Do explanations and sources boost trust in AI answers together, or do they pull in opposite directions?
Do explanations and sources have additive or offsetting effects on user trust?
This explores whether explanations and cited sources work together to raise trust in AI answers, or whether they pull in different directions. Specifically, it asks whether they help people trust the right answers and doubt the wrong ones.
This explores whether adding explanations and sources to an AI answer stacks up into more trust, or whether the two cancel each other out. The most direct evidence says they push in opposite directions. In a 308-person experiment, explanations made people rely more on LLM answers whether the answers were right or wrong. Sources had the reverse effect: they reduced misplaced trust in wrong answers Do LLM explanations actually help people judge answer quality?. So the honest answer is 'offsetting', with one catch. They don't offset by adding and subtracting trust. Explanations raise trust across the board, while sources help people sort good answers from bad ones.
Why would explanations inflate trust? The corpus keeps returning to one idea: fluency feels like correctness. Readers come away more confident than the model's own internal probabilities justify, because ordinary explanation language doesn't signal how unsure the model is Why do readers trust LLM explanations more than they should?. A separate study of 559 people found that how *pleasant* a reasoning trace felt, more than any increase in trust, was what led people to overrate their own understanding Does pleasant-feeling explanation drive overconfidence more than trust?. Agent plans show the same pattern. Wrong plans with clean logical structure misled users as much as correct ones, and being allowed to edit the plan didn't help Does letting users edit agent plans improve trust calibration?.
Sources only help, though, if people actually check them. On their own, citations can become one more surface cue. An analysis of 24,000 search interactions found that irrelevant citations raised user preference almost as much as relevant ones Do users trust citations more when there are simply more of them?. So 'sources calibrate trust' holds only under some conditions. When readers just count citations, sources inflate trust the same way explanations do, and you get two inflating effects stacked together rather than one cancelling the other. Conversational style works in a similar way. Responsiveness and speed build trust in ChatGPT regardless of accuracy Does conversational style actually make AI more trustworthy?.
The more useful lesson is that what calibrates trust is *friction* and *feedback*, not a particular feature type. In the explanation study, inconsistent explanations also reduced misplaced trust, because they gave readers something to snag on Do LLM explanations actually help people judge answer quality?. Format has to fit the task too. Argument-map rationales improved calibration on verbal reasoning but hurt it on visual reasoning Do visual rationales help or hurt how people calibrate trust?. And over time, seeing real outcomes is what fixes trust. Initial bias against disclosed AI partners reversed only when users repeatedly saw the results Does revealing AI identity help or hurt user trust?.
The corpus has no study that crosses explanations and sources in a single factorial design. So the clean interaction effect, whether the combination is more or less than the sum of its parts, remains an open question. The evidence that does exist suggests asking a different question: not 'how much trust does this feature add?' but 'does this feature give the reader a way to catch a wrong answer?'
Sources 8 notes
A 308-person experiment found explanations increased reliance on LLM responses regardless of accuracy, while sources and inconsistent explanations reduced inappropriate trust in wrong answers. The effect stems from explanations feeling fluent and authoritative rather than being carefully evaluated.
Lab experiments on MMLU and TriviaQA show readers' confidence judgments track explanation language poorly compared to the model's internal probability estimates. The gap narrows when explanations are deliberately engineered to reflect internal confidence, but natural language explanations leave readers systematically overconfident.
A mediation analysis of 559 participants found that the pleasantness of reasoning traces—not increased trust in the system—statistically carries the path to inflated self-assessment, consistent with processing-fluency misattribution of understanding.
A 248-person study found that letting users edit LLM agent plans did not improve trust calibration. Instead, trust tracked actual plan quality. Plausible-sounding but wrong plans misled users just as much as correct ones, because clear logical structure feels trustworthy regardless of accuracy.
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.
Show all 8 sources
A focus group study shows conversationality—not accuracy—drives ChatGPT trust through social response activation. Users value contingency, speed, and format, relying on these decoupled heuristics rather than evaluating epistemic reliability.
In an N=204 study, argument-map rationales improved trust calibration on verbal reasoning tasks yet impaired it on visual ones. Subjective ratings (satisfaction, helpfulness) reversed in each domain, suggesting format-task fit matters more than format alone.
Users initially avoid AI partners when identity is revealed, but this preference reverses after repeated interactions with visible results. The learning mechanism—observing consistent outcomes—is essential; disclosure without feedback produces no calibration.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Fostering Appropriate Reliance on Large Language Models: The Role of Explanations, Sources, and Inconsistencies
- Evaluating the False Trust Engendered by LLM Explanations
- LLM or Human? Perceptions of Trust and Information Quality in Research Summaries
- Seeing to Think? How Source Transparency Design Shapes Interactive Information Seeking and Evaluation in Conversational AI
- What Large Language Models Know and What People Think They Know
- Explain-Query-Test: Self-Evaluating LLMs Via Explanation and Comprehension Discrepancy
- The Decision to Verify: How Warmth and User Characteristics Shape Reliance on Conversational Agents for Information Search
- Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent