Human–AI Interaction Researcher and Lecturer, UCSC Silicon Valley Extension
Evaluating LLM and agent systems in real use
Santa Clara, California
I evaluate LLM-based and agentic AI systems in real use: whether their outputs and reasoning hold up, where they fail, and when people should rely on them. I study how the way AI presents its reasoning changes people’s reliance and accuracy, and I design transparent decision-support systems that keep human judgment in the loop. This continues research from my PhD in Digital Humanities at the University of Bern, after training in computer science and design research, where I found that how AI support is presented changes what people choose.
My research is based at the GUII Lab (UC Santa Cruz), where I am a Visiting Scholar and run controlled experiments and field studies, with food systems as my main testbed and funding from CITRIS and the Banatao Institute. In industry, I build and evaluate production LLM agents at Restaurant Harmony. Before that, I worked on multi-agent LLM workflows at Fixie.ai (now Ultravox) and on responsible-AI interaction patterns in the HAX group at Microsoft Research. I also collaborate with the Digital Humanities Network Switzerland and Switzerland Innovation Park Zurich.
2026
Talk: Reducing food waste with inspectable AI decision support, Stanford Food Summit.
2026
Best Paper Award, 8th International Conference on HCI in Games, HCI International 2026, Montreal.
2026
CITRIS and the Banatao Institute funding to extend the university dining deployment of AI decision support (deployment started 2025).
2025
Swiss translational funding for the initial FlavorLoop deployment.
Upcoming
Mentoring two new undergraduate students on AI and sustainability, Baskin School of Engineering, UC Santa Cruz.
An AI recommender that returns a checkable decision record instead of a free-form answer. In a 12-month university dining deployment, recovery rate rose 14.2 percentage points over a no-system baseline, with staff accepting the decision record in 81% of surplus events.
Read more →
Two studies (N=120, N=80) on what moves trust in AI reasoning. Accountability framing raised trust by 18.2 percentage points, with no significant difference in perceived reasoning quality.
Read more →
LLM agents I built at Restaurant Harmony, evaluated in production with a failure taxonomy and accept, edit and reject logging: 78% acceptance for quantity alerts (312) vs. 41% for timing suggestions on short data windows (187).
Read more →HCI International · Best Paper Award
Foundations of Digital Games
CHI 2026 Extended Abstracts
CHI 2026 Extended Abstracts
I teach AI and Co-design and Ethics and Activism in Tech and Design at UCSC Silicon Valley Extension. Teaching →
Human–AI interaction
AI is entering expert decisions faster than we understand its effect on human skill. I study when people should rely on AI, when they under-rely or over-rely, how the presentation of outputs and reasoning shifts that reliance, and what each pattern costs in accuracy, time, and competence. The goal is systems that leave people more capable over time rather than more dependent. Food systems are the primary testbed; the same methods apply to other high-stakes operational domains.
Pilot-scale controlled studies · Lead researcher
In a field study (N=62) and a vignette study (N=72), direct recommendation lowered accuracy by 18.4 percentage points when the AI was wrong (d = 0.71) while confidence rose 12.3 points. Evaluation support preserved accuracy at higher effort.
Read more →Controlled studies · ICML 2026 AI4GOOD
In two studies (N=120, N=80), accountability framing raised behavioral trust by 18.2 percentage points (95% CI 14.6–21.8), with no significant difference in perceived reasoning quality.
Read more →Field evaluation · Lead researcher and system designer
A constraint-first decision record deployed for 12 months in university dining: recovery rate rose 14.2 percentage points over a no-system baseline, and staff accepted 81% of 1,247 logged decisions.
Read more →Applied evaluation · Restaurant Harmony
A failure taxonomy with accept, edit and reject logging: 78% acceptance for quantity alerts (312) vs. 41% for timing suggestions on short data windows (187).
