The short answer
AI can influence how confident people feel without reliably revealing whether a judgment is correct. The goal is not to trust yourself instead of AI, but to calibrate both: preserve your pre-AI view, investigate disagreement, require identifiable reasons for confidence changes, and compare forecasts with real outcomes.
Start with the distinction
Confidence is a signal, not a verdict.
Confidence describes how certain a person feels. Accuracy describes whether the judgment is correct. High confidence can protect independent thought or preserve a human error; low confidence can make useful correction possible or lead to unnecessary surrender.
An AI answer creates the same problem from the other side. Fluent language, detailed explanations, or an expressed level of certainty can influence reliance, but none guarantees that the answer is correct, complete, current, or suitable for the present case.
The useful objective is calibration: confidence that changes for identifiable reasons and becomes better aligned with outcomes over time.
Do not choose between blind self-trust and blind AI trust. Make both positions answer to evidence.
What research shows
Reliance can follow confidence even when confidence is poorly placed.
A 2022 experimental study examined how confidence in oneself and confidence in AI changed through experience. Within the studied tasks, self-confidence played an important role in whether participants adopted AI advice, and some participants misattributed poor outcomes to themselves and continued relying on a poorly performing AI. This does not show that AI always reduces self-confidence; it identifies one feedback loop that can distort reliance.
The 2025 survey of knowledge workers found a related association: higher task-specific self-confidence corresponded with more self-reported critical-thinking activity, while higher confidence in generative AI corresponded with less. Because the evidence was observational and self-reported, the direction of causality remains unresolved.
A preregistered CHI 2025 experiment found that explanations increased reliance on both correct and incorrect LLM responses. Sources and visible inconsistencies helped reduce reliance on incorrect responses in the tested objective-question tasks. Confidence should therefore be grounded in inspectable support, not the completeness of the narrative.
Notice the pattern
Five mechanisms can pull confidence away from accuracy.
These mechanisms are possibilities, not diagnoses. Naming them gives a decision owner concrete questions to ask when an answer changes how the case feels.
| Pattern | What happens | Corrective question |
|---|---|---|
| Anchoring | The first generated answer becomes the reference point for every later option. | What did I believe and consider before seeing this answer? |
| Fluent explanation | A complete narrative feels better supported than a fragmented but evidence-based view. | Which evidence and assumptions actually carry the conclusion? |
| Asymmetric visibility | You see your own uncertainty but not the model's missing context or weak provenance. | What relevant information was absent from the prompt? |
| Misattributed failure | A poor joint outcome is blamed on the visible human contribution rather than the recommendation or interaction. | Which premise, source, forecast, or action specifically failed? |
| False reassurance | An agreeable response reinforces the preferred story instead of testing it. | What is the strongest materially different explanation? |
A practical protocol
Make every confidence update explain itself.
A numerical estimate is not objective truth, but recording one before and after AI makes influence visible. Over a series of decisions, the same record can show where confidence tends to be too high, too low, or grounded in the wrong signals.
- Record the preferred option and a pre-AI confidence estimate.
- Name whether that confidence comes from direct evidence, reports, experience, interpretation, assumptions, or values.
- If AI disagrees, identify the observation or test that would distinguish the two positions.
- Ask under what conditions the recommendation would fail and what missing information could reverse it.
- Record which new evidence or reasoning changed confidence; fluency alone is not a valid reason.
- On the Reality Date, compare the forecast, confidence, rationale, and outcome.
Confidence is social
Protect independent judgment before the room converges.
Confidence is shaped by other confident voices as well as by AI. Team members should record their unique knowledge, prediction, concerns, confidence, and what would change their mind before seeing the group's positions or an AI summary.
After independent input is preserved, disagreement becomes useful evidence rather than friction to average away. An advisor can challenge the case without erasing client ownership, while a decision lead can see whether apparent agreement reflects shared evidence or shared influence.
The product connection
Keep the reasons behind a change in confidence.
For your next confidence update, identify the evidence or reasoning that earned it. Reality Skill is a decision-making workspace that preserves the initial position, confidence, evidence, alternatives, independent perspectives, AI contributions, final commitment, and result. Your starting view remains open to testing, not entitled to a veto.
A decision journal records what changed and the reasons you accepted. On the Reality Date, compare those expectations with the result. Across suitable decisions, these records can help you examine whether confidence was well grounded; learning may mean becoming less certain as well as more certain.
Trust is not the target. Appropriate reliance and accountable judgment are.
Research and sources
Evidence behind this guide.
- Human Confidence in Artificial Intelligence and in Themselves: The Evolution and Impact of Confidence on Adoption of AI Advice Computers in Human Behavior
An experimental study and quantitative model in bounded tasks; it should not be read as evidence that every AI interaction lowers self-confidence.
- The Impact of Generative AI on Critical Thinking Microsoft Research / CHI 2025
Reports confidence-related associations in self-reported work examples, not causal direction.
- Fostering Appropriate Reliance on Large Language Models: The Role of Explanations, Sources, and Inconsistencies Microsoft Research / CHI 2025
Tests reliance in objective-question tasks and shows why explanation presence alone is an inadequate confidence cue.
- AI Risk Management and Human-AI Interaction NIST AI Resource Center
Guidance on differentiating human and AI roles and accounting for variation in human-AI configurations.
- Appropriate Reliance on Generative AI: Research Synthesis Microsoft Research
Explains why calibrated self-reliance and AI reliance both matter.