Anthropic runs one of the biggest experiments in the world on how people actually learn to use AI well, since millions of Claude users give the company a constant stream of data on what works. Kristen Swanson, who leads education and research there, started this year by tracking specific behaviors: did someone tell the AI who they were writing for, did they explain the role they wanted it to play. By February, that approach was already breaking down. The models had gotten good enough to figure most of that out on their own, so teaching the tricks stopped being useful.
Her team's replacement approach drops the checklist and focuses on three things at once: the beliefs people bring to AI, whether the organization actually gives them working tools and permission to use them, and how much effort it takes to get a task done well. Training any one of these alone, the team found, did far less than working on all three together.
The more interesting finding sits underneath the new approach: trust, not skill, is the real bottleneck. Atlassian's internal Teamwork Lab tested what happens when someone admits to using AI at work. The odds a coworker would call them lazy jumped from 3 percent to 24 percent once AI use was disclosed. That is not a training gap. That is a culture problem, and no course fixes it.
This matches what several other studies have found this year. Separate surveys from firms including WalkMe and a global study covering tens of thousands of workers across dozens of countries found that large shares of employees, often close to half, quietly hide their AI use from managers and peers. People are not confused about how to use the tools. They are worried about what happens if they admit they are using them.
That fear has a real cost. If employees will not disclose how they actually work with AI, teams cannot build the shared systems Swanson describes, where an AI agent sits inside a shared workspace and acts on documented team goals. Her point, that if a task is not written down for an agent it does not exist, only works if people are willing to write things down in the open.
There is a matching problem on the hiring side. Companies have rushed to make AI fluency a top requirement in job postings, in many cases ranking it above years of experience in a candidate's actual field. A large recent hiring study found that even with formal AI fluency requirements in place at most companies, a majority still ended up making a bad hire, someone who talked well about AI in an interview but could not apply it on the job. Fluency, it turns out, is hard to fake but just as hard to test for with a checklist.
The pattern across all of this is consistent. Companies that are spending real money on AI training, and training budgets nationally ran past 100 billion dollars in the US alone last year, are discovering that the money does not buy adoption by itself. What buys adoption is whether leaders visibly use the tools themselves, whether admitting you used AI carries no penalty, and whether people believe learning to work with AI adds to their value instead of replacing it. Skip that part, and the fanciest training program in the world will not move the needle.