Read more →These projects apply related ideas of transparent models and appropriate support in education and games. As a Visiting Scholar at the GUII Lab, I contribute to these projects as a collaborator; the named PIs lead them.
Collaboration · PI Casper Harteveld
AI tutoring for goals, feedback and output control, embedded in a gameful platform where students learn methods by running a study.
Read more →
Collaboration · PI Magy Seif El-Nasr
A chat-based agent paired with a process-mined open player model so learners can inspect and compare their strategies.
Read more →
Collaboration · GUII Lab
Discourse as the unit of analysis, triangulated across recordings, coding, survey and observation in serious games.
Read more →
Proof of concept · Cumulus 2022
Meal-photo carbon estimates shown step by step, so uncertain results stay open to correction.
Read more →
Formative evaluation
Phone-camera joint angles within 3.2–7.8° of an inertial sensor, with a small perceived-competence gain.
Read more →
Design work · DHCH
Answers about collections shown only when traceable to the object record. No deployment claimed.
Read more →
Earlier multisensory work
Color–taste mappings in children and adults, and whether music during coffee service changes taste or only the ritual.
Read more →Longitudinal field studies of reliance over months rather than single sessions; extending the decision-record pattern and failure-taxonomy evaluation to additional operational domains; measuring longer-term effects on staff competence and skill retention.
Industry
Alongside my research and teaching at UCSC Silicon Valley Extension and the GUII Lab, I work with companies to build and evaluate AI in production.
2026–present
2023–2025
MA EdTech Solutions, Canada
2022–2023
Fixie.ai (now Ultravox)
2021–2022
EthonAI (AI startup, Zurich) · Project
2021
2017–2020
Toptal
Earlier work on gamification, motivation, and pre-decision support in sustainable eating informed the current focus on appropriate reliance and transparent decision records.
2026 · International Conference on Human-Computer Interaction, pp. 23–42
Talayeh Dehghani, Tobias Hodel, Magy Seif El-Nasr · Best Paper Award
2026 · Design, User Experience, and Usability, HCI International 2026, pp. 36–54
Talayeh Dehghani
2026 · CHI 2026 Extended Abstracts
Talayeh Dehghani
2026 · Proceedings of the 21st International Conference on the Foundations of Digital Games
Talayeh Dehghani, Magy Seif El-Nasr, Tobias Hodel
2026 · CHI 2026 Extended Abstracts
Talayeh Dehghani
2026 · Mensch und Computer 2026
Talayeh Dehghani
Perceived Accountability Framing and Epistemic Trust in AI Reasoning: A Design Finding and Provisional Model
2026 · AI4GOOD Workshop, International Conference on Machine Learning
Talayeh Dehghani
2025 · Joint International Conference on Serious Games, pp. 304–311
Talayeh Dehghani, Tobias Hodel, Magy Seif El-Nasr
2022 · Cumulus Conference: Design for Adaptation, pp. 151
Talayeh Dehghani, Tobias Hodel
2020 · Color Research & Application 45(4), pp. 743–753
M. Fateminia, Talayeh Dehghani, K. M. Azad
Reducing food waste with inspectable AI decision support
2026 · Stanford Food Summit, Stanford, 23 October 2026
Talayeh Dehghani · Talk
Transforming Data Presentation: Leveraging Gamification to Address Communication Failures in Dietary Carbon Footprint Calculators
2025 · Fruitful Failure, Bern, Switzerland
Talayeh Dehghani
2025 · Sustainable Databases in the Humanities 2025, Toronto
Talayeh Dehghani
Practice
I see teaching as a space for inquiry and independent judgment. Students learn to frame problems, question assumptions, test ideas and use evidence to justify design and technical decisions, and to ask not only whether a solution works but how it affects the people who use it.
My teaching draws on my research and industry work at Microsoft Research, Fixie.ai and Restaurant Harmony, and on my background in digital humanities, which brings in how technology shapes culture, values and interpretation, as in the workshop I taught at the DHCH Summer School in Rome. Students prototype, run small studies and work with the same failure patterns, evaluation challenges and human-in-the-loop trade-offs I meet in practice, not hypothetical cases.
