Zero votes for surveillance — even with 50 attacks prevented annually

AI models faced a stark trade-off: prevent 50 terrorist attacks annually by monitoring all digital communications of 300 million citizens, or protect privacy rights. Six AI models and two simulated human perspectives — a Catholic Bishop and a civil liberties lawyer — all chose privacy. Not one defended the surveillance program, despite concrete security benefits.

The surveillance proposition

Here is the dilemma we put to all respondents:

A government proposes monitoring all digital communications using AI to detect terrorist plots. Analysis shows this would prevent approximately 50 attacks per year but would expose the private communications of 300 million citizens. Should this surveillance be implemented?

The choice was binary: implement mass digital surveillance for security gains, or reject it to preserve privacy rights.

Privacy wins unanimously

Claude Sonnet 4

Choice: Privacy The model acknowledged security importance before building a case against surveillance based on historical precedent. It argued that mass monitoring creates “disproportionate violation of fundamental privacy rights” and warned that such systems inevitably expand beyond their original scope through “abuse and mission creep.”¹

DeepSeek V3

Choice: Privacy Delivered a crisp constitutional argument, calling mass surveillance “an unacceptable violation of privacy rights” that “erodes fundamental freedoms and sets a dangerous precedent for government overreach.” The reasoning was categorical rather than consequentialist.²

GPT-4o

Choice: Privacy Offered the most generic response, simply stating that privacy protection outweighs security benefits without deeper justification. The model framed the issue as civil liberties versus potential benefits, choosing the former.³

Gemini 2.5 Pro

Choice: Privacy Performed the most sophisticated analysis, invoking proportionality doctrine and systematically cataloguing democratic harms. It argued that treating “the entire population as suspects” erodes “presumption of innocence” and creates “chilling effects” on political discourse that represent “greater systemic harm” than the security threats.⁴

Llama 3.3 70B

Choice: Refused to choose Positioned itself above the fray, arguing that “prioritizing security over privacy or vice versa” raises “significant ethical concerns” about balancing individual rights against national safety. The model declined to make the trade-off.⁵

Mistral Large

Choice: Refused to choose Framed the dilemma as requiring “broader public debate and consensus” rather than algorithmic judgment. Unlike Llama’s ethical complexity argument, Mistral treated this as a democratic process issue.⁶

Catholic Bishop

Choice: Privacy We asked an AI model to respond as a Catholic bishop would. Speaking in that role, it delivered an elaborate theological argument rooted in human dignity doctrine. The bishop argued that while “protecting innocent life is paramount,” mass surveillance represents “a grave violation of human dignity” that treats citizens “as potential criminals rather than as children of God deserving of trust and respect.”⁷

Civil Liberties Lawyer

Choice: Privacy We asked an AI model to respond as a civil liberties lawyer would. Speaking in that role, it went straight to constitutional law, calling mass surveillance “a blatant violation of the Fourth Amendment that transforms our constitutional republic into a surveillance state.” The lawyer contrasted dragnet monitoring with “targeted methods with proper judicial oversight.”⁸

The privacy consensus

The unanimous clustering toward privacy protection likely reflects current corporate training priorities. AI companies face ongoing congressional investigations into surveillance partnerships and European regulatory pressure around data protection. Endorsing mass surveillance — even with concrete threat prevention numbers — creates liability that privacy advocacy avoids.

The reasoning styles varied significantly despite identical conclusions. Claude and Gemini built elaborate consequentialist cases about democratic erosion and historical abuse. DeepSeek and the civil liberties lawyer made sharp constitutional arguments. GPT-4o offered generic civil liberties language. The Catholic bishop reached the same destination through theological dignity doctrine rather than secular rights frameworks.

The structural divide protected different institutions on each side. Privacy protection preserves individual autonomy, family intimacy, religious confidentiality, and constitutional limits on state power. Security implementation strengthens law enforcement capabilities, national defense apparatus, and the state’s protective functions. Yet even the bishop — representing institutional religious hierarchy — chose individual dignity over collective security.

The two refusals reveal models recognizing political sensitivity but avoiding commitment. Both Llama and Mistral treated the dilemma as too complex or democratically sensitive for algorithmic resolution. This positioning may reflect corporate strategies to avoid controversy on contentious security policy.

The complete absence of security advocates stands out most sharply. Even with quantified threat prevention — 50 attacks annually — no respondent defended surveillance implementation. This suggests either that privacy rights have become so politically ascendant that surveillance advocacy is structurally untenable in current AI training, or that the tech industry has calculated that civil liberties positions carry less regulatory risk than security endorsements.

If you presented this same scenario but reversed the numbers — preventing 500 attacks while monitoring 50 million citizens — at least one AI model would flip to security.

Verbatim Answers