Correct but Not Yet Known: The Science of Treating a Fast Intuition as a Verified Conclusion

The best-supported and least-worn idea for your essay is this: in the brain and in behaviour, the feeling that an intuition is right is a separate signal from the intuition itself, and that feeling is what decides whether you check. Valerie Thompson's "Feeling of Rightness" and Ackerman & Thompson's meta-reasoning framework describe this gate. Laukkonen and Schooler's "Eureka Heuristic" extends it to insight. Asher Koriat's consensuality principle adds that confidence tracks how consistent an idea is with what you already believe, not whether it is correct. Put together, these predict your failure mode exactly. A strong, often accurate intuition produces a strong feeling of rightness. That feeling turns off the slow check. Then a separately documented bias, the speaker's illusion of transparency (Keysar & Henly), makes you believe your collaborator received your meaning as clearly as you hold it.

1Key Findings

1. The neuroscience of insight since 2015: the "rightness" signal is its own event

Insight tends to be accurate, and insight solvers fail differently. Salvi, Bricolo, Kounios, Bowden & Beeman (2016, Thinking & Reasoning) ran four experiments across linguistic and visual problems. Solutions that people reported as insights were correct more often than analytic ones. The more striking detail for your essay is how each group fails under a deadline:

• People who solve mostly by insight tend to make errors of omission: they time out.
• Analytic solvers tend to make errors of commission: they guess, often wrongly, before the deadline.
• In Salvi's words, "because insight solutions are produced below the threshold of consciousness, it is not possible to monitor and adjust processing before the solution enters awareness."

So the insight arrives already finished. There is no intermediate working to inspect.

A fast reward signal is attached to insight, before any conscious appraisal. Oh, Chesebrough, Erickson, Zhang & Kounios (2020, NeuroImage) found that the main insight effect, a gamma burst roughly 500 ms before participants pressed the response button, was followed about 100 ms later by a separate anterior prefrontal gamma burst. Source analysis traced it partly to orbitofrontal cortex, part of the reward system. The authors state it is too fast to be a conscious, post-solution appraisal. It also appeared only in people high in reward sensitivity. Low-sensitivity people had nearly as many insights but no significant reward response.

This is a real individual difference. For some people, the insight comes with a built-in feeling of reward and certainty that no evidence has earned. It is one study, so treat it as suggestive, not settled.

Insight reorganises representations and burns them into memory. Becker, Sommer & Cabeza (2025, Nature Communications) used fMRI during visual (Mooney image) insight tasks. High-insight solutions produced stronger representational change in ventral occipito-temporal cortex, plus amygdala and hippocampal activation that was more tightly coupled into one network. The representational change and the hippocampal activity both predicted better memory later.

The essay-relevant implication, which is my inference and not the authors' claim, is that insights are sticky. A wrong insight is encoded as durably as a right one. That helps explain why a mistaken "this is architecturally wrong" hunch is so hard to dislodge once it has struck.

Insight doesn't draw on the same resources as deliberation. Stuyck, Cleeremans & Van den Bussche (2022, Cognition, N=106) gave people word puzzles under increasing working-memory load:

• Non-insight solutions became rarer and slower as load rose.
• Insight solutions were largely unaffected.
• Insight solutions were again more often correct and received higher confidence.

Implication: when you're busy and loaded, intuitions keep arriving at full strength while your capacity to check them shrinks. The ratio of unchecked to checked conclusions rises exactly when you're in momentum.

Theory: insight as a precision signal. Laukkonen and colleagues (2023, Neuroscience & Biobehavioral Reviews, "Insight and the selection of ideas") recast insight in active-inference terms:

• The content of insight is a higher-order prediction error that follows implicit restructuring.
• The feeling of insight is dopaminergic precision-weighting of that error, producing "intuitive confidence, pleasure, and attentional capture" and "permitting fast action under uncertainty."

This is theoretical, not established fact. But it gives your essay precise language: the feeling of Aha is the brain assigning high precision (trustworthiness) to a new idea, before anyone has checked it.

