OshenDavidson
Why AI Safety Matters to Me
As AI capabilities increase and systems operate with more autonomy, our efforts to understand model behavior need to scale alongside them. This means expanding interpretability research, examining capabilities and limitations, and identifying vulnerabilities in existing defenses. This also means strengthening our ability to anticipate emerging risks and develop safeguards to address them.
That understanding also needs to account for AI's influence on people and society. How a model presents information, expresses uncertainty, and responds to us can affect belief formation, trust, and human autonomy. Building effective safeguards depends on both understanding the depth of these systems' capabilities, as well as how these effects can impact individual behavior.
Through empirical studies and behavioral evaluations, my work is one contribution to a shared effort to better understand AI. That effort grows stronger as people share knowledge, build on one another's work, and bring different perspectives to the same challenges. If you're thinking about these questions too, I'd love to connect, compare notes, or collaborate on a study.
Studies on how language models find information, use sources, and generate answers.
Citation Selection in Commercial AI Search: A Stage-Based Analysis of Retrieved Pages
What separates pages that ChatGPT cites from pages it retrieves and discards, conditioned on commercial intent. 217,508 retrieved pages across 7,500 commercial prompts, segmented by buying-journey stage and analyzed under four cross-checked analytical passes.
From Search to Citation: Measuring How ChatGPT Retrieves and Selects Sources
How ChatGPT moves from search to citation. 548,534 pages retrieved across 15,000 prompts analyzed to separate the retrieval decision from the citation decision, with rank-controlled six-model decomposition and DA × rank tier cross-stratification.
Community Sources in AI Search: A Cross-Model Analysis of 5.5 Million Answers
How user-generated content and community platforms shape brand visibility across ChatGPT, Perplexity, Gemini, and Google AI Mode. 5.5M answers and 61K queries analyzed across 821,000 cited domains.
Selected Work

Controlled Baseline for Agent Evaluations
A control-arm execution platform for comparing agent workflows with a conversational baseline. Preserves staged instructions and resolved inputs, records execution details, and prepares paired outputs for human evaluation.

Human Evaluation Tool
A human evaluation app for blinded comparisons of generated outputs. Collects rubric-based judgments alongside automated scores to examine where human and model assessments agree or differ.

Content Quality Data Pipeline
15,000-query empirical research pipeline measuring which content quality signals predict ChatGPT citation. Six idempotent pipeline steps collect ChatGPT responses through the real consumer interface, join against Google SERP data for original and fan-out queries, extract 24 content quality signals per page, and run multi-method statistical analysis.
Research Coverage & Citations
- Zyppy
- Search Engine Land
- MSN
- eMarketer
- MSN Deutschland
- Forrester





