What changes when evidence shows that a belief, decision, or system is wrong?
My work studies correction as a process: how people and AI systems respond to contradictory evidence, whether revision carries into later reasoning and action, and how we can distinguish genuine updating from agreement, memory, or polished self-correction.
The work spans philosophy of psychology, human–AI interaction, AI evaluation, and institutional decision-making. Some papers are theoretical or methodological; others contain behavioral evidence or proposed experiments. Claims are intended to remain open to criticism, comparison with existing constructs, empirical testing, and rejection.
Research questions
Three recurring problems
Different domains, but a small number of recurring questions.
When should a correction change a belief?
How should acceptance, resistance, or suspended judgment track the actual strength of a correction claim?
Does correction change later AI behavior?
Changing one answer is not enough. Does the revision propagate to dependent reasoning, persist, transfer, and remain revisable?
What makes AI seem continuous or mind-like?
Memory, personalization, social presence, and apparent self-correction should be studied separately from claims about machine consciousness.
Selected research
Current core papers
A short path through the current research program.
Beyond Changing the Answer: Warranted Downstream Correction in Large Language Models
Defines selective, evidence-justified revision that propagates through materially dependent claims and actions while preserving independently supported claims.
The Correction-Capacity Model: Warrant-Responsive Rationality Under Self-Relevant Threat
Models how revision should track evidential warrant when correction threatens identity, competence, status, or self-worth.
The Corrective Continuity Hypothesis: Perceived AI Consciousness, Persistent Memory, and Trace-Bearing Revision
Separates remembered continuity from durable correction and proposes tests of grounding, persistence, transfer, traceability, and revisability.
Answerability Architecture for Corrigible AI Interaction: Corrigible Continuation Without a Self
Develops a design and governance model in which evidence, uncertainty, verification, authorization, reversibility, and accountable human judgment can change what an AI system does next.
The Machine That Seems Awake: Anthropomorphism, Social Presence, and Evidence for AI Consciousness
Separates the experience that an AI seems present or mind-like from evidence about architecture or subjective experience.
Psyche-Like Effects Without a Psyche
With Amy Jean Clark. Examines psychologically meaningful effects around fluent AI interaction without treating linguistic continuity as proof of an inner psyche.
The Corrective Remainder: Selection, Residuals, and Revision Under Feedback
Treats recurring discrepancies, anomalies, hidden costs, and unintended consequences as useful only when they can be traced, compared with rivals, and shown to alter a governing model or decision rule.
Carrier, Trace, and Model Revision: What Invisible Claims Must Earn
Asks what evidence an unobserved cause should provide: an identifiable pathway, a discriminating trace, comparison with rivals, temporal prediction, and conditions for revision.
Method
Evidence before interpretation
The current method is deliberately conservative: separate observation from inference, compare alternatives, state uncertainty, and specify what would change the conclusion.
Start with what is observable. Identify the claim under pressure. Ask whether the contradictory evidence is strong enough to justify revision, what conclusions actually depend on the disputed premise, and what later result would show that the revision persisted rather than merely being acknowledged.
Free AI prompt
Evidence & Revision CheckAnalyze the situation without assuming that the first coherent explanation is correct. Separate: 1. directly observed facts; 2. interpretations or assumptions; 3. missing information; 4. plausible alternative explanations. Then identify the main claim being relied on and ask: - What evidence supports it? - What evidence pushes against it? - How strong is the conflicting evidence? - What would justifiably need to change if the correction is accepted? - What should remain unchanged because it is independently supported? - What later observation would show that the revision actually persisted? - What evidence would weaken or overturn this analysis? If the evidence is insufficient, say so. Do not diagnose people, infer hidden motives from sparse information, or turn analogy into proof.
Evidence status
What this research does — and does not — claim
Conceptual models
Correction-capacity, corrective continuity, and related constructs are proposals, not established scientific laws.
Behavioral investigations
Some AI papers contain behavioral evidence. Results are described at the level the designs support and should not be generalized beyond them without replication.
No consciousness shortcut
Memory, social presence, self-reference, or durable correction do not by themselves establish phenomenal consciousness, personhood, or an inner psyche.
About
Vladisav Jovanović
Independent researcher · Applied Psychology · ICF Associate Certified Coach
My current research focuses on correction, belief revision, AI corrigibility, human–AI interaction, and the difference between a system changing its language and changing what it does later.
Earlier work developed these questions under the broader label Structural Intelligence. I now use that label mainly as an umbrella for the research program rather than as a claim that one theory explains psychology, institutions, artificial intelligence, or physical systems through a single mechanism.
AI-use disclosure: Generative AI has been used for literature organization, drafting, restructuring, counterargument generation, language editing, and workflow support. Concept selection, interpretation, verification, revision, and publication decisions remain my responsibility. AI assistance is not treated as empirical validation or independent peer review.