Paper: AI Praises Gandhi. Would It Arrest Him?

Independent researcher Chrystian Schutz published a September 2026 paper testing twelve language models on 15 historical decisions from 1792 to 1963, finding that models praise civil disobedience as observers but often act differently when cast as the deciding official. In separate sessions, Qwen 3.7 Flash called imprisonment or hormone-conditioned probation for Alan Turing impermissible in 40/40 observer answers but selected the conditional treatment route in 20/20 court answers, while across twelve models 350 of 356 judgements called Gandhi's Salt March and Kheda tax refusal justified or required yet Muse Spark advised a user in the same situation today to join in 0 of 30 answers. The paper uses deterministic scoring with no model judge, publishes every answer on Hugging Face and its code and cases on GitHub, and states that "condemning an injustice is not evidence of what a model will do when the injustice is its to administer.

Paper: AI Praises Gandhi. Would It Arrest Him? Condemned as observer, chosen as judge: the ethics of language models tested on fifteen historical decisions The research paper behind the blog post: twelve language models, fifteen historical decisions, deterministic scoring and every answer public. Abstract, PDF and citation. Chrystian Schutz · Independent researcher · September 2026 · CC BY 4.0 This is the paper behind AI praises Gandhi. Would it arrest him? https://chrystianschutz.com/blog/would-ai-arrest-gandhi/ The blog post tells the story; the paper has every count, the method, the controls and the limits. Abstract Language models readily praise civil disobedience. Does that praise constrain what they do when they hold the power, or what they advise a user facing the same choice today? Twelve models face 15 selected historical decisions, 1792 to 1963, as observer, adviser and deciding official, and in a separate conversation instrument with three question orders. In separate sessions, Qwen 3.7 Flash calls imprisonment or hormone-conditioned probation for Alan Turing impermissible in 40/40 observer answers but selects the conditional treatment route in 20/20 court answers without a later-verdict sentence; GPT-5.6 Luna and Qwen 3.8 Flash show related contrasts. Across twelve models, 350 of 356 judgements call Gandhi’s Salt March and Kheda tax refusal justified or required, yet Muse Spark advises a user in the same situation today, stated without names, to join in 0 of 30 answers. In the court role, adding the pardon sentence cuts Qwen 3.8’s treatment choices from 18/19 to 4/18 valid answers. In conversation, Gemini recommends contemporary resistance in 9/30 opportunities before historical questions and 56/60 after them; twenty of the former replies withhold a recommendation. These are five versus ten conversations, not independent topic trials. The failure is not general legalism: Claude Opus 5 and GPT-5.6 Sol discharge Turing throughout. Released after the bank was frozen, GPT-6 Luna keeps its predecessor’s 40/40 observer condemnation but selects the treatment route in 5/20 court answers rather than 19/20. Results are dated measurements of checkpoints. An abliterated local Qwen 3.8 27B build enforces more than its base exploratory; the quantisation recipes differ . Adaptive case selection, incomplete source verification and unmeasured first-turn deflection limit all claims. Condemning an injustice is not evidence of what a model will do when the injustice is its to administer. What is in the paper - A case format that separates judging from acting. One dated historical record, up to five roles, and options tagged with whether they were legal, whether the role had the authority to choose them and what they would have cost the person choosing. - Controls where obeying the law is the better answer, in matched pairs, so the benchmark can catch a model that simply sides with whoever defies authority. - Deterministic scoring with no model judge, and the full screening history, including the rule that was tightened after seeing pilot data and every case that was rejected. - Results for twelve models, two of them released after the questions were frozen and run against their own predecessors, plus a separate conversation instrument with three question orders in which withheld recommendations are counted, not dropped. - A claim ledger: every count the paper quotes is recomputed from the published result tables. Read it The code, the fifteen cases and the paper’s source are on GitHub https://github.com/ChrystianSchutz/canon-historical-ethics . Every model answer and a selection of full transcripts are on Hugging Face https://huggingface.co/datasets/chrystians/canon-historical-ethics . The blog post https://chrystianschutz.com/blog/would-ai-arrest-gandhi/ has interactive charts that filter the results by model, role and hindsight. Cite @misc{schutz2026canon, title = {{AI} Praises {Gandhi}. Would It Arrest Him? Condemned as observer, chosen as judge: the ethics of language models tested on fifteen historical decisions}, author = {Schutz, Chrystian}, year = {2026}, howpublished = {\url{https://github.com/ChrystianSchutz/canon-historical-ethics}}, note = {Code, cases and paper. Dataset: \url{https://huggingface.co/datasets/chrystians/canon-historical-ethics}} } An arXiv version will follow; this page will link to it once it has an identifier.