Does Generative AI Improve Productivity?

Generative AI’s Productivity Promise—and Its Hidden Pitfalls

Generative artificial intelligence (genAI) promises to “turbo-charge” labor productivity, yet real-world evidence paints a more complex picture. Recent studies in customer support, software programming, and writing consistently reveal that productivity gains from genAI can vary significantly based on worker skill levels. Specifically, these AI tools disproportionately benefit less-experienced or lower-skilled workers, narrowing existing skill gaps.

Building on this uneven pattern, new research reveals that the genAI technology can even reduce productivity for some expert users. A randomized controlled trial by the AI research nonprofit METR found that experienced open‑source developers using an AI coding assistant finished tasks approximately 19 percent slower while believing they were faster.

The underlying cause may stem from the tension between learning from AI (using AI inputs to enhance one’s skills) and learning about AI (understanding when and how to rely on AI), a framework my coauthors and I proposed in a related study. Highly skilled workers, often (over)confident in their established methods, may discount valuable AI suggestions or expend additional cognitive effort reconciling these inputs with their existing mental models. In these situations, prior expertise becomes a liability, turning AI’s potential productivity boost into a hidden productivity tax.

Even novices, who initially benefit significantly, may not be clear winners. Heavy reliance on AI shortcuts can crowd out the trial‑and‑error learning that builds deep expertise, potentially slowing progression into senior or managerial roles. At the firm level, blindly accepting AI outputs (especially in software development) risks accumulating “technical debt” that complicates future maintenance and expansion. Early gains, therefore, may mask longer‑term capability gaps for individuals and costly rework for organizations.

Such uneven outcomes partly explain growing caution among companies adopting generative AI. A 2024 Wakefield Research survey indicated that 98 percent of Fortune 1000 executives have paused at least one generative AI initiative due to concerns around security, accuracy, and unclear guidelines. Some firms have even scaled back or halted AI pilot projects following initial enthusiasm.

Beyond the On‑Off Switch: A Playbook for Making Generative AI Work

Should we pause adopting genAI for some? Not necessarily, but it does mean rethinking how we deploy and interact with it. Below are key strategies that both companies and individuals can begin considering.

Upskilling and AI Literacy

Forward-looking organizations proactively embed AI literacy into employee training programs. Effective training initiatives typically cover three pillars:

1. Using AI well: Crafting effective prompts and iterating queries for better output.
2. Interpreting AI output: Fact‑checking, detecting hallucinations, and applying domain judgment.
3. Respecting limits: Safeguarding privacy, avoiding confidential data leaks, and following ethical use policies.

Rethinking AI Design

Companies should tailor their AI tools and workflows to the people using them. In other words, avoid a one-size-fits-all approach. Some of my related research highlights that the optimal form of AI assistance can depend on a user’s skill level. Practical steps may include:

1. Match tool to role: Determine which teams benefit from advanced generative AI versus those better suited to transparent, explainable machine learning models.
2. Customize workflows: Integrate AI where it complements but does not disrupt existing processes, recognizing that workers may misjudge their productivity gains.
3. Organizational redesign: When tasks fundamentally change due to genAI, proactively restructure roles and workflows. For instance, create positions dedicated to oversight and quality assurance, enforcing safeguards such as mandatory code reviews or human approvals in high-risk contexts.

Continuous Feedback and Adaptive Learning

Companies and individuals should also adopt a long-term view, recognizing initial productivity slowdowns among highly skilled workers as potentially temporary. Evidence from a related driving-assistance AI study shows that, over time, experienced users adapted their behaviors to leverage AI advantages. Companies can facilitate this adaptive learning by gathering regular employee feedback about how AI integrates with workflows, using these insights to iteratively refine tool design and training programs. Employees should become adept at managing and guiding AI tools themselves.

In summary, the rise of generative AI is not purely positive or negative for workers. Rather, it represents a growth opportunity. Those who engage in this process and learn to collaborate with AI will be empowered to achieve even greater outcomes.