An AI system may correct a wrong answer after producing it. Could it instead identify signs of uncertainty, insufficient justification, or unwarranted confidence before presenting that answer to a user, and trigger an appropriate check?
ONSeZI (Turkish: ÖNSEZİ) is a conceptual proposal for early risk detection. Despite its name, it does not claim human-like intuition, consciousness, or foreknowledge. It explores whether measurable signals could support a risk assessment that prompts verification, qualified language, a brief clarification question, or a proportionate pause.
1. WHY CONSIDER PREVENTIVE REVIEW?
Two stages can be distinguished: post-answer auditing and pre-answer preventive review. The former looks for errors after they are produced; the latter looks for early warning signals before an answer is delivered. These approaches are complementary rather than competing.
Uydurulizm is a conceptual critique of claims presented with confidence or apparent justification beyond what the evidence supports. ONSeZI could assess early signs of that problem. However, its proposed scope is broader: lost conversational context, contradictions, misunderstood intent, outdated information, and high-impact decisions could also be considered.
2. A PROPOSED ONSEZI WORKFLOW
Stage A — Interpret request and context: What is the user asking? What constraints follow from prior messages? Could a typo materially change meaning? Separate assumptions from known facts.
Stage B — Identify risk indicators: Does a claim require external evidence? Is that evidence available? Are there internal contradictions, currency or date uncertainties, unjustified confidence, or consequential ambiguities?
Stage C — Select a proportionate intervention: Respond normally at low risk; ask one short question for material ambiguity; check sources for verifiable claims; disclose missing evidence; or recommend stronger verification and accountable human oversight for high-impact, hard-to-reverse consequences.
Stage D — Check consistency before output: Compare the strength of the claim to the evidence. Verify not merely that a source exists, but whether it actually supports the claim.
Stage E — Conclude with reasons and capture evaluation lessons: Avoid needless repetitive warnings. Subject to appropriate privacy protections, errors and corrections may inform future evaluation criteria; this does not authorize storing personal data without permission.
These are proposed conceptual stages. They do not demonstrate that SOSYU or another AI system currently runs an independently validated module under this name.
3. ONSEZI, NOTICE, AND ISTISARELISM
The previously published SOSYU article “ÖNSEZİ ve NOTICE: Bağlamı Kontrol Eden Bir Yapay Zekâ Tasarımı” introduced an initial framework involving context and ambiguity of intent. This article expands its early risk detection dimension.
ONSeZI could serve as the layer assessing potential problems; NOTICE could refer to a brief, reasoned warning shown only when warranted. Not every risk assessment should generate an interruption. Istisarelism's ideas of proportionate objection and reasoned termination may help limit unnecessary repetition.
Example: A user asks whether a newly proposed concept is broadly accepted by researchers. If the available evidence consists only of its creator's publications and a DOI record, the system should distinguish publication from independent academic acceptance. The risk indicator is the gap between the strength of a claim and the character of its supporting evidence.
4. HOW COULD IT BE TESTED?
Hypothesis: A well-designed preventive risk assessment might reduce unsupported confidence and citations that do not support claims. It could also introduce false alarms, delays, and user fatigue.
Proposed comparison: On an independent, held-out evaluation set, compare (i) a baseline answer, (ii) post-answer checking, and (iii) ONSeZI-like pre-answer risk assessment. Evaluations should not rely exclusively on examples invented by the system's proponents.
Potential metrics include: rate of factually incorrect verifiable claims; rate of unwarranted certainty; proportion of citations that actually support their claims; risk detection and missed risks; false positives; needless clarification questions; latency; user burden; and appropriate pausing in consequential cases.
Counterexamples are essential. Does the mechanism wrongly inhibit creative fiction or humor? Does it mark correct and consistent answers as suspicious? Does the model over-trust its own risk score? Could the auditor itself misread a source?
5. LIMITATIONS AND RESPONSIBILITY
ONSeZI is not a claim of prediction, consciousness, or flawless self-monitoring. An AI system cannot be assumed to foresee every error. Internal confidence estimates alone do not establish real-world correctness. Independent evidence, calibration, human review, and evaluations of adverse effects remain necessary.
In health, legal, security, and high-stakes financial matters, an absence of detected risk must not be treated as a safety guarantee. The purpose is not to replace human agency but to make consequential uncertainty more visible before decisions.
6. CONCLUSION: ASKING BEFORE ANSWERING
Uydurulizm-oriented auditing may help examine the evidential basis of an answer already produced. ONSeZI asks whether part of that examination could occur before the answer is delivered.
Its central question is: “Before giving this answer, what reasons must I verify, and what uncertainty should I explicitly disclose?”
This article presents a conceptual model proposal open to development and independent testing. It does not announce an implemented, patented, peer-reviewed, or experimentally validated AI architecture.
RELATED PUBLICATIONS
SOSYU — ONSEZI and NOTICE: https://sosyu.com/makale/onsezi-notice-baglam-kontrolu (Turkish)