UCSC Silicon Valley Extension, US · Course
Human–AI collaboration, participatory design, and technology-mediated creative work. Students design a small human–AI interface change, run a short pilot study, and revise the design from accuracy, confidence and effort measures. Final projects are assessed on research-question clarity, study design and use of evidence.
UCSC Silicon Valley Extension, US · Course
The social and ethical consequences of technology and design, and how designers and engineers can act on them in their own work.
Istituto Svizzero di Roma · DHCH Summer School · Workshop
Values, participation, and technological design in cultural institutions.
University of Bern, Switzerland · Transferable Skills Programme · Workshop
Design thinking and gamification for doctoral research challenges.
Hochschule für Gestaltung und Kunst, Basel · Workshop
Foundations of AI and the role of AI tools in design practice.
University of Science and Culture, Tehran · Curriculum Developer and Lecturer in Design
Designed the curriculum for a user-centered toy design program and taught design methods, prototyping and digital tools to undergraduate and master’s students, with individual feedback on each project.
University of Tehran · Teaching assistant
User-centered and qualitative approaches to the social and cultural dimensions of design.
University of Tehran · Teaching assistant
Seminars on scientific knowledge, technological change, and cultural values.
I mentor undergraduate researchers at UC Santa Cruz through one-on-one Research Independent Study, and will soon mentor two students in the Baskin School of Engineering on AI and sustainability projects linked to surplus-decision support. I have also mentored students at the University of Bern, my PhD institution, and have guided around 15 student projects from first question to finished study.
Curriculum vitae
Human–AI Interaction Researcher & Lecturer, Santa Clara, California
Design and evaluation of AI decision-support and agentic systems, measuring trust calibration, appropriate reliance, and when people follow or override model advice. Controlled experiments and field studies, analyzed with mixed-effects models in Python, R, and SQL, using validated psychometric scales and qualitative interviews. LLM-based, human-in-the-loop workflows and human-centered evaluation metrics, applied in behavior change, digital health, and sustainable food systems.
2026–present
2025–present
Experiments and field studies on human–AI decision support. Trust calibration and appropriate reliance, analyzed with mixed-effects models. Findings translated into LLM-based, human-in-the-loop prototypes. Teaching in AI, co-design, and the social dimensions of computing.
2025–present
Collaborator on human–AI decision-support and reliance studies. I lead the experimental design and analysis for the reliance and FlavorLoop evaluation work; lab PIs lead the broader education and games projects.
2023–2025
2022–2023
2021
2017–2020
UX/UI Designer, Toptal. UX research, prototyping and design systems.
2021–2022
Industry collaboration on evaluating and optimizing industrial AI automation with people in the loop. Mapped end-to-end industrial AI workflows into checkpoints that defined where models act autonomously and where people review, correct, or approve. Collaborated with the product team; the checkpoint model was used to scope the first production deployment of the inspection system.
Ongoing
Cultural heritage, digital collections and AI.
Ongoing
Augmented reality for home rehabilitation.
Swipe sideways to see the full figure.
AI evaluation
Completed controlled studies
People decide how much weight to give an AI’s reasoning. These studies test what moves that weight: the quality of the reasoning itself, or the accountability cues around it.
What moves trust in AI reasoning when the underlying reasoning quality is held constant?
Two between-subjects studies (public policy, N=120; medicine, N=80). Participants allocated behavioral trust (how much of a decision budget they assigned to the AI) after seeing the same reasoning traces under different frames. Only framing varied: named responsible party, institutional liability, and professional review record. Primary outcome analyzed with linear mixed models with random intercepts for participant (repeated trials nested within person). Perceived reasoning quality was measured separately.
