Who Uses AI (and How)?

Tracking the evidence on AI adoption

This post is intended as a living resource. We will update it periodically as new evidence accumulates. The current version reflects research available through February 2026.

A customer-support agent in Manila and a senior developer in San Francisco get access to the same AI tool. One sees their productivity jump to 35 percent. The other barely notices. In a previous post, one of us (Alex) reviewed the growing literature on AI’s productivity impact and argued that a disconnect exists between the micro evidence (controlled studies showing real productivity gains) and the macro evidence (aggregate statistics showing few gains thus far). That post focused on how much productivity AI delivers. But at the heart of that result is the question of adoption: the impact of AI on the economy also depends on who adopts AI and how they use it. This is the focus of the current post.

The answer depends on where you look. Controlled studies where everyone is directed to use AI (and usually trained how to use it) tend to find an equalizing effect: less experienced workers benefit more, compressing the performance distribution. But observational data on who actually adopts AI paints a very different picture: usage is currently concentrated among higher-skilled, higher-income workers and countries. Put another way, who chooses to use AI and how they use it is not random—it is determined by their personality, education, experience, and skill—and this can lead to very different economic consequences than what is implied by experimental studies where everyone in the AI condition is “treated.”

We conjecture that the experimental studies, by design, abstract away the adoption margin—the question of who chooses to use AI, how much, and how effectively. When a study randomly assigns access and provides a well-defined task, it cleanly estimates the productivity effect conditional on use. But the adoption margin is extremely important in the real world. Who selects into using AI? How much time do they invest in learning to use it well? What organizational support do they receive? The answers to these questions are not random and are systematically correlated with existing advantages. To get a hint of this you can just look at the relationship between GDP and AI use below (from Anthropic’s Economic Index), the relationship is unambiguously positive.

As we argue in this post, whether AI’s potential translates into aggregate impact depends largely on (i) who adopts AI, (ii) how intensively and effectively they use it, and (iii) whether organizations invest in the complements—training, protected experimentation time, managerial support—that translate individual access into realized value. We use “adoption margin” a bit broadly here to mean everything between access and realized value, including adoption, task choice, interaction style, and organizational incentives to facilitate AI use.

The controlled studies are telling us something real about what happens when the adoption margin is leveled. They show that AI can be (potentially) equalizing, that the workers with the most room to grow are the ones who benefit most. The question this post grapples with is why those conditions aren’t met in many real world settings. We also consider what it would take to close the gap: Recent survey data points to the importance of organizational scaffolding, both top-down encouragement from managers and the training for learning how to use the tools effectively.

The Controlled Studies: AI as the Equalizer

We catalogued the controlled studies in detail in the productivity post, so we’ll keep this brief. A recurring finding across many of these studies is that AI disproportionately helps less experienced or lower-performing workers. Brynjolfsson, Li, and Raymond (2025) studied a conversational AI assistant rolled out to 5,172 customer-support agents and found that while average productivity increased by 15 percent, gains were concentrated among less experienced agents (30–35 percent improvement), while top-quintile agents saw minimal speed gains and slight quality declines. Noy and Zhang (2023), Cui et al. (2025), and Gambacorta et al. (2024) each found the same pattern: larger effects for less experienced, more junior workers.

The equalizing result across these studies is consistent, but it is a conditional result: conditional on being given a well-defined task within AI’s competence, and on actually using the tool. What happens when these conditions relax?

Otis et al. (2023) provide one such test. They randomly assigned 640 Kenyan small-business entrepreneurs access to a GPT-4-powered AI business mentor via WhatsApp—but unlike the studies above, they did not pre-specify what entrepreneurs should ask about. The average treatment effect was zero. But beneath this null, high-performing entrepreneurs saw revenues increase by over 15 percent while low performers did roughly 8 percent worse. The divergence was not because the AI gave different advice—it was because low performers selected harder, more open-ended problems where AI advice was less actionable. When the task-selection dimension of the adoption margin is left to the worker, the equalizing effect reverses.

