Why AI privacy is confusing
The single most important fact about AI privacy is that the same company often treats your data differently depending on which product you use. A provider’s consumer chat app may use your conversations to improve its models by default, while that same provider’s paid API and enterprise tier contractually commit not to train on your inputs. People get burned because they assume “I trust this brand” covers every surface — but the free app, the mobile app, the API, and the enterprise deployment can each have different defaults for collection, training, and retention. To compare tools honestly you have to compare the specific product and account type, not just the logo.
The four questions that actually matter
For any AI tool, answer four questions. What is collected? — usually your prompts, the outputs, and metadata like timestamps and device info. Is it used for training? — the headline question; consumer tiers often default to yes, business and API tiers usually to no. How long is it retained? — ranging from days for abuse monitoring to indefinite unless you delete, and deletion is sometimes delayed rather than immediate. Can you opt out? — most consumer apps now offer a training toggle, a “temporary chat” mode, or a data-controls page, but you have to find and set it; it is rarely off by default. Run every tool through these four and the marketing language stops mattering.
The general pattern across major tools
Without fixating on figures that change, a reliable pattern holds across the big names. Consumer chat apps (free and personal paid tiers of the popular assistants) lean toward using conversations for improvement unless you opt out, and they keep data for monitoring even when you do. Developer APIs from the same companies generally do not train on your inputs by default and retain data only briefly for abuse detection. Enterprise and team plans add binding no-training terms, admin controls, configurable retention, and sometimes data residency. Local or on-device models (open-weight models you run yourself) are the privacy ceiling — nothing leaves your machine — at the cost of setup effort and weaker capability than frontier hosted models.
Practical guidance
Match the tool to the sensitivity of the data. For casual, non-sensitive use, the consumer apps are fine — just turn off training if you would rather not contribute. For work involving customer, financial, or regulated data, use an enterprise agreement or the API, never paste secrets into a free chatbot, and confirm the no-training and retention terms in writing. For maximum privacy, run an open model locally. Whatever you choose, treat the provider’s current published terms as the source of truth — these policies change, defaults flip, and a setting that protected you last year may have moved. The safe habit is simple: assume anything typed into a hosted tool is stored, and only relax that assumption when the terms explicitly say otherwise.