AI Recommendations: This Time It’s Personal - Harvard University

News

Key Takeaways

- A new Harvard study shows AI recommendations that can adjust to individual users could combat growing over-reliance on AI for decision-making support.

- Researchers tested an experimental AI recommendation model with more than 1,000 human participants. They found that reinforcement learning helped achieve higher accuracy in decision-making.

From doctors diagnosing symptoms to judges intervening in court cases, humans make complex decisions every day. Increasingly, artificial intelligence tools are being used to help in those decisions.

There’s an insidious downside to this type of “help.” Research shows that over-reliance on AI for decision-making can lead to worse or inaccurate choices, and moreover, loss of expertise: over time, a person learns less about the subject, creating a cycle of over-reliance and under-achievement.

Computer scientists at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) offer a potential solution to this cycle. In a recent paper, they argue that AI shouldn’t offer one-size-fits-all decision support, as is typical today, but rather should be adaptive, or able to adjust to the situation and to the uniqueness of each user.

Now, they've developed an AI recommendation model that incorporates reinforcement learning, a machine learning method in which an AI system learns which actions to take based on continuous feedback. This model doesn’t just spit out answers — it decides in the moment how and to what extent to help the human.

In online experiments with more than 1,000 participants, the researchers demonstrated that reinforcement learning improved human-AI performance more than any other type of AI support that’s used today.

The research, led by Zana Buçinca, a recent Harvard computer science Ph.D. graduate and current MIT faculty member, is published in ACM Transactions on Computer-Human Interaction and will be presented later this year at the ACM Symposium on User Interface Software and Technology (UIST).

“Given this worrisome trend of human over-reliance on AI, we wanted to instead design AI that accounts for and optimizes for how the human processes its advice,” said Buçinca, who co-authored the work with former advisor Krzysztof Gajos, the Yahn W. Bernier and N. Elizabeth McCaw Professor of Computer Science at Harvard, and Maja Malaya, a student at Technical University of Łódź in Poland.

How humans and AIs make decisions together

Earlier work by Buçinca and colleagues tested how humans and AIs work together to make decisions. They uncovered that humans tend to over-rely on AI recommendations by accepting incorrect suggestions, even when they could have made the right decision on their own.

Their previous work also uncovered individual differences in how people received information from AI. Some people naturally have more “need for cognition,” — that is, they enjoy and are motivated by analytical thinking – while others desire less to think deeply. This finding underscored the need for AI assistance to adjust to situational context, including the individual characteristics of the human decision-makers.

In the new system, the reinforcement learning agent observes the human-AI as a pair, accounting for the person’s dynamic assessment of the skill, their need for cognition, and how confident the AI model is. It chooses from several interaction strategies, such as showing a full recommendation; providing a partial explanation; or withholding an answer altogether. The system is trained to choose among these options to maximize a specified objective, whether that’s immediate accuracy to the task, or longer-term human learning.

To evaluate their approach, the team designed a decision task modeled on real health-care data. Human participants were shown vignettes of fictitious patients with different health needs and goals and were then asked to select the most appropriate prescription for exercise, like pilates, weight-lifting, etc.

Across two online experiments of 316 and 964 participants each, participants first completed a baseline assessment to measure their initial skill on the task, and they answered survey questions to measure need for cognition. They then made a series of decisions about different patients, but under different sets of conditions, such as reinforcement learning optimized for accuracy; reinforcement learning optimized to support longer-term learning; and baseline non-adaptive AI supports that always provided decision recommendations accompanied by explanations.

In both experiments, people interacting with reinforcement learning optimized for accuracy achieved significantly higher decision accuracy than those with non-adaptive support that gave them the answers. In many cases, the learned policies enabled human-AI complementarity, where the human-AI team outperformed both humans alone and the AI system alone.

The findings have implications for emerging AI regulation, Buçinca added. Many policy frameworks, including the European Union AI Act, call for human oversight of high‑stakes AI systems with the implicit assumption that adding a human decision-maker on top of an algorithm will mitigate errors.

Buçinca and Gajos’s research shows that, without intentional design of how the human and AI interact, human oversight of AI systems could be undermined. Reinforcement learning could be a practical tool to discover assistance strategies that both improve accuracy and support human skills, they contend.

“Our paper shows that psychological needs are not just intangible things beyond the direct grasp of computer scientists, but rather, they are things we can model and incorporate as objectives for our optimization algorithms,” Gajos said. “In other words, as respectable engineers, we can optimize for human happiness.”

Center for Human-driven AI Research and Methods at Harvard

The study is part of the research agenda of a new center at Harvard, the Center for Human-driven AI Research and Methods, or CHARM. The initiative brings together different areas of science to develop AI systems that advance human values, rather than simply automating tasks. Within CHARM, researchers like Gajos and Buçinca are focused on “worker-centric AI”: systems that not only help people do their jobs better today, but also support their long-term competence, autonomy, and sense of meaning at work.

The research received federal support from the National Science Foundation under grant No. IIS-2107391 and by the Office of Naval Research under agreement No. N00014-24-1-2726. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or the Office of Naval Research.

Topics: AI / Machine Learning, Computational Science & Engineering, Computer Science, Research

Cutting-edge science delivered direct to your inbox.

Join the Harvard SEAS mailing list.