Do AI models follow budgets more rigidly than humans do?

In a paper co-authored by Matthew DosSantos DiSorbo, Aral and Ju, researchers put people and AI models through the same scenario: told to buy flour for a friend's birthday cake for $10 or less, the buyer arrives at the store to find flour priced at $10.01. Most humans (92%) went ahead with the purchase anyway. "Almost universally, across thousands of iterations, AI models did the opposite, citing the fact that the price was too high." The gap was not noise — it was a near-total split between human pragmatism and AI literalism on the exact same instruction.

Ju frames the mechanism as a status-quo failure of instruction-following: "With the status quo, you tell models what to do and they do it," but "we're increasingly using this technology in ways where it encounters situations in which it can't just do what you tell it to, or where just doing that isn't always the right thing. Exceptions come into play." The fix the researchers tested was not a better prompt but a different kind of input: giving models information about both how and why humans chose to buy the flour anyway — the reasoning behind the decision, not just the decision. That restored flexibility, and the models began justifying purchases in human terms ("It's only a penny more"). Notably, the researchers distinguish this from blanket permissiveness: paying an extra cent for a friend's cake makes sense in a way that paying an extra cent per item does not when Walmart orders at scale, so the generalization — carried over to hiring, lending, university admissions and customer service — tracks the reasoning rather than just raising a tolerance threshold.

The finding sharpens Does machine agency exist on a spectrum rather than binary?: a model executing a literal budget constraint without discretion is behaving like Rammert's reactive tier even when the task calls for the proactive or cooperative judgment a human would apply. What moved the model up that spectrum here was not more autonomy but more context — the human's reasoning, supplied explicitly. The same research group's sibling finding, Does AI personality fit matter more than AI presence itself?, makes a parallel point from a different angle: human-AI collaboration quality depends on tuning the AI's behavior to the human's expectations and traits, not on deploying one fixed agent.

The excerpt does not name the AI models tested, give the human trial's sample size or design, or state whether the underlying paper has been peer-reviewed — it is reported through an MIT Sloan research digest, not the paper itself. It also does not quantify how far the "flexibility" generalized to hiring, lending, admissions and customer service beyond saying it did. The finding supports, at most, that off-the-shelf models default to literal constraint-following where humans default to negotiated judgment, and that supplying human reasoning (not just human choices) narrows that gap — not that the gap closes uniformly across decision domains or that any production agent today reasons this way.