--- title: "Unverified hunches in AI-assisted software work: the certainty handoff and one-sided memory" kind: "Article" topic: "AI-directed development" date: "2026-10-09" url: "https://canopticon.com/articles/unverified-hunches-in-ai-assisted-software-work" author: "Chris Telford" license: "CC BY 4.0" --- You ask an AI coding agent to rename something because the name is wrong. You know why it is wrong — you have been living in this project for weeks — so you do not say. The agent renames it, guesses at the reason, and the guess spreads through the code. An hour later you are undoing work that was built on a hunch you never checked and never explained. This article names the two failures behind that hour. Neither is new on its own; what is new is how they compound when the other party is a machine that forgets.

- The handoff | The certainty handoff | Passing on a hunch that feels right as though it had been checked — and assuming the other side understood it as clearly as you hold it. | Builds on research into the feeling of rightness and the illusion of transparency. - The memory | One-sided memory | You remember the whole project; the agent remembers only this session; both act as if the ground were shared. | Researchers have begun calling this “asymmetric persistence”. ## The feeling that you are right is a separate signal The research behind this article starts from one finding: the feeling that an intuition is right is a separate signal from the intuition itself. The idea arrives, and a sense of certainty arrives with it. It is that sense, not the idea, that decides whether you stop to test the idea. Usually the feeling is a good guide, which is why trusting it feels rational. But it can be wrong in a way you cannot see from inside: 37% of incorrect solutions got Aha ratings higher than the average for correct ones. And the feeling does not track truth so much as fit: confidence tracks how *consistent* an idea is with what you already believe, not whether it is *correct*. > **Objection: Intuitions are usually right, so why not trust them?** > > They often are: Aha solutions are correct more often than analytic ones. That is the trap, not the defence. A signal that is right most of the time is exactly the one you stop checking. In the research’s words, The same signal that makes intuition useful also stops you from verifying it.

“believed they were understood 72% of the time, while listeners actually misunderstood them 39% of the time” The signal arrives before any thought about it. Brain recordings found a reward-system burst about 100 milliseconds after the insight itself, and the authors state it is too fast to be a conscious, post-solution appraisal. It is not the same for everyone: It also appeared only in people high in reward sensitivity. The research treats that as one study, so treat it as suggestive, not settled Insight also keeps arriving when the capacity to check it is used up. Under heavy mental load: Non-insight solutions became rarer and slower as load rose. By contrast: Insight solutions were largely unaffected. The research draws the consequence: The ratio of unchecked to checked conclusions rises exactly when you're in momentum. Part of what makes an insight feel like one is that it seems not to need checking. In one study, participants were told an Aha meant feeling "relatively confident that your solution is correct without having to check it once more… if you feel you still need to check it that would not be an Aha!" And for open-ended questions, where there is no single right answer, the feeling may mislead more: the feeling of insight inflated people's judgments of idea quality. ## The gate that decides whether you check Reasoning researchers have a name for the gatekeeper: the Feeling of Rightness. Across several reasoning tasks, a low FOR predicted longer rethinking and more answer changes. And it is driven by the wrong thing: FOR is driven largely by answer fluency, meaning how quickly and easily the answer came to mind, not by its validity. Medicine shows what that costs. In an analysis of 100 diagnostic errors: Premature closure, "the failure to continue considering reasonable alternatives after an initial diagnosis was reached," was the single most common cause. The errors were not ignorance: Faulty or inadequate knowledge was uncommon. A tested remedy exists, in which clinicians systematically recheck their findings against the first diagnosis and its alternatives; it has improved accuracy in several randomised experiments, especially on complex cases ## The certainty handoff: you think you were understood Handing the hunch to someone else adds a second, separate error. When people try to get a specific meaning across, they misjudge how well it landed. In the classic experiment, speakers believed they were understood 72% of the time, while listeners actually misunderstood them 39% of the time. The authors concluded that speakers "underestimate the ambiguity of their own utterances and overestimate the extent to which their disambiguating cues make their intention transparent."

