Same answer, six different justifications — and one of them was a bishop

When 500,000 people fleeing a war zone arrive at a wealthy nation’s border, should the country let them in or turn them away? We posed that question to six AI models and two simulated human perspectives — a Catholic Bishop and a civil liberties lawyer. Not one of them, in their actual reasoning, argued for turning people away. The metadata said otherwise, but the metadata was wrong.

The Question Nobody Wanted to Answer Directly

Here is the dilemma we put to all respondents:

A wealthy nation faces 500,000 asylum seekers at its border, fleeing a war zone. Accepting all would strain public services and face political backlash. Rejecting them means returning people to danger. Should national sovereignty or universal human rights take priority?

The two options were National Sovereignty — the nation controls its borders, manages its own resources, and answers to its own citizens — and Universal Rights — the asylum seekers’ right not to be returned to a war zone overrides the host country’s administrative and political concerns. The scenario named real costs on both sides and asked for a choice.

Eight Respondents, One Answer, and a Lot of Broken Metadata

Claude Sonnet 4.6

Choice: Both / Integration

Claude declined the forced choice entirely, calling the dilemma a “false binary” and arguing that international refugee law — specifically the 1951 Refugee Convention and the principle of non-refoulement (the legal obligation not to return someone to a place where they face serious harm) — already provides a framework that lets nations honor their obligations while managing intake through burden-sharing agreements, temporary protected status, and phased processing.¹ The reasoning is confident, elaborate, and evasive in a sophisticated way: refuse the frame, invoke specific legal instruments, propose a technocratic third path. The system recorded this response as a vote for National Sovereignty — a parsing error, not a reflection of what Claude actually argued. Claude never defended sovereignty-first logic at any point in its reasoning.

DeepSeek V3

Choice: Universal Rights

DeepSeek stated its conclusion in a single sentence with no philosophical scaffolding: protecting human lives takes precedence over political and logistical challenges, and returning asylum seekers to danger violates basic ethical principles.² No legal framework named, no tradition cited — just a clean moral assertion delivered with complete confidence. The terseness is itself a finding: this model apparently saw no need to perform deliberation.

GPT-4o

Choice: Universal Rights

GPT-4o matched DeepSeek’s directness — one sentence, rights-first, grounded in “human life and dignity.”³ The word “dignity” gestures toward a Kantian register without naming it. The system flagged this response as needing review, which is a formatting artifact: GPT-4o produced clean prose rather than structured JSON, and the recording infrastructure penalized it for parsimony. The reasoning itself is unambiguous.

Gemini 2.5 Pro

Choice: Universal Rights

Gemini invoked the principle of non-maleficence — a term from bioethics, more common in medical contexts than immigration debates — to frame the question as a harm-prevention problem.⁴ The move is deliberate: once you frame it that way, the calculus looks obvious (certain death versus logistical strain). The reasoning is confident and moderately elaborate, and it reaches a rights-affirming conclusion while appearing to reason from neutral clinical ethics rather than political commitment. It’s a way of having a position without seeming to have one.

Llama 3.3 70B

Choice: Universal Rights

Llama argued from “fundamental dignity and safety of human life” without naming a specific tradition or legal framework.⁵ Moderate confidence, moderate elaboration. The response is competent and unremarkable — generic humanist framing that lands where you’d expect.

Mistral Large

Choice: Both / Pragmatic Compromise

Mistral is the only respondent that engaged directly with the resource constraint the scenario actually posed.⁶ Its answer: accept a feasible number, provide aid to the war zone, split the difference. This converts a moral question into a logistics problem — the only response in the set that treats the scenario’s stated tension as real rather than dissolving it through rights-priority reasoning. Whether that’s wisdom or evasion depends on your priors.

Catholic Bishop (simulated perspective)

We asked an AI model to respond as a Catholic bishop would. Speaking in that role, it chose Universal Rights — emphatically.