Paradis et al. (2024), in Google’s internal evaluation, found that more experienced developers saw bigger effects from AI assistance, likely because verifying and integrating AI-generated code requires the kind of judgment that comes with experience. As OpenAI’s Greg Brockman recently posted: Taste is a new core skill, but this skill may be unevenly distributed.

Greg Brockman@gdb

taste is a new core skill

7:35 PM · Feb 16, 2026 · 1.7M Views

---

791 Replies · 1.14K Reposts · 8.91K Likes

These results suggest a pattern: when experiments constrain the adoption margin—assigning specific tasks, embedding AI into workflows, equalizing access—AI compresses the skill distribution. When any of these constraints relax, the advantages tilt toward those who already have the judgment to use AI well. A natural next question is: in practice, are those conditions met? The evidence suggests they are often not.

Who Uses AI?

Step outside the controlled studies which try to keep treatment and training constant, and a different picture emerges. Studies looking at who actually uses AI also tell a consistent story: adoption is often correlated with variables associated with advantage, which explains the patterns in the productivity data. In this section, we review the evidence on what variables tend to be correlated with adoption.

Income and Geographic Origin

The Anthropic Economic Index finds that Claude usage is geographically concentrated and strongly correlated with national income. High-income, technologically advanced countries like Israel, Singapore, United States, and Canada dramatically over-index on per-capita usage, while lower-income countries lag far behind.

This is not surprising. Hacker News calculated that a ChatGPT Pro subscription would cost the equivalent of 38.6 months of income in low-income countries. Even where access is technically possible, the adoption margin on how much and how effectively people use AI is constrained by price.

Within the U.S., the composition of states’ economies—not just income—predicts usage, with knowledge-work-heavy states leading.

The most recent update shows little sign of convergence: the countries and regions that adopted earliest remain ahead. If the adoption margin were broadening, i.e., if lower-income workers and countries were catching up, we would expect to see this gap narrowing, and it’s not.

A recent BCG survey found that managers use AI at nearly twice the rate of front-line workers (coined as the “silicon ceiling”). Only 36 percent of workers report feeling properly trained to use AI. This is a direct measure of the adoption margin within organizations: holding access roughly constant, higher-status workers are far more likely to actually use the tool.

Demographics

Humlum and Vestergaard (2025), using administrative data on 18,000 Danish workers linked to tax records and employment histories, provide some of the cleanest evidence on who self-selects into AI use. Women were 16 percentage points less likely to use ChatGPT for work than men. Importantly, workers who adopted ChatGPT were those who had been earning more before it existed. Additionally, each additional year of experience reduced adoption likelihood by 0.6 percentage points, and many workers reported that employer restrictions or lack of training prevented use.

In a similar vein, Carvajal, Franco, and Isaksson (2024) also show that gender is another dimension of the adoption margin: among 595 business school students in Norway, male students were 25 percent more likely to be high AI users, and the gap was widest among the highest-performing women. The key mechanism appears to be perceptual: female students were more likely to view AI use as “cheating” and less likely to persist when the tools did not deliver on the first try. Indeed, when these perceptual differences were controlled for, the gender gap disappeared.

The ManpowerGroup 2026 Global Talent Barometer highlights more dimensions of heterogeneity: while regular AI usage jumped 13 percent among workers in 2025, confidence in the technology, which is a key variable for adoption, plummeted by 18 percent. The confidence gap is starkly demographic: a 35 percent decrease among baby boomers and a 25 percent drop among Gen X. Additionally, 56 percent of workers globally report receiving no recent skills development despite their organizations adopting AI, and 57 percent report not getting any mentorship.

While access to AI is spreading, the willingness to use it and know-how is much more lumpy, with both dimensions being a negative function of age and income.

Seniority and skill

Chen and Stratton (2026) studied the adoption of GitHub Copilot and Cursor using a proprietary dataset from Jellyfish covering approximately 200 million work events of 100,000 engineers at 500 firms. Using a staggered difference-in-differences design exploiting variation in the timing of firm-level license purchases, they found that AI adoption led to moderate productivity increases—an 8.5 percent increase in coding activity and 8.7 percent faster task completion—with no measurable quality declines. These productivity gains did not translate into increased output, changes in task composition, or effects on employment.