caption: How often speakers are understood - What speakers believed | 72% | 72 | #2f55c4 - What actually happened | 61% | 61 | #c2603f note: Listeners misunderstood 39% of the time So the certainty handoff has two halves. You are too sure of the idea, and too sure you conveyed it. With an AI agent, the research puts it plainly: Your intention is transparent *to you* because you hold it. The agent gets only the words. Asking yourself to explain a mechanism step by step exposes the gap: people think they understand mechanisms far better than they do until they try to explain them step by step, at which point their self-ratings drop. The timing matters, though, because words can damage a pattern still forming. The research's conclusion is about where the check goes: let the intuition form, then articulate and test it before it becomes a directive to someone else, not while it's still forming. Researchers in human–computer interaction describe the same gap from the interface side. One team names a Gulf of Envisioning, the distance between a person's initial intentions and a prompt that can produce the desired output. Another finds that working with these tools imposes demands like those on "a manager delegating tasks to a team." The research is candid about the gap in the evidence: No study I found directly measures rework caused by unverified human intuitions given to coding agents. ## One-sided memory: only one of you remembers Between people, misunderstandings get caught because the listener asks. AI models mostly do not. Measured against human conversation, they were "three times less likely to initiate clarification and sixteen times less likely to provide follow-up requests than humans," and part of that comes from how they are trained: "training on contemporary preference data leads to a reduction in generated grounding acts." Now add memory. You have weeks of decisions in your head; the agent has whatever is in front of it this session. A 2026 paper calls this "the repeated re-establishment of a working ground under asymmetric persistence." The result is that neither side asks: the human presumes accumulated ground, and the model is trained to presume ground too, so neither party triggers repair. > **Objection: Write the context down and the problem goes away** > > Partly. Writing notes for the agent is now standard practice, but the evidence says notes alone are weak: the best controlled evidence shows that naive artifacts (prose context files) do not raise task success. What helps are the ones that constrain the work itself (tests, types, executable specs), not notes about it. > > The research offers a reading of the mixed evidence, flagged as its own synthesis: prose notes are a marker left beside the work, while tests and types are the work itself. Its prediction: stigmergic direction outperforms dialogic direction only when the traces are constraining rather than descriptive. One researcher's framing puts it in a line: "Technical debt lives in code. Cognitive debt lives in people. Intent debt lives in artifacts," People also misjudge what the agent remembers. A study of users' mental models of AI memory found "a widespread acknowledged lack of understanding", with uncertainty persisting "regardless of the user's technical background." The research points to an old idea that fits a fresh session exactly, the overhearer: every fresh agent session is structurally an overhearer of the project — it sees the repository (referential context) but not the history of negotiated meanings that produced it Experienced people seem to know this already. In a study of professionals delegating long tasks, they gave people a compass of high-level intent but gave AI "rigid, exhaustive instructions to minimize ambiguity and deviation," because of a perception "that current AI struggles to infer intent." The research's inference is that rework happens precisely when the relationship felt long enough to switch to compass mode ## Time pressure turns both failures up Both failures get worse when you are rushing, and directing an agent always feels like rushing. Across 102 studies of software teams, time pressure usually raises productivity but lowers quality, mainly because developers drop reviews and testing. The check that would catch an unverified hunch is the first thing to go. The numbers in that review are specific. 75% of practitioners named time pressure as the top obstacle to reviews. Defects followed the same curve: Expected defects rose 47.1% from the lowest to the highest time pressure. The cost persists: one longitudinal study found developers "waste, on average, 23% of their development time due to TD"— technical debt. The research cautions that much of this evidence comes from surveys and correlations. ## Why an expert’s gut check stops protecting them Experienced practitioners decide by recognising a familiar situation and imagining one workable option play out. The researcher who described this warned about it himself: "we become too confident in the [simulations] we construct". The method works because the person imagining the action is the person who takes it. Handing it to an agent breaks that: The developer simulates their implementation of the recognized pattern; the agent executes its interpretation. And the warning signs arrive late: With an agent, the violation surfaces only when the developer reviews the output — often after it has been built on — which is where rework cost compounds. The same research sees the remedy in the tools themselves: coding agents make external simulation cheap. A rough draft or a test run turns the imagined check into a real one before anything is built on it. ## What the two names explain They explain why rework with an AI agent so often traces back to something you knew and never said: the certainty handoff made you sure you had said it, and one-sided memory meant the agent had no way to know otherwise. They explain why more notes rarely fix it, and why tests do — a test is a check the agent cannot skip and you cannot forget to explain. And they put the fault in the right place. It is not that the agent is careless or that you are. It is that the feeling of being right travels easily, and the reasons for it do not.