CHAPTERS
- 0:00 – 0:30
Why it’s worth asking what AI “knows” about you
Zoe sets the context: people share sensitive, personal, and workplace information with AI that they wouldn’t normally broadcast. She frames the core questions—does the AI remember you, is anyone reading chats, and could your inputs affect ads or other outcomes.
- •Common sensitive prompts include health, finances, and anxious email drafts
- •Key user concerns: memory across time, human access to chats, and advertising use
- •Goal: explain what happens to your info when you share it with AI
- 0:30 – 1:01
A practical framework: where your data goes, why, and for how long
She introduces the idea that user control should be tied to understanding data destinations and retention. Information can be transient (within a chat) or persist to improve your experience, and sometimes may benefit broader systems.
- •Policies vary by provider—always verify specifics
- •Focus on three questions: where data goes, why it’s used, and retention duration
- •Some data disappears after the chat; some persists for product experience; some may help others
- 1:01 – 1:31
Concrete example: a “birthday gift” prompt and what happens after you close the chat
Using a gift-shopping scenario (hiking-loving partner near Denver), Zoe illustrates how the model uses details during the conversation and tees up the four main “places” those details might go afterward.
- •Example details: partner likes hiking; you live near Denver
- •Within a chat, the AI uses whatever you provide as active context
- •Sets up the four distinct data uses/destinations
- 1:31 – 2:01
Use #1 — The conversation context (session-only understanding)
Zoe clarifies that the model’s baseline behavior is not long-term memory: it holds context only within the ongoing conversation. Starting a new chat is essentially a blank slate unless additional product features store information.
- •Model context is limited to the current conversation
- •A new chat typically starts with no knowledge of prior details
- •User control is primarily what you choose to type into the chat
- 2:01 – 2:32
Use #2 — Product memory saved to your account (cross-chat personalization)
She distinguishes “memory” features from the model itself: details may be stored in your account and made available in future chats. Users often can edit, clear, or disable memory depending on product settings.
- •Memory is an account-level feature, not the model inherently “remembering”
- •May be enabled by default or opt-in depending on the tool
- •Future chats can reuse saved details (e.g., hiking interest)
- •Typically offers controls: edit, delete, or turn off memory
- 2:32 – 3:02
Use #3 — The provider’s internal systems (operations, safety, research, ads, improvement)
Zoe explains that providers may retain information within their own systems for service operation and governance functions. Exact access and retention vary, so the dependable step is checking privacy policies and retention settings.
- •Possible uses: service operations, safety/abuse review, bug fixing, research
- •May include commercial purposes like serving ads (provider-dependent)
- •May also include model improvement where applicable
- •Best practice: read privacy policy and review retention/settings
- 3:02 – 3:32
Use #4 — Training future models (patterns, opt-outs, and privacy handling)
She describes how some providers use conversations to train future model versions. While personal data is generally removed and chats aren’t stored to be replayed to other users, the key idea is that content may contribute as training patterns, with opt-out options in many products.
- •Some providers train future models on conversation data
- •Personal data is generally removed before training use
- •Outputs aren’t meant to reveal your chat to other users
- •Your words become training “patterns,” not retrievable chat logs
- •Many providers provide an opt-out for training
- 3:32 – 3:33
What control looks like in practice (policies, settings, and Anthropic defaults)
Zoe reinforces that user control comes from understanding the provider’s policy and the product’s settings. She notes Anthropic’s approach: making choices understandable, and in organizational Claude deployments, training is off by default.
- •Privacy policy explains categories of data use
- •Product settings provide practical toggles and controls
- •Anthropic aims for clear, understandable choices
- •For org/enterprise Claude use, training is off by default
- 3:33 – 4:02
Habit 1–2: Review settings, then share based on comfort and use-cases
She recommends spending a few minutes auditing key settings and then making sharing decisions based on sensitivity and where data might flow. The emphasis is on intentionality rather than blanket rules.
- •Check: memory, chat history, connected apps, and training toggles
- •Locate these in account/privacy settings
- •Decide what to share based on comfort and intended uses
- •Think across: in-chat use, saved memory, and potential training
- 4:02 – 4:33
Habit 3–4: Minimize sensitive details and choose the right plan for the data
Zoe advises reducing exposure by using placeholders and limiting identifying details when they’re unnecessary. For confidential or regulated work, she recommends business/enterprise plans with different data-handling terms, verified in documentation.
- •Leave out what you don’t need to share to get value
- •Use placeholders (e.g., omit real names in email rewrites)
- •Match tool/plan to sensitivity: consumer vs business/enterprise
- •Confirm data-handling terms in product documentation
- 4:33 – 4:50
Bottom line: what AI knows depends on your inputs and your settings
She concludes that, in the moment, AI “knows” what you provide in the conversation, and longer-term durability depends on defaults and settings like memory and training. A small upfront review plus thoughtful usage helps you stay in control.
- •Baseline: AI knows what you put into the current chat
- •Persistence depends on product defaults and user settings
- •Upfront settings review improves control and clarity
- •Thoughtful engagement reduces unnecessary data exposure
