AI, Agency, and the Future of HCI

Artificial intelligence is making computers more capable of interpreting language, generating images, predicting what we want, and acting on our behalf. That creates enormous possibilities for human-computer interaction. It also raises a question that I think should sit at the center of HCI: As computers become more capable, what decisions should remain with people?

I spent this quarter in my UC Irvine HCI+D program exploring that question through the design of a speculative communication system for people who use implanted speech brain-computer interfaces, or BCIs. The project began with a fairly simple idea. Current speech BCIs can decode attempted speech and turn neural activity into text (Card et al., 2024; Willett et al., 2023). What if that decoded speech could also be used to create images, giving a person another way to communicate emotion, memory, humor, or desires?

Three prototypes grew out of this work, with each one building on the last. Whose Words focuses on speech-to-text, which current BCIs can already do. I built this to help the average able-bodied individual get a taste of the trade-off between absolute accuracy in choosing one’s words and accepting what the AI suggests from the speech motor cortex signal. BCI Emotion Capture uses those decoded words to create an image, giving someone a way to show a feeling, memory or desire instead of having to spell it out. The Speech-to-Image Communication Flow is more like a map and looks at the bigger system around the technology: who makes decisions, who approves what gets shared, when another person can help, and what the AI should or should not do on its own.

The project eventually became less about image generation and more about something larger: how people and AI should work together when technology is participating in an act as personal as communication. That shift changed my view of the future of HCI.

I no longer think our primary challenge is designing systems that can “do more for people.” We also need to design systems that know when not to act, when to ask, when to wait, and when a decision belongs to a human being.

Research in speech neuroprostheses has made a lot of progress. Scientists have demonstrated systems that decode attempted speech from neural activity with increasing accuracy and speed (Card et al., 2024; Willett et al., 2023). For a person who has lost the ability to speak because of ALS, stroke, or another neurological condition, restoring access to language could be life-changing. But conversation is not made of words alone.

People communicate through facial expression, gesture, timing, tone, shared references, humor, visual imagery, and emotion. A sentence on a screen can communicate what someone said while leaving out much of how they wanted to say it.

My initial design explored whether AI-generated images could help fill some of that gap. A speech BCI would decode the person’s intended words. An AI system could then create several image options based on those words. The person would decide whether any of the images represented what they meant and could approve, revise, reject, or regenerate them before sharing anything.

At first, I treated this primarily as an interaction design problem. By the end of the quarter, I understood it as a collaboration design problem. The important question was not simply whether AI could create a useful image. It was who should decide what the image means, who can approve it, who can receive it, what happens when the system is wrong, and whether another person should ever be allowed to assist.

Research on human-AI interaction has shown that people need to understand what an AI system can do, where it may fail, and when they should rely on it (Amershi et al., 2019). Research on judge-advisor relationships also distinguishes between receiving advice and giving up authority to the advisor (Sniezek & Van Swol, 2001). That distinction became central to my design.

In the workflow I developed, the AI has several narrow jobs. One part decodes neural signals into candidate words. Another generates image options. Another delivers an approved message. The person using the BCI decides whether the image actually represents what they intended. The AI does not get to make that final decision.

This becomes especially important in light of recent work on agency and speech ownership in neurotechnology (Freudenburg et al., 2024; Sankaran et al., 2023). If technology participates in producing someone’s communication, what makes that communication legitimately theirs?

An AI that drafts an email, summarizes a medical conversation, suggests a reply, edits a photograph, recommends a decision, or generates language in someone’s voice is doing more than completing a task. It may be helping shape how that person is represented to others. The future of HCI therefore needs to make a clear distinction between assistance and authorship.

Designing for a person who cannot reliably use their hands also changed how I thought about accessibility. In my concept, the user would make choices through the BCI itself. They could review image options, choose a recipient from friends and family, approve a message, ask the system to try again, or stop the interaction.

Accessibility, in this context, is not simply adding another input method. It means preserving the same meaningful choices even when someone interacts differently. This matters because accessibility features can sometimes reduce autonomy while appearing to provide support.

A caregiver may be able to complete a task more quickly. A clinician may understand the technology better. An AI may be able to predict what the person probably wants. But convenience for the system or the people surrounding the user should not replace the user’s authority.