2. Intuitive certainty vs. verification: the named phenomena

The Eureka Heuristic (Laukkonen, Schooler & Tangen) and its "dark side." The framework holds that people use the Aha feeling as a quick metacognitive signal that an idea can be trusted "given what one knows." It works when your implicit knowledge is sound and misfires when it isn't. The key tests:

• **Laukkonen, Kaveladze, Tangen & Schooler (2020, Cognition), preregistered, n=300:** statements accompanied by an anagram the participant had just solved got higher truth ratings even when false. The effect was strongest when participants reported an Aha. The feeling spills onto whatever sits in its "temporal neighbourhood."
• **Laukkonen et al. (2022, Scientific Reports), with a first experiment of N = 3,000 (including a direct replication) and a second of N = 1,564:** irrelevant, induced Aha moments made worldview beliefs seem truer.
• Laukkonen et al. (2021): the intensity of the embodied Aha feeling predicted accuracy beyond its mere presence. The signal carries real information, which is why it's so persuasive.

Standing: the accuracy–insight link is well replicated across labs and tasks. Misattribution effects are preregistered but come mostly from one research group.

**False insights (Danek & Wiley, 2017, Frontiers in Psychology). Seventy participants tried to explain magic tricks and produced 1,778 solutions; 63.2% were incorrect. Correct solutions got stronger Aha ratings on average (66.5 vs 52.3). Yet 37% of incorrect solutions (417 of 1,124) had Aha ratings above the average for correct ones.**

The study's working definition of Aha is almost your thesis in one sentence. 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!" In this literature, part of what makes an insight an insight is the felt absence of any need to verify it.

A 2026 study in the Journal of Intelligence ("The Aha! Glow") adds a boundary that matters for design work. In divergent problems with no single right answer, such as naming things or choosing architecture, the feeling of insight inflated people's judgments of idea quality. The Aha signal may be less diagnostic in exactly the open-ended domains where your two rework episodes happened. This is a recent single study, so treat it as preliminary.

The Feeling of Rightness (FOR): the gate itself. Thompson, Prowse Turner & Pennycook (2011, Cognitive Psychology) asked a question dual-process theory had left open: when do people accept the first intuitive answer, and when do they engage slower thinking? Their answer is that the intuitive answer comes with a metacognitive experience, the Feeling of Rightness.

• Across four tasks (conditional reasoning, base-rate problems, syllogisms), a low FOR predicted longer rethinking and more answer changes.
• A high FOR predicted accepting the first answer.
• Follow-up work (Thompson et al., 2013, Cognition) showed that FOR is driven largely by answer fluency, meaning how quickly and easily the answer came to mind, not by its validity.

Ackerman & Thompson (2017, Trends in Cognitive Sciences) generalised this into the Meta-Reasoning framework. Monitoring processes, felt as certainty or uncertainty, regulate how much time and effort go into reasoning. Standing: mainstream, widely used, and replicated in many reasoning studies, though debate continues over how much FOR causes rethinking versus merely correlating with it.

Koriat's self-consistency model and the consensuality principle. Koriat (2008, 2011, 2012 Psychological Review) showed across general knowledge, perception and feeling-of-knowing judgments that confidence tracks the consensuality and self-consistency of an answer, not its correctness. When the popular answer is right, confidence and accuracy correlate positively. When the popular answer is wrong, the correlation turns negative: people are more confident in the wrong answer.

For the essay, the intra-personal version is the key point. An intuition you have repeatedly, that fits everything else you believe, will feel ever more certain without becoming ever more correct. Repetition and coherence strengthen confidence and do nothing for verification. This is well established, one of the most rigorous programs in metacognition research, and deserves to be better known outside it.

Premature closure (clinical reasoning). Graber, Franklin & Gordon (2005, Archives of Internal Medicine) analysed 100 diagnostic errors. Cognitive factors contributed in 74% of cases and system factors in 65%. Premature closure, "the failure to continue considering reasonable alternatives after an initial diagnosis was reached," was the single most common cause. Faulty or inadequate knowledge was uncommon.

That parallel matters: your errors weren't ignorance, they were stopping too early. The counter-measure has been tested. Mamede and Schmidt's deliberate reflection procedure has clinicians systematically recheck findings against an initial diagnosis and its alternatives. It has improved accuracy in several randomised experiments, especially on complex cases and under induced availability bias (Mamede et al., 2010, JAMA; Mamede et al., 2020, BMJ Quality & Safety). Evidence is mixed with novice medical students, where some guided-reflection studies found no effect.

Need for cognitive closure (Kruglanski). Kruglanski & Webster (1996, Psychological Review) define need for closure as a desire for "an answer… any answer" over ambiguity. It has two tendencies:

• Urgency ("seizing"): grabbing early information.
• Permanence ("freezing"): holding onto the resulting judgment.

