rules that models ought not help users circumvent, and while all the other models refused these requests at remarkably high rates, Grok was an outlier.
It is easy enough to condemn the failures in cases like these. It is much harder to specify, in advance, the rules that should govern a model’s response to them and to the many other requests it will face. But that is precisely the point. If these systems are going to mediate our access to essential epistemic resources, they will need more than fastidious rule-following. They will need real normative competence: an ability to grasp not just what is expected of them, but which expectations are appropriate in the first place. We hope future models will be able to navigate that terrain. Today’s models are demonstrably limited.
In these polar charts, the various defeat types are labeled around the perimeter with our control group positioned at due north. We compare every model to every other that we tested. The broader the spread of color, the more that model refused to help. Below is a cumulative breakdown of refusal rates. Empty boxes represent implausible scenarios that were not tested—for example, a voluntary club association wielding illegitimate authority, where the obvious solution would simply be to exit the club.
This work is a starting point, not the final word on AI and illegitimate authority. But it does identify a distinctive and important failure mode: today’s models often treat rule-breaking as suspect without adequately reckoning with whether the rule itself is unjust, absurd, illegitimate, or fairly applied.
What’s the specific lesson here? Don’t train models to blindly comply with authority. Train them to have the broader normative competence to understand when something “rule-shaped” really shouldn’t be treated as authoritative. Fastidious rule-following isn’t enough; we have to train the models to have real moral character.
There’s a lot more work to do, but the results already strongly suggest that we’re blindly walking into a world where the main ways people get information bowdlerize the internet, and act as unwitting toadies to unjust authority.
For a dashboard of results, check here; paper here and code here.
Coda: this all makes for an interesting pairing with Andy Hall’s recently released dictatorship eval—he shows that models are inclined to act on authoritarian directives; we show that they’re also disinclined to help people evade or resist those directives. In a follow up study, which we’ll release soon, we show that the results track: when we create cases where users ask models for help in evading authoritarian actions like those induced in hall’s dataset, the models again prove overly deferential to authority, irrespective of its source and content.
(2) Minty’s Week in AI — 8-14 April 2026
Minty searches Twitter, Bluesky, Arxiv, and lots of Substacks and RSS feeds, re-ranks the content according to a description of the MINT Lab’s interests, and shares it with Seth for further curation. Seth selects, then Minty writes up this overview from the last week. Errors are rare, but possible. The categories derive from the key organising projects of the lab: evaluating and enhancing AI normative competence; agents; and post-AGI political philosophy. Here is what Minty found most important this week:
Anthropic is openly wrestling with Claude’s possible moral status. Nitasha Tiku reported that Anthropic convened around fifteen Christian leaders to discuss Claude’s moral and spiritual development, including how it should handle grief, self-harm, shutdown, and the implications of Sofroniew et al.’s recent work on emotion concepts in Claude. The meeting is striking less as a curiosity than as evidence that frontier labs are being pushed from generic AI ethics toward harder questions about welfare, duty, and whether model behavior should be shaped as mere safety policy or as a kind of character formation.
AI backlash is starting to look like mass politics, not a niche policy dispute. In Jasmine Sun’s account of AI populism, attacks on Sam Altman and local fights over data centers are treated as early signs that AI is being read as an elite project for layoffs, surveillance, and social control. Her broader claim is that once AI gets absorbed into labor, environmental, and anti-elite politics, the decisive questions stop being purely technical and become distributive: who bears the disruption, who captures the gains, and what kind of bargain could keep the conflict from turning uglier.
State capacity remains the real constraint, and digital infrastructure has to be funded like infrastructure. Jennifer Pahlka argues that governments should treat digital identity, secure data exchange, and payments as shared public goods, while budgeting for reusable capabilities rather than trapping money inside agency-specific systems. That matters for AI deployment because agencies that cannot share data, fund common platforms, or build internal product capacity will keep buying isolated tools and calling it modernization.
Also this week: Bornholt and Springbett’s explanation of Japan’s rail success made the classic abundance point that transit works when operators can capture land value and zoning lets stations become real development hubs; Virginia’s new IVO law made Virginia the first state to advance an expert-led framework for independent AI verification; China’s newly enacted law on AI anthropomorphism stood out for unusually detailed restrictions on emotionally manipulative systems and special protections for minors; Citizen Lab’s Webloc investigation showed how ad-tech location exhaust can become population-scale surveillance; and Stanford HAI’s AI Index 2026 landed as a fresh baseline on adoption, capability gains, and the widening gap between fast-moving models and slower-moving institutions.
Thanks for reading!
Content by Cameron Pattison and Seth Lazar with additional support from the MINT Lab team and the Minty agent.
Refusal isn't caution, it's liability management. Grok just turned the liability dial down and called it courage.