← Openai Interview Insights

Openai·Software Engineer·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Interviewed for a data science role at OpenAI. One question that stuck with me was about working with messy or historical data when you can't just run a clean experiment.

Questions Asked (1)

Q1

When you can't run a controlled experiment, how do you make decisions using imperfect or historical data?

A/B Testing & ExperimentationAdaptability & AmbiguityProduct Analytics & Metrics
Author's notes

I fumbled the opener a bit, started listing techniques before actually grounding it in a real scenario.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that controlled experiments are ideal but not always feasible, then describe a structured approach to decision-making with imperfect data. Emphasize triangulation, causal inference techniques, and iterative learning to reduce uncertainty.

Pro tip: Highlight that you quantify uncertainty and set decision thresholds upfront, and that you treat decisions as reversible experiments when possible. This shows you balance rigor with pragmatism.

1. Clarify the decision and constraints

Define the decision, its impact, and why a controlled experiment isn't possible. Identify what data is available and what assumptions are being made.

2. Triangulate multiple data sources

Combine historical data, observational studies, qualitative feedback, and domain knowledge to build a more robust picture. Look for converging evidence.

3. Apply causal inference methods

Use techniques like propensity score matching, difference-in-differences, instrumental variables, or regression discontinuity to approximate causal effects from observational data.

4. Quantify uncertainty and simulate

Model uncertainty with confidence intervals, sensitivity analyses, or Monte Carlo simulations. Consider best-case, worst-case, and most likely scenarios.

5. Decide, monitor, and iterate

Make a decision based on the evidence, but set up monitoring and guardrail metrics to detect issues. Treat it as a reversible experiment if possible, and update as new data arrives.

Key Points to Mention

  • Causal inference techniques (e.g., difference-in-differences, propensity score matching)
  • Triangulation of multiple data sources (quantitative and qualitative)
  • Quantifying uncertainty (confidence intervals, sensitivity analysis)
  • Decision reversibility and setting guardrail metrics
  • Iterative learning and updating priors with new data
  • Domain knowledge and expert judgment to validate assumptions

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.