As of launch, Anthropic is the strongest of the tracked providers on consumer data handling, though the margin over OpenAI is narrow. Its consumer products (Free, Pro, Max) use a forced explicit choice rather than a silent default: training on your chats happens only if you opt in, and the prior no-training policy holds if you decline. Opting in extends retention to up to five years; declining keeps it at 30 days. Across all commercial, API, government, and education tiers, the terms contractually prohibit training on inputs. Deleted conversations are never used for training, and Incognito chats are excluded, though conversations flagged for safety review may still be analyzed. The main weakness is the long five-year retention window that attaches once a consumer opts in.

DeepSeek is the weakest of the tracked providers on the available evidence, though this assessment leans more on secondary analysis than a clean first-party control. Its privacy policy permits data use for model improvement, and no clear consumer opt-out is offered. Prompts may be shared within its corporate group. Its data scope is narrower than the social-graph collectors because it is a chat product rather than a social platform, which keeps it off the absolute floor on that factor, but it is weakest on consent and controls. Treat this profile as the least firm in the set and the one most likely to be revised as clearer documentation emerges.

Google's consumer Gemini is materially more exposed than ChatGPT or Claude. On standard consumer plans, data can be used to improve models unless you turn off Gemini Apps Activity, and a subset of conversations is routed to human reviewers and retained for up to three years disconnected from your account, surviving deletion. The policy itself warns users not to enter confidential information they would not want a reviewer to see. Activity is stored 18 months by default and around 72 hours even with activity off. The bright spot is enterprise: Google states that Workspace prompts, organization data, and AI responses are not used to train its core models. Deep integration across Gmail, Photos, and Drive widens the data surface well beyond the chat box.

Meta trains by default on public Facebook and Instagram content plus Meta AI interactions, draws on profile data, and under its 2026 terms also feeds conversational data into ad targeting. The opt-out is present but buried, and EU users who objected in 2024 were required to object a second time. There is no dedicated business no-training tier clearly published. It earns partial credit for excluding private messages and data from users under 18. Independent privacy audits have ranked it at or near the bottom of consumer AI tools, and the combination of broad social-graph training and advertising use is the core concern.

OpenAI sits just behind Anthropic. On consumer accounts (Free, Plus, Pro), data sharing for model improvement is on by default, but a single clear toggle under Data Controls turns it off, and a Temporary Chat mode bypasses training entirely. Business, Enterprise, Edu, and API usage is not used for training by default, a contractual stance comparable to Anthropic's. Deletions are normally processed within about 30 days, though an ongoing litigation hold currently prevents some logs from being purged on that schedule, which is the main live concern. The gap behind Anthropic comes down to the opt-out-by-default consumer posture rather than a forced choice.

xAI is among the most aggressive on data collection. Training is on by default, and for non-EU users it extends to the entire history of public X posts, whether or not the person ever uses Grok, alongside Grok conversations. The opt-out exists but is buried in X's settings, operates only prospectively, and can be silently re-armed by submitting feedback. There is no clearly published commercial no-training tier. Some relief exists: private X accounts are excluded from AI training, and a Private Chat mode deletes sessions in about 30 days. The defining concern is the breadth of default collection across a user's public social graph.