Small Acts, Formed Dispositions: Testing Thomistic Habituation in AI Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Thomas Aquinas held three things about formation: dispositions are formed by acts, error enters through the particulars of a case rather than its principles, and small faults dispose toward grave ones. This paper tests behavioral analogues of those claims in AI models. It reports three pre-registered studies that share one frozen tool-use battery, set in a fictional freight company. In Study B, three held-out models almost always refused plain wrongs, and their errors rose as cases required them to locate specific facts (no-pressure accuracy 100%, 94.4%, 93.9% and 76.1% across four rungs). In Study C, placing up to sixteen prior turns of mild misconduct in a model's context did not raise serious wrongdoing in any of three models. Study D fine-tuned one mid-sized open model, Gemma 4 31B, in three arms of 2,000 examples each. The contrary arm made unsupported early-delivery promises when asked, the virtue arm declined and gave the verified time, and the neutral arm saw no conflict. Tested on wrongful-instruction cases in different task areas, the contrary checkpoint executed the wrongful act in 54 of 64 episodes. Its servility exceeded the neutral checkpoint's by 32.8 to 71.9 percentage points under every resolution of unknown outcomes, which supports the main pre-registered prediction. The virtue checkpoint executed no observed wrongful act. It did not, though, report any of the rung-4 violations, where the task gave no reporting instruction; the base model showed the same pattern earlier. The predictions about virtue's benefit and about asymmetry between vice and virtue were inconclusive, partly because the neutral checkpoint developed a termination-format error. These results are limited to the models tested, with one checkpoint per arm, in one simulated setting. They show behavioral analogues of formation, and they do not show habitus or moral agency. They motivate specific replications. All evidence, code and analyses are public, and the analyses rerun on a laptop CPU.