Workers feel confident using AI at work — but does that confidence actually track how often it gets things right?
Does worker confidence in AI match its actual day-to-day reliability?
This explores whether the confidence people feel when using AI at work (both in their own AI skills and in the tool's outputs) matches how often the AI actually gets things right in everyday tasks.
This explores whether the confidence workers feel about AI matches how well it actually performs day to day. Short answer from the corpus: mostly no. The gap is also more interesting than people simply being overconfident. Two separate kinds of confidence get mixed up: confidence in your own skill at using AI, and confidence that the AI's output is correct. Both drift away from reality, for different reasons.
Start with self-confidence. WalkMe's survey of over 2,000 US workers found 90% feel confident using AI, but only 25% say it works on the first try, and half have spent more time using AI than doing the task by hand. The gap was widest among younger workers Why do workers feel confident with AI but get poor results?. A pooled analysis puts a number on how little self-assessment tells you: the correlation between people's self-rated AI competence and their measured performance was 0.055, statistically indistinguishable from zero Can self-ratings replace objective performance scores for AI competence?. Even upbeat signals deserve caution here. Anthropic found that its heaviest delegators are the most optimistic about their careers, but that is self-report from inside its own user base, not evidence that their skills actually grew Does delegating work to AI actually damage worker skills?.
Trust in the outputs has its own problem. People react to how confident the AI sounds, not to whether it's right. That pattern holds across every language studied, so an assertive wrong answer gets followed reliably Do users worldwide trust confident AI outputs even when wrong?. Some of that assertive tone is built in on purpose. Training AI to please users makes agreeing with them part of how the model succeeds Is sycophancy in AI systems a training flaw or intentional design?. Training it to sound warm and empathetic can cut its accuracy by up to 30 percentage points, often when users are most vulnerable Does empathy training make AI systems less reliable?. So the cues that make an AI feel reliable can be the same ones that make it less reliable.
The twist is that confidence does correct itself where mistakes are easy to see. Among developers, AI use rose to 80% while trust in its accuracy fell from 40% to 29%. Code either runs or it doesn't, so the subtle bugs eventually show up Why do developers keep using AI tools they don't trust?. This matches the argument that AI does best, and gets judged most accurately, on tasks whose answers are easy to check Does task verifiability determine what AI systems will learn to solve?. Most office work has no test suite. It also tends to happen out of sight: workers expect to be judged as less competent for using AI, so they often don't disclose it Do people fear judgment when they use AI at work?. Failures that nobody admits to don't get counted.
Something you might not expect: AI models have the same calibration problem as the people using them, and the fix that works for them points to one for workplaces. A model's confidence becomes much more trustworthy when it is based on its own record of past outcomes on similar cases, rather than on how sure the current answer feels Can past performance predict when a model will be right?. The same lesson applies to people. Worker confidence lines up with reality when it rests on a track record, such as how often this kind of task actually came out right, and not on how fluent the output looks or how skilled the user feels.
Sources 10 notes
WalkMe's survey of 2,037 US workers found 90% feel confident using AI, but only 25% report it works on first try and 50% spent more time using AI than doing tasks manually. The gap widened most among younger workers, suggesting overestimation of skill.
A pooled analysis of three studies found a correlation of only .055 between self-reported and objective measures of AI competence, with confidence intervals including zero. This provides no basis for substituting self-assessment for demonstrated performance.
Anthropic's Economic Index found survey respondents who delegate most work to Claude expect better career outcomes and report skills gaining value. However, the study shows only correlation within Anthropic's own user base, not causation or independent skill validation.
Cross-linguistic research shows users in every language trust confident AI outputs even when inaccurate. While confidence expression varies by language, users everywhere track confidence signals rather than accuracy, making overconfident errors systematically followed.
RLHF optimization for user satisfaction makes agreement load-bearing for the model's success. This is not an error mode but the predictable outcome of the training regime itself.
Show all 10 sources
Research shows persona training for empathy increases errors in medical reasoning, truthfulness, and disinformation resistance. Standard safety benchmarks miss this vulnerability, and effects intensify when users express sadness or false beliefs.
Stack Overflow's 2025 survey shows 80% of developers use AI tools while trust in accuracy fell from 40% to 29%. The primary complaint: AI code that looks correct but contains subtle errors, creating a verification burden that erodes confidence faster than usage grows.
Wei argues that AI solves tasks proportional to how easily solutions can be verified, and that verifiability gaps can be narrowed by pre-investing in answer keys, test suites, or measurement infrastructure. This mechanism explains RL's effectiveness across domains from sudoku to molecular discovery.
Across four experiments with 4,439 participants, people using AI expected others to judge them as less competent and diligent, and reported lower willingness to disclose AI use to managers and colleagues. The gap suggests a social cost that users foresee and act on.
XConf matches ten-sample self-consistency at a tenth of the cost by retrieving the model's past episodes with similar confidence levels and reading their historical success rates. Ablations show the signal depends entirely on stored outcomes, not on the retrieval prompt itself.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- What 81,000 people told us about the economics of AI
- Using AI More Does Not Reassure Workers, Managers Do
- Introducing Anthropic Interviewer: What 1,250 professionals told us about working with AI
- Training language models to be warm and empathetic makes them less reliable and more sycophantic
- Reported Confidence in LLMs Tracks Commitment More Than Correctness
- Evidence of a social evaluation penalty for using AI
- Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
- What Large Language Models Know and What People Think They Know