Workers need AI interaction skills to unlock GenAI productivity gains

Generative AI can raise performance in education and knowledge work, but the gains are not evenly shared, according to a new study by Bharat Anand of New York University and Lihi Idan of Texas A&M University, whose randomized experiment finds that workers benefit most when they know how to prompt, verify and apply AI outputs effectively.

The working paper, Generative AI and the Productivity Divide: Human–AI Complementarities in Education and Knowledge Work, published on arXiv, shows that access to large language models (LLMs) improved average task performance while opening a new divide around AI Interaction Competence, a practical skill set that determines whether users can turn AI access into measurable productivity gains.

AI access improves performance, but the gains are uneven

The findings challenge the idea that simply giving people access to AI tools will produce uniform productivity gains. Access matters, but the experiment shows that access alone does not determine who improves.

The research was designed to mirror the conditions faced by early-career knowledge workers. Participants had to quickly learn unfamiliar technical material, synthesize information and apply it under time pressure. The experiment included 179 participants at Texas A&M University, mainly from engineering disciplines, with a small number of business students also taking part.

Participants were first assessed on baseline knowledge, learning preferences, self-rated knowledge of machine learning and LLMs, and other background characteristics. They were then assigned to a self-study task focused on learning about large language models. One group used traditional resources, including articles, textbooks, videos, peers and recorded training materials. Another group used the free version of ChatGPT as its learning tool.

The self-study period lasted three days, with participants told to spend at least three hours a day learning. A post-intervention exam then measured how much they had learned and how well they could apply the material. The results showed a clear advantage for the AI group. Participants with LLM access achieved an average post-intervention score of 0.56, compared with 0.48 among those using traditional resources. The treatment effect remained statistically significant after controls for baseline performance, GPA, gender and study preferences.

The advantage also appeared in engagement. Twenty participants in the traditional-resource group dropped out before the post-intervention exam, compared with nine in the LLM group. The study treats this imbalance as an important signal: participants did not just perform better with AI access, they also appeared more willing to continue when AI tools were available.

Survey results pointed in the same direction. LLMs were the most preferred learning resource, selected by 69 percent of participants. That placed them ahead of lectures, YouTube tutorials and textbooks. The finding reflects how quickly generative AI has become a favored study and work aid among people preparing for knowledge-intensive careers.

The benefits were uneven. Some participants gained substantially from AI access, while others saw limited or even negative marginal returns. The dividing line was not GPA, prior knowledge or conventional academic achievement. It was AI Interaction Competence.

AI Interaction Competence becomes the new productivity fault line

The paper defines AI Interaction Competence as the ability to use generative AI systems effectively and reliably. It includes three core skills: formulating precise prompts, filtering and verifying outputs for accuracy and relevance, and iterating with the model while combining machine suggestions with human judgment.

GPA did not significantly shape who benefited most from the AI treatment. Prior topic knowledge also did not significantly predict who gained more from LLM access. This means generative AI did not simply reward participants who were already stronger students or more familiar with the material.

Instead, high-AIC participants realized the largest gains. They could ask better questions, evaluate responses more critically and turn AI-generated material into stronger learning outcomes. Low-AIC participants struggled to extract the same value. In some cases, the technology's flexibility appeared to create added difficulty because users had to know what to ask, how to follow up and when to distrust a response.

Generative AI is a new form of human-capital divide, the study insists. In older models of workplace performance, grades, prior knowledge and technical training often served as markers of ability. In AI-mediated work, those markers may become less predictive. The more important distinction may be whether workers can collaborate effectively with a probabilistic machine that produces fluent but not always reliable output.

Many companies are treating AI adoption as a technology deployment issue: buy tools, provide access and wait for productivity to rise. The study suggests that approach is incomplete. Generative AI is also a capability-design problem. Organizations must develop the skills and routines that help workers use AI consistently.

The research also warns against confusing preference with competence. Participants who preferred LLMs generally had higher AIC, but the relationship was only modest. Some users liked AI despite having relatively weak interaction skills. Others had strong interaction skills but still preferred lectures, peer learning or other methods. This means self-reported comfort with AI is not a reliable measure of actual AI ability.

Employees who are enthusiastic about AI may still produce weak or unchecked outputs. Employees who are hesitant may still be capable of using AI well if given the right context. Firms that rely on informal enthusiasm, rather than measured competence and training, risk uneven quality and hidden performance gaps.

The study also found that novice learners can benefit from generative AI, but only under the right conditions. Participants with low prior knowledge gained more from LLM-assisted study when they had high AIC. Novices with low AIC did not receive the same lift, even though they had more room to improve. In other words, AI can help weaker learners catch up, but not automatically. The ability to interact with the system shapes whether the tool becomes an equalizer or a divider.

Generative AI raises average performance while creating a new axis of inequality. The divide is no longer only between those who have access to AI and those who do not - it's between those who can use AI well and those who cannot.

Training and scaffolding can reduce unequal AI outcomes

The study also tests practical interventions that organizations can use to narrow it. The most important was scaffolding: a simple conceptual roadmap that gave novice participants a structured sequence of topics and helped them organize their learning.

The scaffolding intervention improved performance among novices and reduced outcome variance. Scaffolded novices scored an average of 0.45, compared with 0.38 among unguided LLM novices. The more important finding was the reduction in dispersion. The standard deviation among scaffolded novices was 0.14, compared with 0.22 among unguided novices, suggesting that structured guidance helped weaker users close part of the gap.

The effect was strongest for participants with low AI Interaction Competence. High-AIC users already had the skills to structure prompts, sequence learning and verify outputs. Low-AIC users benefited more from a roadmap because it reduced the burden of navigating an open-ended AI environment. The intervention did not replace skill, but it helped compensate for weaker skill.

This finding is highly relevant for workplace AI adoption. Many workers may not need long formal courses to use AI better. They may need short, targeted training and simple operating procedures: prompt templates, review checklists, concept maps, verification steps and standard workflows for using AI-generated material. These tools can help convert AI from an open-ended assistant into a more reliable productivity system.

On the other hand, simply giving users more time did not improve outcomes. Novices assigned to study four hours per day instead of three did not perform significantly better than unguided LLM novices. More exposure to AI did not solve the problem when users lacked effective strategies. Without structure, additional time produced little value and, in some cases, frustration.

Firms should not measure AI adoption by access rates alone. They should measure whether employees can use AI outputs safely, accurately and productively. Training should focus on prompting logic, error detection, verification habits and synthesis. Workflows should specify when AI can be used, how outputs should be checked and how human judgment remains accountable for final decisions.

The paper estimates that GenAI access generated an approximate 17 percent productivity lift, consistent with other evidence from professional work settings. But it also found that dispersion among novices increased sharply under standard LLM use. Scaffolding reduced that dispersion by nearly 40 percent without lowering average performance. This combo makes the intervention crucial: it preserved productivity while making outcomes more consistent.

The findings also carry a warning for education systems. As students increasingly use generative AI for learning, institutions may need to teach AI interaction directly. Traditional exams, grades and prior coursework may not capture the skill that determines who learns effectively with AI. If AI becomes a routine part of study and work, then prompting, verification and critical use of generated content will become foundational learning skills.

Organizations that pair AI access with micro-training, scaffolding and clear workflow rules are more likely to capture the technology's benefits across teams. Those that simply hand workers a chatbot may still raise average productivity, but they risk creating a workplace in which the gains concentrate among skilled AI users while others fall further behind.