- The paper demonstrates that the AI tool raises materials discovery by 44%, leading to 39% more patent filings and a 17% increase in new product prototypes.
- It reveals that the AI tool fosters radical innovation by generating novel compounds and new technical terms, while improving R&D efficiency by 13-15%.
- The paper highlights heterogeneous impacts among scientists, emphasizing that effective AI evaluation relies on strong domain expertise.
This paper investigates the impact of artificial intelligence on innovation within the R&D lab of a large U.S. firm specializing in materials science. The paper exploits the randomized introduction of a new AI tool for materials discovery to over 1,000 scientists across 221 teams. The AI tool, a graph neural network, is trained on existing materials to generate novel candidate compounds predicted to possess desired properties, aiming to partially automate the traditional trial-and-error process.
The analysis reveals a substantial positive impact of the AI tool on the innovation pipeline. AI-assisted scientists discover 44% more materials, which translates into a 39% increase in patent filings and a 17% rise in downstream product prototypes incorporating these new materials. The effects on discovery and patenting emerge within 5-6 months, while the impact on product prototypes lags by over a year. Importantly, the AI tool does not compromise quality; discovered materials show statistically significant improvements in average quality for both atomic and large-scale properties (9-13% increase in overall quality index).
The paper also finds that the AI tool fosters more radical innovation. Model-generated materials possess more novel chemical structures (0.4 standard deviation decrease in similarity to existing compounds). Patents filed by treated scientists introduce more new technical terms (a 22% increase in the share of new terms), which is a leading indicator of transformative technologies. Furthermore, the share of product prototypes representing entirely new product lines (rather than improvements) increases by 3 percentage points (a 22% rise from the baseline of 13%), suggesting that AI helps overcome materials bottlenecks for novel product ideas. Accounting for input costs, the tool boosts R&D efficiency by 13-15%.
However, the benefits of the AI tool are distributed unequally among scientists. Based on pre-treatment productivity, the bottom third of researchers see minimal gains, while the output of top-decile scientists increases by 81%. This disparity leads to a more than doubling of the 90:10 performance inequality. This heterogeneous impact is primarily driven by differences in scientists' ability to evaluate AI-generated candidate materials.
Investigating the mechanisms, the paper shows that the AI tool dramatically changes the research process. It automates 57% of "idea generation" tasks (time allocated drops from 39% to 16%), reallocating scientists to evaluating model-produced candidates (time allocated increases from 23% to 40%). Time spent on experimentation also increases (from 37% to 44%) due to the higher volume of candidates. Scientists with a comparative advantage in judgment tasks shift more effort towards evaluation.
Differences in judgment ability explain the heterogeneous impact. Top scientists leverage their domain knowledge and expertise to effectively prioritize promising AI suggestions, leading to a steeper "discovery curve" (higher probability of a tested material being viable earlier in the testing sequence). In contrast, scientists with weaker judgment struggle to distinguish viable candidates from false positives, testing materials no better than random chance (a flat discovery curve). This difference in prioritization efficiency explains over three-quarters of the variation in post-treatment productivity. Survey evidence indicates that domain knowledge (scientific training, experience with similar materials, intuition) is crucial for effective judgment of AI suggestions.
The findings challenge the notion that AI will render domain knowledge obsolete, suggesting instead a strong complementarity between algorithms and human expertise, particularly in evaluation tasks. The paper highlights the growing importance of judging model predictions as a key skill in AI-augmented research.
Beyond productivity, the AI tool impacts scientist wellbeing. Despite increased productivity, 82% of scientists report an overall decline in job satisfaction. The primary reasons cited are skill underutilization (73%) and reduced creativity or increased repetitiveness of tasks (53%). This challenges the view that AI solely automates tedious tasks, suggesting it can automate creative and engaging aspects of work. Experience with the tool also changes scientists' beliefs about AI, increasing belief in its productivity-enhancing potential while keeping concerns about job loss stable. A significant majority (71%) plan to reskill, recognizing that AI changes the skills needed to succeed.
The paper also notes organizational adaptation. After the paper period, the firm adjusted its hiring and firing criteria to favor scientists with strong judgment, indicating that the estimated effects may understate the longer-run impact as the workforce composition shifts. This highlights the potential for organizational change to amplify or mediate the effects of AI.
In summary, the paper provides causal evidence that AI can significantly accelerate real-world scientific discovery and innovation, leading to more and more novel outputs. However, its effectiveness is contingent on human expertise, specifically the ability to critically evaluate model outputs. This creates disparities in benefits among researchers and necessitates a shift in required skills, while also presenting challenges for scientist job satisfaction by automating creative tasks.