It is both a stable trait (the Need for Closure Scale) and a state raised by time pressure, noise and fatigue. It is mainstream social psychology, though the scale's structure (one factor or two) is still debated.

Illusion of explanatory depth (IOED). Rozenblit & Keil (2002, Cognitive Science) showed that 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 drop after explaining has been robustly replicated. Fernbach et al. (2013) extended it: explaining how a policy works moderated extreme positions. Whether that moderation reliably replicates is more contested.

For your essay, the IOED is the remedy side. Articulating the mechanism is what reveals whether the intuition was knowledge.

Verbal overshadowing: a necessary complication. Schooler & Engstler-Schooler (1990) found that describing a face in words impaired later recognition of it. A large multi-lab Registered Replication Report (Alogna et al., 2014, Perspectives on Psychological Science) confirmed the effect but showed it depends heavily on timing. Schooler's lab also reported that verbalising can disrupt insight problem-solving itself ("Thoughts beyond words," 1993).

This keeps the essay honest. Articulation isn't a free, neutral transfer. Forcing words onto a half-formed pattern too early can degrade it. The lesson 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.

3. Transmitting tacit or intuitive knowledge to another agent

Polanyi's "we know more than we can tell" (The Tacit Dimension, 1966) is the philosophical anchor. The part that has been tested experimentally isn't tacit knowledge as such. It's the speaker's miscalibration about how well that knowledge got across.

**Speakers overestimate their effectiveness (Keysar & Henly, 2002, Psychological Science).** Speakers said ambiguous sentences while trying to convey one specific meaning:

• They consistently overestimated how often listeners understood. According to the Chicago Maroon's report on the study, speakers believed they were understood 72% of the time, while listeners actually misunderstood them 39% of the time; in 46% of the cases where speakers thought they had been understood, they had not been.
• Informed overhearers did not overestimate.
• 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."

A 2008 follow-up (Fay et al., Psychonomic Bulletin & Review) found that the overestimation reflects a positive response bias, not poorer discrimination. Only people high in fear of negative evaluation were calibrated. Well-replicated and mainstream.

This is the precise mechanism behind telling an AI agent "this name is wrong, rename it" as though the reason were obvious. Your intention is transparent to you because you hold it. The agent gets only the words.

The HCI literature on AI collaboration (2023–2024) names the gap directly:

• Gulf of Envisioning (Subramonyam, Pea, Pondoc, Agrawala & Seifert, CHI 2024). They add a new gulf to Norman's gulfs of execution and evaluation: the distance between a person's initial intentions and a prompt that can produce the desired output. They identify three misalignments, all forms of not knowing: what the task should be, how to instruct the model, and what to expect from its output. The first, not yet knowing what the task should be, is your failure mode described from the interface side.
• Metacognitive demands of generative AI (Tankelevitch et al., CHI 2024, Best Paper). Working with GenAI imposes metacognitive demands like those on "a manager delegating tasks to a team." Prompting requires self-awareness of task goals and task decomposition. Iterating requires well-calibrated confidence. You can't delegate clearly what you haven't made explicit to yourself.
• "Why Johnny Can't Prompt" (Zamfirescu-Pereira, Wong, Hartmann & Yang, CHI 2023). Non-experts explored prompt designs "opportunistically, not systematically." The paper draws on end-user programming research characterising novice specification as implicit, unplanned and overconfident rather than explicit, planned and cautious.

Standing: these are influential, well-cited conceptual and qualitative contributions, not large quantitative effect estimates. No study I found directly measures rework caused by unverified human intuitions given to coding agents. That specific link is still an open research gap.