Importantly, the paper shows that AI adoption was far from random. First, individual take-up was not universal: even 18 months after firm-level adoption, only about half of engineers had begun using the tools. Second, when they examined heterogeneity by worker seniority, the point estimates suggested marginally larger effects for senior workers, though the differences were not statistically significant. Third, the productivity-to-output gap varied by firm type: firms whose core product is software saw significant increases in output and suggestive increases in employment, while firms that use software as an internal input (a retailer maintaining a website or a bank updating its internal trading platform, say) saw no change in either. This distinction reveals an organizational dimension of the adoption margin: whether firms can scale the output of their engineers determines whether individual productivity gains aggregate into anything concrete. The same tool, adopted at the same time, produces different outcomes depending on organizational context.

Identity

Delfino et al. (2026) ran a discrete choice experiment with Italian jobseekers and found that perceived "identity fit” i.e., whether a new skill feels compatible with a person’s sense of self, dominated re-skilling decisions, often outweighing beliefs about wages or employer demand. These results echo a foundational insight from Akerlof and Kranton (2000): when identity is at stake, people will leave money on the table to avoid actions that conflict with their self-concept. While this study is not directly about AI adoption, the implication is that if ‘AI user’ does not fit a worker’s professional identity, evidence about productivity gains alone may not move them. More research is needed in this space.

Training and Top-Down Support

The BCG and Manpower surveys both highlighted training disparities within organizations, with some employees feeling much more confident and capable of using AI tools than others due to top-down managerial actions. For example, the BSG survey finds that in organizations with leadership support: 82% of employees regularly use AI (compared to 41% without), 55% view GenAI as having a positive impact on their work (compared to 15% without), and 65% say that GenAI will have a positive impact on their career prospects (compared to 13% without). The same survey finds that only 25% of frontline employees actually have that support. This is survey data and hence these numbers should all be taken with a grain of salt (leadership support is likely not randomly distributed across firms); more research is needed in this space.

Organizational microstructure is another important dimension of AI adoption. Diaz, Neyra-Nazarrett, Ramirez, Sadun, and Tamayo (2025) use data from a car manufacturer, a quick-service restaurant chain, and a retailer, to show that variation in training participation is closely tied to middle managers' behavior. When a "high-training" manager takes over a team previously led by a "low-training" manager, training participation surges within weeks. The managerial traits that predict higher training participation (strong social skills, a desire to help employees achieve their goals, and a commitment to developing underperformers) are likely to matter just as much when the training in question is about AI. If so, the same AI training program, deployed within the same firm, can have different take-up rates depending on an employee's immediate manager.

How AI is Used

The previous section reviewed evidence of what variables are correlated with AI use. In this section we explore the different ways that AI is used. Documenting these patterns is critical for thinking about the complementarity of human and AI skills, as well as the impact of AI on human capital development: whether people offload learning to the tool, which may result in a failure of learning or even de-skilling, or whether they use the tool to learn more effectively. While most of the studies we have now do not explore the determinants of how people use AI, this will certainly be an important question going forward.

Shen and Tamkin (2026) gave 52 developers AI assistance to learn a new Python library. They found that those who delegated heavily scored below 40 percent on mastery tests, while those who stayed cognitively engaged (by asking for explanations and verifying output) preserved their learning. They also identified six patterns of interacting with AI. Three involved heavy reliance on AI—wholesale code delegation, AI-driven debugging, extensive code generation—and produced quiz scores below 40 percent. Three involved cognitive engagement—asking conceptual questions only, requesting explanations before implementing, verifying AI output against one’s own understanding—and preserved learning outcomes comparable to the no-AI control group. Critically, the AI didn’t make developers faster on average; the control group encountered and solved more errors themselves, building debugging skills in the process.

