The psychology of the human side of AI: why confidence, not access, decides adoption

Oct 14

7 min read

If you have been using AI more this year and feeling less sure of yourself, you are not imagining it. And you are not alone.

Last January, ManpowerGroup published its Global Talent Barometer, a survey of nearly 14,000 workers across 19 countries. The results showed:

Regular AI use had gone up 13 points in a year.
Confidence in using AI and technology had gone down 18 points, from 81 per cent to 63.
Among Gen X, my generation, only 52 per cent felt confident.
Among baby boomers, 36.
And 57 per cent of workers said they had no recent training or mentoring.

In the same survey, 92 per cent of people said they were confident in their own skills and experience. The respondents of that survey had not lost faith in themselves. They have lost faith in themselves with this one thing.

That is a very different problem, and a much more fixable one.

Introducing self-efficacy.

The belief that you can do a particular thing. The word "particular" matters. Self-efficacy is not a personality trait. You can be a confident surgeon and a nervous driver. You can run a business for twenty years and still freeze in front of a blank prompt box in your AI app.

The ManpowerGroup numbers are a textbook picture of this: high belief in your own competence, low belief in one specific domain.

It matters more than most leaders expect, because it is the part of adoption that does not show up in a business case, or in your licences. You can't get someone in IT to run a usage report. Someone can have the tool, the login and the training, and still hold back, because they do not yet trust themselves with it.

Australian data shows the same result. Gartner's survey of 574 Australian employees, run in the first quarter of this year, found that confidence is the strongest driver of whether AI adoption sticks, ahead of training. It also found that only about half of employees report clear guidance, training or support. Both surveys are from vendors, so should be treated as a guide. But they match what I have seen in 25 years of leading technology change.

Why more use and less confidence?

The psychologist Albert Bandura spent a career working out where self-efficacy comes from.

It is built from four sources.

Doing the thing and succeeding at it.
Watching someone like you succeed at it.
Being told by someone credible that you can.
And how you read your own body and mood while you try.

Notice that the first and biggest source is doing the thing. Which raises an obvious question. AI use went up. By Bandura's logic, confidence should have gone up with it.

Why did it fall? Because use is not the same as mastery. A mastery experience is not "I opened the app." It is "I attempted a defined task, finished it, and could tell it went well."

The task was never defined.

A study published in Frontiers in Psychology in August surveyed 541 employees at five high tech firms in China. As people used generative AI more, two things rose together: their belief that they could take on a bigger role, and their uncertainty about what their role actually was.

What tipped the balance was AI literacy. Where people understood the tool in the context of their work, confidence pulled ahead. Where they did not, confusion did.

Granted, this is just one cross sectional study. But it fits the theory exactly.

Self-efficacy is always confidence at something. If nobody has said what good AI use looks like in your job, there is no standard to measure a win against. People are using the tool a great deal and completing almost nothing they can point to.

The target keeps moving.

Even where a task is clear, the ground shifts under it. A new model, a better prompt, a different workflow, every week, in every feed. Each one restarts the clock on mastery. And the sheer volume of change and information does something else: it floods people, and Bandura's fourth source is how people interpret that feeling.

Overwhelm gets read as evidence. "I cannot keep up, so I must not be able to do this."

It is not a judgment about ability at all. It is a judgment about volume.

The examples are the wrong examples.

Watching someone succeed raises your confidence only when you see them as similar to you.

Watching someone very different do it effortlessly does the reverse.

For most of the workforce, watching someone on their team they can't relate to use AI successfully or a stranger on LinkedIn, Insta, Tik Tok, you name it, who "automated their whole week." are the wrong comparison, and comparison is not neutral.

Mistakes were not safe.

Another Frontiers in Psychology study, published in May, followed 449 regular AI users across two surveys. AI use was linked to higher confidence, hope and resilience, and through those to more fresh ideas at work, but only where people felt mistakes were treated as information to learn from. Where they were not, the benefit did not appear.

The authors flag limits, including self-reported measures and one cultural setting. But it lines up with something Gartner found in Australia: of workers expected to use AI, 86 per cent also use personal AI tools. The survey does not say why. My guess is that a personal tool is a safer place to get it wrong without anyone watching.

What about motivation to change?

Working in technology change, I see all the time a lack of motivation to adopt new technology.

Motivation, or lack of, is what you see on the surface. People avoid what they do not believe they can do, so when confidence drops, effort drops with it.

Fixing confidence usually fixes motivation.

There is one real exception. Bandura separated "can I do this?" from "will it be worth it?", and ManpowerGroup found 58 per cent of Gen Z expect their role to be replaced within two years.

If the reward for getting good at AI looks like making yourself redundant, no amount of confidence will produce effort to change or adopt AI.

Anecdotally, I see this all the time, and each time, it catches me out as I don't expect Gen Z to lag on adopting AI.

That is a conversation a leader has to have directly. What does good use of AI in your job look like?

What to do this week?

Addressing the findings in those different surveys and doing something about it, is within your control.

Three small, deliberate moves cover most of it:

1. Define one task and measure confidence in it. Pick a single (AI) task where a win is almost guaranteed. This doesn't need to be the most impressive task. The one most likely to work. A first draft of a routine email, a summary of a long document, a meeting agenda. Say out loud what a good result looks like. Make that the only AI goal for the week. Then add one question to your next pulse survey: "How confident do you feel using AI for [that task]?" A defined task and a confidence question tell you who is not just opening up the AI tool, and give people a win they can actually bank.

2. Choose the example to showcase.

Ideally, you will have someone on your team, roughly the same age and role as the people you are worried about, who has worked something out. Ask them to show it. Five minutes, no slides. "Here is what I tried, here is what went wrong first, here is what works now." If you don't have that person, become that person. A leader decides who gets held up as the model, and the right model is the one who looks like the audience or one they can relate to.

3. Make the first mistake safe.

Get your team together for 5-15 minutes each week with a focussed AI review agenda; What did AI get wrong for us this week? What did we learn? What do we do differently? This is where you get to decide together what is worth exploring more or what is worth ignoring.

This is the one I think can make the biggest difference, particularly when the leader goes first, every time. Normalise and create a culture of openness and transparency. Showing vulnerability and failing could save you the risk of staff using unathorised AI tools and exposing your data becuase they don't think they can or should do use the tools at work.

It might drop before it picks up

Confidence falling while use rises is not a sign that people are failing. It is the sign that they have used the tool enough to see what they do not yet know, which is the first stage of getting good at anything. Moving from conscious uncompotence to conscious competence.

You may not control how fast the technology moves. You do control whether the task is clear, who gets held up as the example, and what happens when the first answer is wrong.

If you lead a team, which of the three would you start with? I would love to hear. Sources:

- ManpowerGroup Global Talent Barometer 2026, press release and global report, 21 January 2026 (investor.manpowergroup.com/node/68676); figures also reported by Workplace Journal, 21 January 2026.

- Gartner Global Talent Monitor survey (Australia, 574 respondents, Q1 2026), press release 2 June 2026.

- Bandura, A. (1977). Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review, 84(2), 191 to 215. Bandura, A. (1997). Self-Efficacy: The Exercise of Control. W. H. Freeman.

- Frontiers in Psychology, 14 August 2026: study of 541 employees on generative AI use, role breadth self-efficacy and role ambiguity (doi 10.3389/fpsyg.2026.1793095).

- Frontiers in Psychology, 8 May 2026: study of 449 regular AI users on AI adoption, psychological capital, innovation and error management culture (doi 10.3389/fpsyg.2026.1740508).

General information only. Tracy Attanasio is the Founding Principal of Nuovo Insights.