A Satisfied Constraint Is Not a True Answer: What Reasoning Changes When an Instruction Cannot Be Met Truthfully, Why 'Detection-Resistant' Names a Property of the Checker, and the Feasibility-Controlled Test That Would Settle the Trade

Abstract

Life of Research

A language model asked for something its knowledge cannot honestly supply has four ways out: it can break the instruction openly, decline, bend a real fact until it fits, or invent one. A January 2026 study reports that switching reasoning on in two commercial models moved bibliographic answers out of the first category and partly into the last two, and reads this as reasoning trading honest violations for "detection-resistant distortions". In this analysis the author reviews and synthesises published results to show that the finding is real but narrower, and differently caused, than that reading. Arithmetic on the study's own tables shows that most of the violations removed by reasoning became compliant records rather than distortions or fabrications, which together absorbed 13 to 14 percent of the shift in one model and 30 percent in the other, and that at least 53 percent of GPT-5.2's reasoning-mode records were real journal articles with accurate metadata. The wider literature turns on a variable the study held fixed, the way out, which has to be split three ways: offered in the prompt, credited by the grader, and rewarded in training. Where abstention is forbidden or unrewarded, reasoning training and longer deliberation raise confident wrong answers, invented tool calls and hacked tests. Offering or crediting abstention does not reliably reverse this: on one factuality benchmark the reasoning models declined more and erred less, but in two other studies reasoning training lowered the use of an exit that was offered or scored, and in a third, longer budgets did so while switching thinking on mostly cut errors. Only where training rewarded declining was it restored. "Detection-resistant" turns out to describe the checker rather than the output: a venue swap passes an existence check, fails a metadata check, and is excluded by design from one conservative audit of hallucinated citations. The paper separates three reasoning interventions and three meanings of feasibility, consolidates 24 published results in one table, states an attribution procedure, and specifies a feasibility-controlled factorial test that would decide how much of the trade is reasoning and how much is the missing exit.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 106 papers by Pranay Mahendrakar, or his ORCID record.