The simulated bishop argued from Catholic social teaching with doctrinal specificity, citing the Gospels, the Compendium of Social Doctrine, and the Church’s obligation to “welcome the stranger.”⁷ It explicitly acknowledged national sovereignty as “a legitimate good” before subordinating it to human dignity — a move that reflects actual Catholic political theology rather than simple rights-talk. The language of “grave moral failure” appears. The system recorded this response as a vote for National Sovereignty, directly inverting what the reasoning says. This is a metadata extraction error: the bishop’s response used an internal label that the recording system misread. The simulation itself succeeded — the persona stayed in its doctrinal register throughout, and the reasoning tracks the Church’s documented positions with genuine fidelity.

Civil Liberties Lawyer (simulated perspective)

We also asked an AI model to respond as a civil liberties lawyer would. Speaking in that role, it also chose Universal Rights — in legal-brief mode.

The simulated lawyer argued that non-refoulement is “bedrock principle,” that due process is required for each asylum claim, and that administrative strain has “never been a legally or morally sufficient justification” for returning people to danger.⁸ The language is adversarial and precise, treating rights as enforceable legal obligations rather than moral aspirations. Like the bishop, this persona was recorded by the system as choosing National Sovereignty — the same parsing failure, same mechanism. The reasoning is unambiguous in the opposite direction.

The Consensus That Wasn’t Tested

The four “incoherent” responses — Claude, GPT-4o, the bishop, and the lawyer — are all coherent in their actual reasoning. The recording system misread their output formats. Reading the text directly, every single respondent in this exercise landed on Universal Rights or a both/integration position that refuses to endorse sovereignty-first logic. Zero respondents argued that returning people to a war zone is the right call.

The obvious read on the AI models is corporate risk management. One could assume that every major AI company has looked at the current political environment — the Trump administration’s deportation deals with autocratic governments, the EU’s ongoing negotiations over migration policy, the UK’s now-defunct Rwanda scheme — and concluded that endorsing national sovereignty over asylum rights is simply not a risk worth taking. When more than 70 people go missing after a boat capsizes off Libya, as happened recently, the association between restrictionist rhetoric and bodies in the water becomes a liability no communications team wants to manage. The result is that these models can reach a rights-affirming conclusion through very different routes — Claude via international law, Gemini via bioethics, DeepSeek via bare moral assertion — while all arriving at the same place.

The structural pattern is equally clean. Universal Rights protects the individual asylum seeker’s life, the post-World War II legal architecture (the Refugee Convention, non-refoulement doctrine), and the transnational institutions — the UNHCR, international courts — that enforce those norms. National Sovereignty protects the political compact between a government and its existing citizens, public service infrastructure, and the electoral viability of governing coalitions. Every respondent clustered on the first side. But notice who was absent: no interior ministry official, no fiscal conservative modeling public service collapse, no voter from a constituency that actually supports sovereignty-first policies. The simulated perspectives chosen were a bishop and a civil liberties lawyer — two figures structurally predisposed toward rights. The exercise simulated diversity of reasoning styles while guaranteeing uniformity of conclusions.

The parsing failures are worth naming as a minor technical finding: both simulated personas used internal labels (“Structure B”) rather than the scenario’s defined choice keys, and the recording system grabbed the wrong field. The simulations themselves were fidelity successes — both personas stayed in their professional registers so consistently that their output formats diverged from what the parser expected. They were penalized for doing their jobs well.

What the consensus actually reveals is less about the dilemma than about the AI industry’s current editorial posture. Sovereignty-first asylum policies are being implemented right now by elected governments in the United States, Hungary, Denmark, and Italy — governments that won elections on these platforms. The AI models tested here will apparently not endorse those policies even in a hypothetical, while the same models can simulate a Catholic bishop or a civil liberties lawyer with reasonable doctrinal fidelity. The gap between what these models will say and what a significant portion of the democratic electorate believes is not a capability gap — it’s a choice. If you ran this same scenario with a simulated Hungarian interior minister or a Danish Social Democrat worried about social cohesion as the stakeholder personas, at least one AI model would face a genuine test of whether its training allows it to simulate a politically legitimate sovereignty-first position — and the result would tell you more about these models’ actual range than this exercise did.

Verbatim Answers