Artificial intelligence is becoming increasingly powerful but bigger does not always mean better for every application. A Henry Xie 17-year-old student from Portland, Oregon, is exploring whether smaller AI models can deliver one of the qualities people often associate with advanced chatbots: empathy.

Henry Xie, a student at Westview High School, has developed a method designed to transfer empathetic behaviour from large language models into smaller, more efficient AI systems. His project, titled “Distilling Empathy From Large Language Models,” was selected for the 2026 Regeneron Science Talent Search.

The idea addresses an important challenge in the AI industry: large language models can require significant computing resources, while smaller language models are cheaper and easier to deploy but can struggle with nuanced human interactions.

The problem with smaller AI models

Large language models, or LLMs, have become increasingly capable of understanding context and generating responses that appear sensitive to a person’s emotions. But running sophisticated models can require substantial computing power.

Smaller language models, or SLMs, offer a different proposition. They can be faster, less expensive and more practical for devices with limited resources, including smartphones and other everyday applications.

The trade-off is that smaller models can sometimes struggle with qualities such as emotional sensitivity and empathetic communication.

Xie’s research asks whether it is possible to take one specific capability from larger models—empathy—and transfer it to smaller systems without having to recreate the enormous computing infrastructure behind the biggest AI models.

How Henry Xie’s method works

Xie’s approach uses larger AI models such as ChatGPT and Gemini as sources of examples rather than attempting to build a massive model from scratch.

The first stage involves collecting examples of empathetic responses generated by large language models. These examples are grounded in psychological theory and are used to teach smaller models what stronger empathetic responses can look like.

The second stage uses targeted prompts to help the smaller models distinguish between better and weaker empathetic responses.

The process is essentially a form of knowledge distillation: instead of making the smaller model learn everything a huge AI system knows, researchers can focus on transferring a particular capability.

For Xie, that capability is empathy.

The results caught attention

According to the Society for Science, the smaller models trained using Xie’s approach were judged to be more empathetic than untrained smaller models in at least 90% of cases.

That result is significant because the project is not attempting to make a small model as powerful as a frontier AI system across every task. Instead, it focuses on improving one characteristic that can be particularly important when AI interacts directly with people.

If similar techniques can be developed further, smaller models could potentially provide more human-sensitive responses while retaining their advantages in speed, cost and deployment.

Why AI empathy matters

The question of AI empathy goes beyond making chatbots sound polite.

People increasingly use AI systems to discuss personal problems, ask for advice, communicate with digital assistants and navigate emotionally sensitive situations. In those settings, the way an AI responds can influence how useful—and how trustworthy—the interaction feels.

Xie became interested in the issue after observing what he described as growing confrontation and declining empathy during the pandemic. As AI became more integrated into everyday life, he believed the technology should be developed with human interaction in mind.

That motivation led him to explore whether AI could become not only more capable, but also more considerate in its communication.

Importantly, an AI model producing an empathetic-sounding response does not mean that it actually understands human emotions in the way a person does. Empathetic language can improve communication, but questions around accuracy, safety and appropriate use remain.

More than just an AI project

Xie’s interests extend beyond artificial intelligence.

At Westview High School, he leads the computer science club and is also part of the varsity swim team. He has additionally been a three-time semifinalist in the CyberPatriot National Youth Cyber Defense Competition, reflecting his wider interest in cybersecurity and computing.

He is also the co-founder and president of Youth for Empathetic AI, a nonprofit focused on bringing students and researchers together around ethical and human-centred artificial intelligence.

His work has also appeared in academic research settings. A paper on distilling empathy from large language models involving Xie and researchers from Portland State University was presented through the SIGDIAL 2025 research community.

Could smaller AI become more human-friendly?

One of the most interesting aspects of Xie’s work is its focus on efficiency.

The AI industry has often pursued larger models, more parameters and increasingly powerful computing infrastructure. Xie’s project takes a different route: rather than simply making models bigger, it asks how useful capabilities can be transferred into models that are easier to run.

That could become increasingly relevant as AI moves from cloud-based services into phones, personal computers and other devices.

Smaller models cannot necessarily match the overall capabilities of the largest systems, but they can offer advantages in latency, cost and deployment. Adding better emotional sensitivity to those systems could make them more useful in everyday interactions.

For a 17-year-old researcher, the project represents an unusually focused contribution to a much larger debate about the future of artificial intelligence.

And its central idea is simple: perhaps making AI better does not always mean making it bigger. Sometimes, it may mean teaching smaller systems the capabilities that matter most.

Disclaimer: The results described above are based on the project information published by the Society for Science and related research records. An AI system producing more empathetic responses does not mean it possesses human emotions or genuine emotional understanding. Further research is needed to determine how well such techniques perform across different models, users and real-world situations.

Learn more about our site on contactarchive2in.com

Contact us on our email id contactarchive2in@gmail.com