Pi's privacy policy sets opt-outs, not a no-training default

The design

Pi is a conversational assistant made by Inflection AI, Inc. The company's own privacy policy, as retrieved on 16 September 2026, states it was last updated 11 August 2026 and governs Pi along with Inflection's other services. It covers 'inputs' of text, voice or audio a person sends Pi, plus device, usage and cookie data collected automatically. The policy warns that personal information volunteered in an input 'may be reproduced in outputs,' making the user, not the company, the source of anything sensitive that resurfaces later. Pi's own chat interface separately carries a notice that it is 'intended for users 18 years old or over' and routes new accounts through an age-verification step.

What the evidence says

The policy states Inflection uses inputs to 'develop and train our AI models,' and a user 'may opt out of model training' from account settings. It also says the company may 'use your inputs to make inferences about your emotional state, tone, sentiments,' with a separate opt-out for voice-based emotion inference. These are the company's own design choices, not an independently audited outcome; no regulator's or researcher's review of Pi's retention or training practice appears in either document consulted here. Pi's own account-settings copy corroborates the policy in narrower terms: a toggle for 'Use my data to improve models for everyone' and another for 'Use my voice data for emotion recognition and model improvement,' alongside a claim that 'chats are never shared for ads or marketing.'

What it asks of people

Using Pi means trusting that opting out of training and voice-emotion inference, switches that sit behind a settings menu rather than a default, will be honoured, and that a stated retention window applies after the fact. The policy states 'for users of Pi, we generally retain inputs for at most 15 days after user deletion,' but lists broad exceptions for fraud, security, legal and financial reasons with no further date attached. A user who wants a personal copy can use the in-app option to download their chat history, a self-service alternative to trusting deletion alone.

Privacy and safeguards

What is disclosed: the categories of data collected, a named 15-day post-deletion retention target for Pi inputs, and account controls over training and emotion-inference use. What is not disclosed: how long the fraud, security or legal exceptions typically last in practice, or any third-party audit confirming the opt-outs work as described. Neither document records a regulator's examination of this retention or training regime.

• Has the model-training opt-out been switched on, and does it persist across app updates?
• What counts as a 'fraud or security' reason that extends retention past 15 days, and who decides?
• Does downloading a chat history actually remove the underlying copy Inflection retains?

Read together, the policy and the product's own settings language describe a system built around opt-out controls and a headline retention figure rather than a guarantee that no trace of a conversation persists. That distinction, between a company's stated practice and a verified one, is the one worth carrying into any decision about what to tell Pi.