Headline: New Study Finds ChatGPT, Claude, Grok and Perplexity Leak Chat Data to Meta, Google and TikTok — What Crypto Users Need to Know
A fresh analysis from IMDEA Networks Institute—released May 4 as the LeakyLM project—shows that the four biggest consumer AI chat services (OpenAI’s ChatGPT, Anthropic’s Claude, xAI’s Grok and Perplexity) quietly share user data with third‑party advertising and analytics services, including Meta, Google and TikTok. Researchers detected more than 13 embedded trackers across those platforms, none of which are explained to users in plain language.
What’s being sent and why it matters
- Trackers can “phone home” whenever you open a chat: they report who you are, what page you’re on, and in some cases identifiers tied to your session. The most basic data point is a conversation URL (a permalink) that points to a specific chat.
- Permalinks can be powerful. Several platforms make some conversation URLs publicly accessible by default, which means anyone with the link can view the full chat without logging in. When those URLs are forwarded to ad networks such as Meta or Google, those companies gain the technical capability to access the conversations.
- “Leaking a URL is not just metadata—it can be equivalent to leaking the conversation itself,” the researchers write. They emphasize they have not proven that ad companies actually read chats, but the infrastructure and data flows required to do so exist.
Platform-by-platform highlights
- Grok (xAI): The riskiest in the study. Guest conversations are public by default, and TikTok’s tracker received Open Graph metadata that included verbatim message content—effectively a preview or screenshot of your chat.
- Claude (Anthropic): Stronger access controls (chats aren’t public unless shared), but the platform still transmits conversation permalinks and identifying data to advertising systems. Crucially, data to 11 advertising platforms flows via Anthropic’s servers (server‑side), so browser ad‑blockers won’t stop it.
- ChatGPT (OpenAI): Also not public by default, but sends permalinks and advertising cookies—Google Analytics still runs for free logged‑in users, and rejecting cookies where possible reduces exposure.
- Perplexity: Removed its Meta tracker last month; still included in the study’s findings before that change.
Why crypto users should pay attention
- Crypto traders, developers and privacy‑conscious users frequently paste sensitive information into chat apps—wallet addresses, private API keys, trading strategies or research. Even if an AI provider claims chats are private, leaking a permalink or embedded preview to ad networks could expose that content to third parties.
- Server‑side sharing (as with Claude) bypasses common client‑side protections like ad‑blockers, raising the bar for technical mitigation.
What you can do right now
- Grok: Set conversation visibility to private and revoke any previously shared links.
- Claude: Reject non‑essential cookies to disable the Meta Pixel; be aware server‑side sharing may still occur.
- Perplexity: Set conversations to Private.
- ChatGPT: Reject cookies where possible; realize some analytics (like Google Analytics for free logged‑in users) may still run.
- For stronger protection, follow a full AI privacy guide (researchers and privacy outlets publish step‑by‑step advice).
Context and next steps
- The researchers say providers’ privacy controls “may mislead users by implying stronger protections than are actually enforced.”
- IMDEA Networks submitted the findings to Data Protection Authorities on April 13, 2026, and notified xAI on April 17. As of publication, none of the companies had publicly responded.
- The team plans to extend LeakyLM to Meta AI, Microsoft Copilot and Google Gemini; those were excluded here because those services operate as both AI providers and ad platforms, which complicates the threat model.
This study doesn’t prove ad networks have read your chats, but it does show that the plumbing is in place. For crypto users—where a single leaked link or string of data can have real financial consequences—the findings are a timely reminder to audit what you paste into chatbots and to lock down sharing settings wherever possible.
Read more AI-generated news on: undefined/news