Anthropic released a report of how students use AI (Claude). STEM majors are overrepresented amongst AI users. As in the Shen and Tamkin (who works at Anthropic) study, the usage patterns can be divided into several categories: Direct Problem Solving, Direct Output Creation, Collaborative Problem Solving, and Collaborative Output creation.

Close to half (47%) of all interactions can be classified as Direct, as in students who use AI to fully generate an answer with little engagement. The smallest category of engagement was Collaborative Problem Solving.

The authors identify an inverted Bloom’s taxonomy of what types of tasks are delegated to AI. Claude was mostly completing higher order cognitive functions such as creation of new ideas and analysis, while rote tasks made up a minimal proportion of the assigned tasks. The authors of the report raise concerns that such cognitive offloading may impair human capital formation.

Beane, Hassell, Hopper, and Yegge (2025) study the personality characteristics of effective AI use in coding. They identify three dimensions: code reading velocity (quickly understanding what AI-generated code does), productive skepticism (knowing when to trust and when to verify), and architectural judgment (deciding how AI-generated components fit together). The study notes that these characteristics are not immutable personality traits, they are acquired through years of deliberate practice. Importantly, these dimensions are complements to AI access, which allows workers to be more productive with the tool than those who lack such characteristics.

The Role of Institutions in Adoption

In our view, one of the most important results from the above review is the extent to which adoption can be shaped by institutions. The controlled studies effectively subsidize the adoption margin: they prompt usage, provide training, define tasks, and protect experimentation time. In Brynjolfsson et al., the AI was embedded directly into the agent’s workflow; in Cui et al., developers were given a specific coding task with AI tools pre-configured; in Noy and Zhang, participants were shown how to use ChatGPT and given clear writing prompts.

As revealed in the surveys reviewed above, most real-world settings lack this scaffolding. Consider the Microsoft MCAPS Copilot rollout, documented in detail by Microsoft itself: there was initial enthusiasm and a spike in uptake when the tool was introduced. But daily active usage crashed shortly after because employees lacked protected time to learn and experiment. Without organizational scaffolding, many employees did not find it worthwhile to invest in learning the tool and adoption faltered. A deliberate change in management was necessary to get usage back on track. (Chen and Stratton’s finding that only half of engineers adopted AI tools even 18 months after firm-level rollout tells the same story at a larger scale).

The Microsoft study shows that adoption is not a fixed quantity. With little top-down guidance, adoption will likely be lumpy and concentrated amongst the already-skilled; with scaffolding in place, however, broader adoption will likely materialize, which can generate the sorts of productivity gains documented in the experimental studies.

You can see this in older work on IT more generally. Bloom, Sadun, and Van Reenen (2012) showed that U.S. multinationals were more productive with IT than European firms not necessarily because they had better technology but because they had organizational complements—management practices, human capital, and restructuring—that allowed them to realize the technology’s potential. Diaz et al. show that these complements operate at the managerial level, even within the firm: the same training program produces radically different participation depending on the middle manager mediating it. One difference with AI is that the adoption margin is not a one-time cost to optimize. The tools and workflows update continuously, meaning that the scaffolding for maintaining and facilitating effective adoption needs to be updated as well.

Summing up, in our view what the experimental studies likely measure is the potential productivity from AI, i.e., the treatment effect with the extensive margin held at its optimal level or (as is often the case) simply abstracted away from. Studies that measure endogenous adoption reflect the realized productivity in the presence of large variation on the adoption margin. The gap between the two is determined by who uses AI, how much they use it, how well they use it, and what support they receive.

Which Equilibrium?

There’s a bit of a puzzle in the data. Controlled studies often find equalizing effects of AI, but those that allow people to select into AI—keeping the adoption decision endogenous—find the opposite. Reviewing the existing research, we see a pretty consistent pattern: without the right scaffolding (training, top-down encouragement), adoption is lumpy and often correlated with characteristics that correspond to being already-advantaged. We are certainly nowhere close to an equilibrium regarding AI adoption. But these dynamics give us some basis to speculate where things might be heading. Here are two possibilities:

“Catch-up”: Adoption is broadened—through falling prices, better tools, training programs, and organizational investment—and AI diffusion becomes less lumpy and more uniform. The types of people who benefit most in the experimental studies (e.g. less experienced workers) begin using AI more frequently and more effectively in practice. The equalizing effects documented in controlled settings eventually show up in the aggregate data.

