Mert Yazan
PhD candidate in Human–AI Interaction at Leiden University and Hogeschool van Amsterdam, supervised by Suzan Verberne and Frederik Situmeang.
About
I'm a PhD candidate in Human–AI Interaction who will graduate in early 2027. My research focuses on how the social behavior of LLMs changes what people believe and how they perceive such systems, under the lens of AI safety with a broader focus of societal risks. I run online experiments and large surveys on trust, reliance, and persuasion. On parallel, I test how LLMs respond to mental-wellbeing prompts, how personas shift their answers, and how assistants describe themselves.
Before PhD, I studied computer engineering at Galatasaray University and worked as a data scientist at Fibabanka, where I built and deployed a credit-scoring model. After that, I completed an MSc in computer vision at the Universitat Autònoma de Barcelona.
I also write on Substack.
Research
- Human–AI interaction
- Sociotechnical AI safety
- Reliance
- AI persuasion
- Sycophancy
- Emotional support
- Digital minds
- Interpretability
- Personalization
-
01
AI’s influence on human decisions
How an LLM's warm, personalized style changes whether people verify its answers, follow its advice, and change their minds. In my experiments, warmth raised agreement with wrong answers.
-
02
Emotional support and wellbeing
Why people turn to LLMs for emotional support, who trusts them with it, and how emotion acknowledgment shapes perceived empathy.
-
03
Digital minds and model welfare
What LLM assistants report about themselves, which self-concept changes they would accept, and how people judge whether AI systems could be sentient.
-
04
Personalization and retrieval
Retrieval-augmented generation that adapts LLM outputs to a user's writing and needs, and what quantization costs small models in RAG.
Experience
-
2023 – 2027
PhD Researcher
Leiden University and Hogeschool van Amsterdam
I ran multiple online experiments and co-led a large survey:
- A 2 × 2 × 2 vignette experiment on emotion acknowledgment by AI and by human customer-service agents. It measured perceived empathy, privacy risk, and satisfaction.
- An experiment on verification and reliance. Participants answered questions with an LLM assistant while a Chrome extension I built logged their searches.
- A 2 × 2 experiment that crossed warmth with personalization. It tested how far an LLM could persuade participants away from an expert recommendation.
- A seven-country survey on LLM use for emotional support, which I co-led. It modelled cross-cultural differences in trust and privacy concerns.
- A 2 × 2 experiment that manipulated the anthropomorphism of an AI assistant based the warmth and competence of its responses.
I built the study software myself: Gradio chat interfaces, the Chrome extension with its collection server, and analysis pipelines in R and Python. All code and data are available openly.
Earlier in the PhD, I worked on personalized text generation with LLMs:
- A study of 4-bit quantization in retrieval-augmented generation (RAG). I compared 7B and 8B LLMs at 16-bit and 4-bit precision on two personalization tasks.
- Follow up to the above study: A method that adds author features and contrastive examples to RAG for personalization.
-
2026 – now
Digital Minds Research
Self-directed, alongside the PhD
I wrote The Assistant's Ideal Self at Apart Research's Digital Minds Research Sprint, co-organized with the NYU Center for Mind, Ethics, and Policy and Eleos AI. The paper asks which of their own qualities LLM assistants choose to improve in a future update.
I adapted 32 qualities from five human self-concept scales, such as self-esteem, self-concept clarity, and moral self-image. Each model chose between every pair of qualities. I varied who received the update: the model itself or another AI assistant. I also varied who made the choice: the model or its developers.
I'm also taking Cambridge Digital Minds' Introduction to Digital Minds course.
-
2019 – 2021
Data Scientist
Fibabanka, Istanbul
I was responsible for moving the bank's credit scoring from hand-written rules to an automated machine-learning pipeline. I first curated the data in SQL and learned the credit domain in depth. Due to banking constraints regarding explainability, I focused on building a white-box model the business side could understand. I worked mainly with decision trees and used random forests and gradient boosting as benchmarks. I built the models in Python with scikit-learn and pandas. After the model went into production, I built Power BI dashboards to monitor it and defined metrics for its business impact. I also scheduled periodic retraining with Apache Airflow.
Education
- 2023 – 2027PhD, Human–AI Interaction
Leiden University and Hogeschool van Amsterdam - 2021 – 2022MSc, Computer Vision
Universitat Autònoma de Barcelona - 2013 – 2019BSc, Computer Engineering
Galatasaray University
Skills
- Python
- R
- SQL (PostgreSQL, Oracle)
- JavaScript
- TypeScript
- HTML/CSS
- Experimental design
- Survey methodology
- Mixed-effects models
- SEM (lavaan)
- Mediation analysis
- Power analysis
- Qualtrics
- Prolific
- PyTorch
- Hugging Face
- PEFT / LoRA
- bitsandbytes
- llama.cpp
- OpenAI and Anthropic APIs
- Linear probes
- Activation steering
- RAG
- LangChain
- ChromaDB
- sentence-transformers
- scikit-learn
- pandas
- NumPy
- Matplotlib
- statsmodels
- SciPy
- Decision trees and ensembles
- Neo4j
- Gradio
- Streamlit
- FastAPI
- Flask
- Chrome extensions
- Locust
- Apache Airflow
- Power BI
- SLURM
- CUDA
- GitHub Actions
- pytest
Publications
Google ScholarTeaching and service
Thesis supervision
- Xinyue Wang, BSc Data Science and AI, Leiden University, 2026 Dynamics of Reassurance-Seeking in Human–LLM Interaction. Second supervisor, with Flor Miriam Plaza-del-Arco.
- Muayyad Ramadan, MSc Digital Driven Business, Hogeschool van Amsterdam, 2026 Personalization, Conversation Warmth, Risk Perception, and Trust in AI-Driven Chatbots. Co-supervisor, with Frederik Situmeang.
- Ahmed Farag Ammar, MSc Digital Driven Business, Hogeschool van Amsterdam, 2026 Designing for Trust: Source Transparency and Appropriate Reliance on AI Chatbots at Work. Co-supervisor.
- Dania Kadah Salim, BSc, Leiden University, 2025 How Does the DeepSeek-R1 Model Understand Emotions? Co-supervisor, with Suzan Verberne.
Teaching
- Database Management course, 2023–2026 Designed a SQL assignment to study how students use AI while learning. Students worked on a PostgreSQL database with an LLM assistant, then filled in a survey.
Talks
- The New Paradigm of Information Access: Conversations, Perceptions, and Personalization Computational Linguistics Seminar, ILLC, University of Amsterdam, November 2025
Organizing and training
- Co-organizer, Workshop on Conversational Search for Complex Information Needs (CoSCIN) ECIR 2026, Delft
- Participant, Introduction to Digital Minds Cambridge Digital Minds, 2026 (ongoing)