Skip to content
Maren Lian
Research

I study how people understand, trust, and act through complex systems.

My work spans human-AI interaction, behavioral measurement, and research for digital products and services. Across those settings, I make hard-to-see behavior observable and turn evidence into decisions people can act on.

Current research

COoKIE Lab, University of Toronto

Principal Investigator: Dr. Anastasia Kuzminykh

At COoKIE Lab, I investigate conversational dynamics in human-AI interaction: how a system's language changes under user pressure, how those changes shape trust, and how interaction design can make model behavior easier to assess.

That work is one part of a broader research practice that also includes bilingual interviews, usability evaluation, behavioral analysis, and translating evidence into product and service decisions.

Research areas

A broad practice connected by one question: what evidence helps people make better decisions?

  • 01

    Human-AI interaction

    How conversational systems shape trust, judgment, and the way people interpret information over time.

  • 02

    Behavioral measurement

    Turning difficult concepts such as alignment, trust, and conversational drift into observable, testable signals.

  • 03

    Research into decisions

    Using qualitative and quantitative evidence to guide product direction, service design, and responsible interaction patterns.

Featured paperAccepted · HAI 2026

The Sycophantic Slide: From Static Detection to Dynamic Trajectories in Multi-Turn LLM Conversations

The 14th International Conference on Human-Agent Interaction · HAI '26 · Osaka, Japan · November 16–19, 2026

The paper reframes sycophancy as a trajectory, not a switch: linguistic agreement can accumulate across turns before a model visibly changes its stance.

Corpus
~2,500 five-turn conversations per primary model
Primary models
Claude 3.7 Sonnet · GPT-4o · DeepSeek-R1
Analysis
LIWC-22 · random forest · logistic regression · longitudinal GLM
Validation
Out-of-sample testing with Llama 3.3 and Qwen 2.5
What the study found
  1. 01

    The drift starts before the flip

    Conversations ending in a stance change showed compounding assent across turns; stable conversations stayed comparatively flat.

  2. 02

    A small signal set travels across models

    Eight shared lexical markers, plus the turn-to-turn change in assent, captured a cross-model pattern of compliance and resistance.

  3. 03

    Detection can become proactive

    Tracking the trajectory creates an opportunity to surface conversational drift before the model explicitly abandons its position.

The public paper link will be added when the camera-ready version is available.

Methods I work in

  • Mixed-methods research design
  • Human-centered AI evaluation
  • Computational linguistic analysis
  • Longitudinal conversation analysis
  • Semi-structured interviews
  • Moderated usability evaluation
  • Thematic analysis
  • LIWC-22
  • Random forest feature selection
  • Logistic regression
  • Survey design
  • SQL and Python for behavioral analysis
Contact

I'm interested in research where human behavior, responsible technology, and product decisions meet.

Ask about the research →