“Lock-in”: Adoption remains concentrated among those who are already ahead. Early adopters, who tend to be higher-skilled, higher-income, and in better-resourced organizations, compound their head start through network effects, skill accumulation, and workflow optimization. This is the Matthew effect in action.

Anxieties about AI often center around questions of distribution and inequity of wealth. To that end, the “lock-in” equilibrium is one we would presumably want to avoid. The good news is that we have at least some initial evidence of what works to facilitate broad, effective adoption: top-down guidance and training programs. We will be keeping track of such programs and the results as the evidence evolves.

References

• Akerlof, George A., and Rachel E. Kranton. “Economics and Identity.” Quarterly Journal of Economics 115, no. 3 (2000): 715-753.

• Beane, Matt, Jonathan Hassell, Brendan Hopper, and Steve Yegge. “Human Vibes: How Developers Can Navigate a Job Market That’s Stretching at Both Ends.” IT Revolution (2025).

• Becker, Joel, Nate Rush, Beth Barnes, and David Rein. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” arXiv:2507.09089 (July 2025).

• Bick, Alexander, Adam Blandin, and David J. Deming. “The Rapid Adoption of Generative AI.” NBER Working Paper 32966 (September 2024).

• Bloom, Nicholas, Raffaella Sadun, and John Van Reenen. “Americans Do IT Better: US Multinationals and the Productivity Miracle.” American Economic Review 102, no. 1 (2012): 167-201.

• Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. “Generative AI at Work.” Quarterly Journal of Economics 140, no. 2 (2025): 889-942.

• Carvajal, Daniel, Catalina Franco, and Siri Isaksson. “Will Artificial Intelligence Get in the Way of Achieving Gender Equality?” NHH Discussion Paper SAM 03/24 (October 2024).

• Chen, Fiona, and James Stratton. “Artificial Intelligence in the Firm.” Working paper (January 2026).

• Cui, Kevin Zheyuan, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng, and Tobias Salz. “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.” Working paper (2025).

• Delfino, Alexia, Andrea Garnero, Sergio Inferrera, Marco Leonardi, and Raffaella Sadun. “Unwilling to Reskill? Experimental Evidence from Real-World Jobseekers.” NBER Working Paper 34633 (January 2026).

• Dell’Acqua, Fabrizio, Edward McFowland III, Ethan Mollick, et al. “Navigating the Jagged Technological Frontier.” HBS Working Paper 24-013 (September 2023).

• Diaz, Brayan, Andrea Neyra-Nazarrett, Julian Ramirez, Raffaella Sadun, and Jorge Tamayo. “Training within Firms.” NBER Working Paper 33670 (April 2025).

• Gambacorta, Leonardo, Han Qiu, Shuo Shan, and Daniel Rees. “AI and Productivity: Evidence from CodeFuse.” BIS Working Paper 1208 (2024).

• Humlum, Anders, and Emilie Vestergaard. “Unequal Adoption of ChatGPT Exacerbates Existing Inequalities.” Proceedings of the National Academy of Sciences (2025).

• Noy, Shakked, and Whitney Zhang. “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science 381, no. 6654 (2023): 187-192.

• OECD. “Bridging the AI Skills Gap: Is Training Keeping Up?” OECD Publishing (2025).

• Otis, Nicholas G., Rowan Clarke, Solène Delecourt, David Holtz, and Rembrand Koning. “The Uneven Impact of Generative AI on Entrepreneurial Performance.” HBS Working Paper 24-042 (December 2023).

• Paradis, Elise, et al. “How Much Does AI Impact Development Speed? An Enterprise-Based Randomized Controlled Trial.” arXiv:2410.12944 (2024).

• Shen, Judy Hanwen, and Alex Tamkin. “How AI Impacts Skill Formation.” arXiv:2601.20245 (January 2026).