Pedro Martins de Bastos

Ph.D. Student · Language and Learning Analytics Lab · UC Irvine

Pedro Martins de Bastos

I study how AI teammates change the way people talk, think, and connect in small groups.

I want AI teammates to support people’s ideas and their relationships with one another.

Explore my researchGet in touch

About

I’m a Ph.D. student in Education at the University of California, Irvine, specializing in Teaching, Learning, and Educational Improvement. I work with Nia Nixon in the Language and Learning Analytics Lab.

My research brings together learning analytics, language analysis, and computer-supported collaborative learning. I study what happens to the people in a group when an AI joins the conversation. I’m also one of two principal developers of TRAIL, a platform for studying human–AI teams.

Research & featured work

What changes when an AI becomes part of a team? My work examines the conversation, the experience of belonging, and the tools we use to study both.

HICSS-60 · Accepted · 2027

The Social Cost of an AI Teammate: How an Artificial Teammate Reshapes Human-Human Communication in Small-Team Decision-Making

Nixon, N., Choi, J., Martins de Bastos, P., Samadi, M. A., Park, S., & JaQuay, S.

An AI can take part in a conversation while the people around it feel less connected. In a text-chat rescue task, students on AI-inclusive teams participated less and reported lower belonging, status, and connection than students on all-human teams.

Read the paper on arXiv

Preprint (arXiv)

TRAIL: A Platform for Configurable Human–AI Teaming Experiments

Samadi, M. A., Martins de Bastos, P., Choi, J., JaQuay, S., Park, S., & Nixon, N.

TRAIL lets researchers configure an AI teammate, run group conversations, and connect the discussion with survey responses. Its classroom demonstration showed different patterns for a persona that challenged ideas and one that encouraged inclusion.

Read the paper on arXiv

Two AI roles, different patterns

Descriptive summary of reported findings

Explore the experiment flow
  1. Configure persona: Researchers store a graded Big Five profile, backstory, communication style, expertise, education, model, and triggers.
  2. Selective participation: The message pipeline batches messages, integrates memory, decides whether to speak, and generates a persona-aligned reply.
  3. Dual memory: Recent conversation context is paired with periodic reflective summaries of the team’s progress and dynamics.
  4. Chained sessions: Researchers link sessions into ordered sequences for studying teams over time.
  5. Export & analytics: Condition-labeled exports connect chat logs, persona specifications, decision traces, and survey responses for analysis.

Platform components, not a timeline of measured outcomes. Dual memory is integrated within the message pipeline.

Observed persona comparisons; no significance tests were reported.

Directions compare these two configured personas, not all possible AI teammates.

Observed persona comparisons; no significance tests were reported.

All reported comparisons
MeasureCognitive-scaffolding agentSocially-supportive agent
AI message lengthLongerShorter
Topic similarity to studentsHigherLower
Language-style matchingHigherLower
Rated contribution to thinkingHigherLower
Building on the AI’s ideasHigherLower
Feeling includedLowerHigher
Reported over-relianceHigherLower
AI share of messagesSimilar participation ratesSimilar participation rates
TRAIL lets researchers configure an AI teammate and study the conversation around it. In this classroom demonstration, the cognitive-scaffolding persona received higher contribution ratings, while the socially-supportive persona coincided with greater inclusion and lower reported over-reliance. Read the paper on arXiv

Preprint (arXiv)

Out of sync: Accommodation, evaluation, and emergence in five weeks with an agentic AI teammate

Nixon, N., Choi, J., Martins de Bastos, P., & Samadi, M. A.

A teammate can be rated less helpful while the team’s view of its collaboration improves. Across repeated sessions, language, helpfulness ratings, and perceptions of collaboration followed different patterns.

Read the paper on arXiv

Helpfulness and collaboration can diverge

Descriptive summary of reported findings

Relative ordering only — positions do not show scores or effect sizes.

Primary analyses used sessions 1, 2, 4, and 5.

Session 3 was excluded after a technical issue; session 6 used a substantially longer task prompt.

Both agent configurations changed after session 1.

Positive-affect words describe reflections on the collaboration; they do not measure liking of the AI.

This small exploratory study leaves the reason for the crossover unresolved.

Session 1: Cognitive-Agent teams were higher on perceived collaboration process and positive-affect words in reflections. The Cognitive Agent was also rated more helpful.

All reported comparisons
MeasureSession 1Session 5
Perceived collaboration processCognitive Agent: Higher; Supportive Agent: LowerSupportive Agent: Higher; Cognitive Agent: Lower
Positive-affect words in spoken reflectionsCognitive Agent: Higher; Supportive Agent: LowerSupportive Agent: Higher; Cognitive Agent: Lower
Perceived AI helpfulnessCognitive Agent: Higher; Supportive Agent: LowerCognitive Agent: Higher; Supportive Agent: Lower
Supportive-Agent teams began below Cognitive-Agent teams and ended above them on perceived collaboration process and positive-affect words in reflections. Students still rated the Supportive Agent less helpful throughout; the study does not establish why team perceptions changed. Read the paper on arXiv

Service · AIED 2026

Human-AI Teaming: From Tools to Teammates II

Designing AI That Belongs

Co-organizer · June 28, 2026 · Seoul

I co-organized the second annual workshop on human–AI teaming, a half-day workshop at AIED 2026.

Read the workshop summary

Experience & education

  1. Ph.D. in Education · UC Irvine

    Research in the Language and Learning Analytics Lab, with a specialization in Teaching, Learning, and Educational Improvement.

  2. Data Engineer · Atypical AI

    Developed systems that retrieved relevant material for AI tutoring and personalized learning support.

  3. Data Science Associate · Adtalem Global Education

    Built models to identify students who might need academic support, along with data pipelines and dashboards for academic teams.

  4. B.S. in Data Science and Statistics · Minerva University

    Minor in Brain and Behavior.