There may be situations when a family member or communication partner should help. My design allows that possibility only when the person has chosen it, and the system makes the assistance visible. That principle applies broadly: support should increase a person’s ability to participate, not make their participation unnecessary.

One of the most important changes in my thinking involved evaluation. Speech BCI research appropriately measures technical performance: decoding accuracy, word error rate, communication speed, calibration, and reliability over time (Card et al., 2024; Willett et al., 2023). Those measures are necessary, but they are not enough to evaluate the kind of human-AI system I am proposing.

We would also need to ask whether the generated image actually matches what the person meant. How many attempts did it take? Did the person feel in control of what was sent? Did the recipient understand the message? Could the person correct a mistake? Did the system increase the user’s joy? And ultimately: did the technology help the person communicate something they could not communicate as easily before?

The impact might not appear as a better error rate. It might appear in a teenager telling a joke, a parent sharing a memory, a patient showing fear without spelling out a sentence, or someone choosing an image that expresses a memory they want to share. Those outcomes are harder to quantify. They are also why the technology matters.

I do not think there is a perfect design that eliminates every conflict. More control can mean more steps. Faster automation can mean fewer opportunities to correct the AI. Personalization can make a system more useful while also requiring more data, and therefore time to set up. Logging system decisions can improve accountability while creating the possibility of surveillance. A generated image may communicate emotion more effectively than text while also introducing ambiguity.

These are not problems we can solve with a single interface pattern. They require ongoing involvement from the people affected by the technology: people who use brain-to-computer interface technology or other kinds of augmentative and alternative communication systems (eye-gaze systems, switch-access systems, gestures, signs, or symbol systems), family members, speech-language pathologists, clinicians, disability advocates, and others who understand the realities of communication outside a laboratory. Responsible innovation should therefore be treated as a continuing design activity, not something checked at the end of development.

My vision for HCI is not a future in which computers disappear or AI anticipates every need. It is a future in which increasingly powerful systems make human participation more meaningful rather than less necessary.

AI should be able to reduce repetitive work. It should help people explore possibilities, translate information, create alternatives, notice patterns, and accomplish things that were previously difficult or impossible. But when a decision concerns a person’s meaning, identity, relationships, consent, or representation, HCI should make sure that person still has the final decision.

For me, that is the larger lesson of this project. The measure of a good human-AI system should not simply be how much intelligence we can put into the technology. It should be whether the technology helps people express more, participate more, understand more, and retain authority over the parts of life that should remain theirs. That is the future of HCI I want to help create.

References

Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for human-AI interaction. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (pp. 1–13). Association for Computing Machinery. https://doi.org/10.1145/3290605.3300233

Card, N. S., Wairagkar, M., Iacobacci, C., Hou, X., Singer-Clark, T., Willett, F. R., Kunz, E. M., Fan, C., Vahdati Nia, M., Deo, D. R., Srinivasan, A., Choi, E. Y., Glasser, M. F., Hochberg, L. R., Henderson, J. M., Shahlaie, K., Stavisky, S. D., & Brandman, D. M. (2024). An accurate and rapidly calibrating speech neuroprosthesis. New England Journal of Medicine, 391(7), 609–618. https://doi.org/10.1056/NEJMoa2314132

Freudenburg, Z. V., Berezutskaya, J., & Herbert, C. (2024). Editorial: The ethics of speech ownership in the context of neural control of augmented assistive communication. Frontiers in Human Neuroscience, 18, Article 1468938. https://doi.org/10.3389/fnhum.2024.1468938

Sankaran, N., Moses, D., Chiong, W., & Chang, E. F. (2023). Recommendations for promoting user agency in the design of speech neuroprostheses. Frontiers in Human Neuroscience, 17, Article 1298129. https://doi.org/10.3389/fnhum.2023.1298129

Sniezek, J. A., & Van Swol, L. M. (2001). Trust, confidence, and expertise in a judge-advisor system. Organizational Behavior and Human Decision Processes, 84(2), 288–307. https://doi.org/10.1006/obhd.2000.2926

Willett, F. R., Kunz, E. M., Fan, C., Avansino, D. T., Wilson, G. H., Choi, E. Y., Kamdar, F., Glasser, M. F., Hochberg, L. R., Druckmann, S., Shenoy, K. V., & Henderson, J. M. (2023). A high-performance speech neuroprosthesis. Nature, 620(7976), 1031–1036. https://doi.org/10.1038/s41586-023-06377-x