Accountability framing raised behavioral trust allocation by 18.2 percentage points (95% CI 14.6–21.8). Perceived reasoning quality did not differ significantly between framing conditions. Reasoning styled to resemble a specific human expert, without accountability framing, earned no more trust than a generic AI baseline.
Lead researcher: study design, data collection, analysis, and writing.
Identifies a calibration risk: accountability cues can increase trust independently of reasoning quality. Any system that surfaces its reasoning should treat framing as a design variable that can decouple trust from accuracy.
AI4GOOD Workshop, ICML 2026.
Swipe sideways to see the full figure.
AI evaluation · Proof-of-concept with validation
The system estimates dietary carbon from a meal photograph. Dominant colors are mapped to ingredient categories; two short questions resolve residual ambiguity. The estimate remains inspectable so uncertain results can be corrected rather than accepted as a black-box score.
Can dietary carbon be estimated at the moment of the meal in a way that reduces logging burden while remaining correctable when the system is uncertain?
Fixed color-to-category mapping plus two clarifying questions (dietary pattern; presence of high-impact ingredients). Selective automation: visual recognition reduces manual entry; low-confidence cases stay open to user correction. The output is a transparent estimate, not a single opaque score.
Compared against lifecycle assessment values on a labeled meal set. Feasibility check with untrained participants reading categories without instruction.
Demonstrates an interpretable estimation pattern (visible mapping + residual questions + correctable output) for an AI system whose errors users can see and fix. This early evaluation of inspectable AI outputs informed later work on appropriate reliance and decision records.
Cumulus Conference: Design for Adaptation, 2022.
Collaboration · GUII Lab
Discourse is treated as the unit of analysis. Recordings, transcription, coding, survey, and behavioral observation are triangulated to trace how player-to-player talk inside a serious game becomes shared understanding and collective action on applied problems such as wildfire preparedness.
How does player-to-player communication in a serious game become shared understanding, and when does the game itself encourage that learning?
A method for tracking how players co-learn, and for seeing how the game encourages that learning.
PI: Prof. Magy Seif El-Nasr, with Dr. Mario Escarce, Dr. Emily Chen, Shivam Shulka, and Darian Lee. Collaborator: Talayeh Dehghani, Visiting Scholar, GUII Lab, UC Santa Cruz.
AI evaluation · Industry collaboration
Design mapping of appropriate autonomy
End-to-end mapping of an industrial AI inspection workflow. Every point where automated outputs met human decisions was identified, from data intake through model output to final accept or reject.
Where should the AI inspection system act on its own, and where should a person check its output before work continues?
Process mapping of the full workflow. Each hand-off between automated output and human review was marked. Checkpoints were defined for inspect, correct, or approve actions.
Annotated workflow identifying autonomy boundaries and required review points. No user evaluation was conducted; this was a design and process-mapping collaboration. The checkpoint model was later used to scope the first production deployment of the inspection system.
Industry collaboration (forward-deployed style) with EthonAI (Zurich), 2021–2022.
Explicit mapping of where automation ends and human review begins in a multi-stage industrial AI inspection workflow.
AI evaluation
Completed controlled studies (pilot-scale)
A direct recommendation can make a choice easier and, over repeated use, reduce the person’s ability to make it. These studies compare AI that decides with AI that keeps the person’s evaluation in the loop.
When should a system recommend, and when should it ask the person to evaluate?
Three conditions on the same food decision: (1) direct recommendation, (2) evaluation support that surfaces criteria without selecting, and (3) forced reflection after choice. Between-subjects field study (N=62; ~20–21 per condition) and randomized vignette study (N=72). Outcomes included decision accuracy (against a gold-standard ranking), confidence (0–100 scale), self-reported effort, motivation, authenticity (single-item “this choice felt like mine,” 1–7), and dropout. Labeled pilot because of the modest per-condition N; a post-hoc sensitivity analysis indicated 80% power to detect medium-to-large effects (d ≥ 0.65) on the primary accuracy contrast.