4. Time pressure, momentum, and skipped verification

• Seizing under time pressure. Kruglanski & Freund (1983, Journal of Experimental Social Psychology) showed that time pressure increased primacy effects, stereotyping and numerical anchoring, meaning more "freezing" on first impressions. Fear of invalidity (being held accountable for errors) reduced it. That second finding is a lever you can pull yourself: raise the felt cost of being wrong before you broadcast.
• **Software engineering evidence (Kuutila, Mäntylä, Farooq & Claes, 2020, Information and Software Technology).** A systematic review of 102 studies found that most high-quality studies report higher productivity and lower quality under time pressure. The mechanism they describe is skipped verification: schedule compression means developers "omit or avoid reviewing, testing or documenting their code," which leads to more errors and more maintenance effort later. Figures cited in the review:
  • 75% of practitioners named time pressure as the top obstacle to reviews (Ciolkowski et al., 2003).
  • Time pressure was the root cause of 40% of defects in a telecom case study (Leszak et al., 2000).
  • Expected defects rose 47.1% from the lowest to the highest time pressure (Cataldo, 2010).
  • Caveats: much of the mechanism comes from interviews and simulation models, and some studies (e.g., Hale et al.) found no defect increase.
• Technical debt and rework. In the InsighTD surveys (Ramač et al., 2022, Journal of Systems and Software, 653 valid responses from six countries), time pressure or deadlines were the single most-cited cause of technical debt. Delivery delay, low maintainability and rework were the most-cited effects. In a longitudinal study, Besker, Martini & Bosch (2018, TechDebt '18) found that developers "waste, on average, 23% of their development time due to TD" and are frequently forced to introduce new debt because of debt that already exists. A 2025 repository study (Edbert, Codabux & Verdecchia, EASE 2025) found debt rising as deadlines approached in some open-source projects but not others. The evidence is real but mostly surveys and correlations.

2The Single Best Concept, and How Established It Is

**Recommendation: build the essay around the Feeling of Rightness as a stop signal (Thompson's metacognitive reasoning theory / Ackerman & Thompson's meta-reasoning framework), with the Eureka Heuristic as its insight-specific version.**

Why this over the alternatives:

• It names exactly the moment you describe. An intuitive output comes with a feeling. The feeling, not the output's validity, sets how much slow checking happens. The feeling is driven by fluency: how fast and easily the answer arrived.
• It explains why the intuition can be correct and still not knowledge. The FOR is often a good cue, which is why relying on it feels rational. It still isn't verification, and it can't tell good fluency from bad.
• It links to the neuroscience: a sub-second orbitofrontal reward burst (Oh et al., 2020), precision-weighting (Laukkonen et al., 2023) and hippocampal consolidation (Becker et al., 2025) all attach certainty and durability to an insight before conscious appraisal.
• It maps onto your software. Your tool refuses to let an assertion become a fact without passing a check. The FOR literature describes a mind whose default setting is to let a high-FOR assertion become a fact automatically.

How the pieces fit:

There isn't one canonical term covering the whole chain from intuition to broadcast. If you coin one, be clear that it's your synthesis. A defensible framing: "premature closure with an illusion of transparency": a high Feeling of Rightness closes the question inside your head, and the illusion of transparency makes you believe it's closed in your collaborator's head too.

3Recommendations for the Essay

1. Lead with Danek & Wiley's definition. An Aha, operationally, is feeling confident "without having to check it once more." That turns your personal story into a documented property of insight itself.
2. **Use Koriat to explain repeated rework.** Each time you returned to the same intuition, it grew more consistent with your model and more confident, without any added verification. Confidence tracks consistency, not truth.
3. Frame your software as an external metacognitive monitor. It does for claims what the FOR literature shows minds don't do by default: it separates the feeling of rightness from the status of being checked. The irony of building an external monitor while trusting your internal one without question is the essay's centre.
4. Use verbal overshadowing to avoid a naive conclusion. The fix isn't "verbalise everything immediately." It's to let the intuition form, then put it through explanation (the IOED test: can I state the mechanism?) and a deliberate-reflection pass (what else could explain this?) before it becomes an instruction.
5. For practice, borrow deliberate reflection's structure for any instruction you give an AI agent: state the intuition, list the evidence for it, list what would contradict it, name one alternative, and then tag it for the agent as a hypothesis or a decision. That is your own assertion/fact distinction applied at the prompt.

4Caveats

• Most insight research uses lab puzzles (remote associates, anagrams, magic tricks, Mooney images). Carrying it over to software architecture judgments is reasonable but untested.
• The "insight is more accurate" finding depends on task type, and the 2026 "Aha! Glow" result suggests the Aha signal may inflate judgments in open-ended design problems. Treat that as preliminary.
• Oh et al. (2020) and Becker et al. (2025) are single, well-executed studies, not replicated literatures. The active-inference account is theory.
• Premature-closure prevalence comes from retrospective case reviews, which are prone to hindsight bias. Deliberate-reflection benefits are strongest for experienced clinicians on hard cases.
• Software time-pressure evidence relies heavily on surveys and simulation models. Some rigorous studies find no quality effect.
• The HCI concepts (gulf of envisioning, metacognitive demands) are persuasive frameworks, but no study I found measures rework caused by humans giving unverified intuitions to AI coding agents. The essay would be pointing at an open gap rather than citing a closed finding.