Workers who hand AI the most tasks feel the most optimistic about their careers — but does that feeling mean their skills actually improved?

This explores whether workers' sense that AI is good for their careers can stand in for evidence that their abilities actually changed. The corpus says no. Self-report and real skill come apart in several places, and the gap shows up in human and machine studies alike.

The starting point is Anthropic's Economic Index finding that the people who hand the most work to Claude are also the most optimistic. They expect better career outcomes and say their skills are gaining value Does delegating work to AI actually damage worker skills?. That's a correlation inside one company's own user base, and nobody independently checked the skills. A pooled analysis of three studies makes the problem concrete: people's ratings of their own AI competence correlated with their measured performance at only .055, and the confidence interval included zero Can self-ratings replace objective performance scores for AI competence?. In other words, how capable people feel with AI tells you almost nothing about how capable they are. If self-assessment can't track current competence, it can't be trusted to track *changes* in competence either.

You might expect the job market to correct for this, but it partly rewards the self-report. In a conjoint experiment with 1,725 recruiters, listing AI skills raised interview invitations by 8 to 15 percentage points, and a certificate added only a little over simply claiming the skill Do AI skills help candidates get more job interviews?. So optimistic heavy delegators may be right that their careers will benefit, and still have no better skills. Feeling valued, being treated as valuable and being more capable are three separate things.

The less obvious lesson comes from AI research, where the same trap has been studied in models. Models trained to imitate ChatGPT picked up its confident, fluent style well enough to fool human evaluators, while their factual accuracy and ability to handle new tasks didn't improve Can imitating ChatGPT fool evaluators into thinking models improved?. Research on self-improvement reaches a similar conclusion: a system judging its own progress tends to go in circles. The methods that work bring in an outside reference point, such as an earlier version, a third-party judge or feedback from a tool Can models reliably improve themselves without external feedback?. There is a hopeful counterpoint. Models *can* be trained so that their stated confidence matches their real reliability, but only by checking those self-judgments against actual performance Can models learn to judge their own performance accurately?. That suggests a design principle for studying workers too: self-reports become useful once they're calibrated against something measured.

The gap in the corpus matters here. It has no longitudinal study that tests workers' skills before and after heavy AI use, so the question of whether delegation builds or wears down skill is still open. The material on hand shows that optimism is a real and possibly useful signal about how people experience their work, and that it is a poor stand-in for measured skill.