Staffing an AI Project and Handling Personal Identifiers in LLM Training Data

This question has two parts. First, describe the roles on your AI project team and how they work together. Second, explain how you would approach training a large language model when the training data contains usernames and email addresses. ### Clarifying Questions - If I have not been on an AI project team, may I describe the team I would assemble for one? - Where does the data come from: public web text, logs from a consumer product, or content that business customers store in Microsoft services, and what did users agree to when it was collected? - Is the plan to train a model from scratch or to fine-tune an existing one for a single feature, and does that feature ever need to output real email addresses? - Which jurisdictions do the affected users live in, since that changes consent and deletion obligations? ### Part 1 — Who does what on the team Name each role on the team, what it owns, and where you as the product manager fit in. ```hint Follow the model's lifecycle Trace the work from data through training, evaluation, launch and monitoring, and ask who owns each step. Some of the owners sit outside engineering. ``` #### What This Part Should Cover - The core build roles and what each owns across data, modeling, evaluation and serving. - The partners outside engineering, such as privacy, legal, security and responsible AI review, and when they get involved. - How decisions are made, and where the PM adds value versus deferring to specialists. ### Part 2 — Usernames and emails in the training data The dataset you plan to train on includes usernames and email addresses. How would you handle it? ```hint Question the need first Before picking a technique, ask whether the model needs these identifiers at all and whether you are allowed to use them. The answers decide between removing, replacing or protecting them. ``` #### What This Part Should Cover - Whether the data may be used at all: consent, purpose and data minimization. - How identifiers are detected and removed or replaced, and what each option costs in model quality. - How to verify that the trained model does not leak identifiers, and how to handle later deletion requests. ### What a Strong Answer Covers - Clear decision rights: who decides on data use versus model quality, and how disagreements between them are escalated. - Privacy controls across the whole pipeline, from collection through training to model outputs, rather than a single "anonymize it" step. - An explicit trade-off between privacy protection and model usefulness, backed by measurement. - Checks that must pass before launch and monitoring that continues after it. ### Follow-up Questions - Your scientists say removing usernames hurts the model because usernames carry who-replied-to-whom context in discussion threads. How do you resolve it? - After launch, a user asks for their data to be deleted, and their email address was in the training set. What do you do? - A red-team test gets the model to output a real email address from the training data. Do you delay the launch, and which fixes do you consider?