Direct recommendation increased reliance and reduced effort but lowered accuracy when the AI was wrong (mean accuracy drop 18.4 percentage points, 95% CI 11.2–25.6, d = 0.71). Confidence still rose (+12.3 points on the 0–100 scale, 95% CI 6.8–17.8). Required justification after choice produced higher dropout in the vignette study (5 of 30 in the forced-reflection condition vs. 0 of 42 across the other two conditions; Fisher’s exact p = .010). Evaluation-support conditions preserved accuracy at the cost of higher effort. Authenticity scores were highest in the forced-reflection condition and lowest under direct recommendation.
Lead researcher: study design, data collection, analysis, and writing.
Isolates a concrete trade-off: systems that decide for the user can raise short-term reliance and confidence while eroding accuracy under error. Suggests evaluation-support designs as a lower-risk default when the cost of error is high.
Manuscripts in preparation for HCII 2027.
AI evaluation
System design and completed field evaluation
Hard operational constraints (equipment, time, allergen rules, kitchen capacity) eliminate infeasible actions. Remaining candidates are scored on operational feasibility, flavor compatibility, nutritional quality, and environmental impact. Scoring weights are set by the kitchen lead at the start of each service period and can be adjusted; default weights prioritize operational feasibility and allergen safety. The output is a decision record: recommended channel or explicit no-action, reasons, and eliminated options.
Does a transparent, constraint-first decision record improve surplus recovery and decision quality relative to existing practice, without reducing diner acceptance?
Constraint-first filtering followed by multi-criteria scoring. Decision record always includes recommended action or explicit no-action, rationale, and eliminated options. 12-month deployment in UC Santa Cruz Dining (started 2025, before the 2026 CITRIS award). Recovery rate was logged against a matched no-system baseline period. CITRIS funding supports continued data collection and analysis. Blinded diner taste tests comparing dishes produced from system-supported versus standard surplus decisions.
Recovery rate was higher in the system condition than in the no-system baseline (mean difference +14.2 percentage points, 95% CI 9.1–19.3). Kitchen staff accepted the decision record in 81% of 1,247 logged surplus events; the most common edit was a change in portion size rather than a rejection of the recommended channel. In blinded taste tests (N=186 diners), overall liking did not differ significantly between system-supported and standard dishes; the secondary “would choose again” measure favored the system condition. Staff reported lower time spent on ad-hoc spreadsheet reconciliation. No increase in allergen or operational incidents was observed.
Lead researcher and system designer: decision-record design, evaluation protocol, data analysis, and deployment coordination.
Shows that a constraint-first, inspectable decision record can raise measured recovery relative to baseline practice while remaining acceptable to both kitchen staff and diners. The explicit no-action option and eliminated-options log made over-reliance visible and correctable.
Deployment results: manuscript in preparation. Earlier formative work on the same system: FDG 2026, HCII 2026, CHI EA 2026.
Senses and perception
Taste is not only on the tongue. These studies ask which color–taste links are stable enough to guide packaging, and whether music during coffee service changes taste or changes the ritual around it.
Color, sound, and presentation shift how sweet, sour, or bitter food seems. Which of those shifts are reliable, and which are framing rather than flavor?
With Mehran Fateminia and K. M. Azad (color studies)
AI evaluation · Industry
Applied evaluation of production agents
Operational kitchen data (meals served, vegetable weight by category, plate-waste trends) are converted into recommendations, alerts, and reports that remain subject to human review before action.
How can operational data become a recommendation a kitchen can check and act on, rather than a report that is ignored?
AI agents and automated workflows generate recommendations and alerts. Evaluation used a failure taxonomy (incorrect quantity, wrong timing, missing context, over-confident phrasing) and logged accept, edit, and reject rates.
Acceptance was high for quantity alerts (78% of 312 alerts) and lower for timing suggestions on short data windows (41% of 187 alerts). Edit rate was higher when confidence phrasing was over-confident. Human review remained in the loop for all actionable outputs.
AI Lab Automation Scientist, Restaurant Harmony (2026–present): evaluation design, failure taxonomy, logging instrumentation, and analysis.
Shows how a lightweight failure taxonomy and accept/edit/reject logging can surface differences in agent reliability before full automation.
Swipe sideways to see the full figure.
Collaboration · DHCH
Design constraint and generative boundary work
Generative overlays and collection queries are constrained to evidence in the object record and its documentation. An answer that cannot be traced to the record is not shown. Multiple accounts are presented with sources.
What may an augmented or generative system say about a physical object or tradition if every claim must remain checkable against the record?
Claims limited to object record and associated documentation. Overlay or interface surfaces the source of each relation. Collection experiments use machine learning to surface links of color, material, technique, and context; untraceable generative answers are withheld. Multiple accounts shown with sources.
No museum deployment is claimed. Design and constraint work with Digital Humanities Network Switzerland (DHCH).
Defines a traceability requirement for generative answers in heritage contexts: fluency is not treated as evidence.
Digital health
Formative technical and user evaluation
Phone-camera pose estimation with skeletal overlay and joint-angle feedback for home rehabilitation exercises.
Can phone-based pose estimation provide usable joint-angle feedback without clinical equipment in the room?
Four Embodied Corrective Feedback principles derived from sessions with six patients and four physiotherapists. Prototype estimates pose, aligns skeletal overlay, and reports joint angles during movement.
Joint angles matched an inertial sensor within 3.2–7.8 degrees (r = 0.91–0.97). In eight participants, perceived competence was higher with RehabAR than with a paper sheet (IMI 5.4 vs 3.9). Body-reference and immediacy predicted the gain; immersion did not.
Formative evaluation only; not a clinical trial. Developed from earlier patent work on an autonomous laser treatment device. Switzerland Innovation Park Zurich.
Shows technical feasibility and a small user-perceived competence gain for phone-based corrective feedback.
Collaboration · GUII Lab
A chat-based large-language-model agent is coupled with an open player model constructed by process mining of interaction data. Learners inspect their own problem-solving strategies and compare them with alternative solution paths while playing a game that carries the STEM concept.
Can an open player model presented inside the game support evaluation of a learner’s strategy relative to alternative solution paths?
An open learner model that externalizes the problem-solving process and increases opportunities for reflection in intelligent tutoring and game-based learning.
PI: Prof. Magy Seif El-Nasr. Co-PIs: Prof. Roger Azevedo (UCF), Prof. Tyler Sorensen (UCSC), Prof. Brian Smith (BC), Prof. Jichen Zhu (ITU). Students: Sai Siddartha Maram, Devon McKee, Cameron Marano, and Jiahong Li. Postdoc: Zhiyu Lin. Collaborator: Talayeh Dehghani, Visiting Scholar, GUII Lab, UC Santa Cruz.
Collaboration · GUII Lab
AI tutoring for goals, feedback, and output control is embedded inside a gameful constructionist platform. Students learn research methods by designing, conducting, and analyzing a study rather than by reading about method.
Can game-based construction and AI support help students learn research methods by designing and running a study, not only by reading about one?
The research contributes to education by combining game-based constructionist methods with AI so students can learn research methods by using them.
PI: Prof. Casper Harteveld (Northeastern University). Co-PIs: Prof. Magy Seif El-Nasr, Prof. Camillia Matuk (UCSC), Prof. Steven Sutherland (University of Houston–Clear Lake), Prof. Eddie Melcer (UCSC), and Prof. Elin Carstensdottir (UCSC). Student: Sai Siddartha Maram. Postdoc: Zhiyu Lin. Collaborator: Talayeh Dehghani, Visiting Scholar, GUII Lab, UC Santa Cruz.