Friends don’t let friends feed their data to AI agents

Friends Don’t Let Friends Feed Their Data to AI Agents

The rise of AI‑powered personal assistants promises convenience, but it also expands the reach of surveillance to anyone a user interacts with. Meta’s Muse, a recently launched chatbot that has already topped app charts, illustrates how these tools can turn friends, family, and even casual acquaintances into inadvertent data sources for a corporate AI.

How AI Assistants Harvest Third‑Party Data

Muse advertises itself as a helpful “genie” that can draft emails, schedule appointments, and even remember the plumber you hired last summer. To deliver that level of personalization, the app asks for deep access to a user’s device—bank accounts, contacts, chat logs, and more. The catch is that the assistant doesn’t stop at the installer’s own data. It pulls in information about anyone the user has communicated with, building detailed profiles of friends, coworkers, and past dates.

The underlying mechanism is a continuous profiling engine that updates a “relationship graph” hourly. It tracks how close you are to each contact, the topics you discuss, and even inferred preferences. This data is then used to generate recommendations, automate tasks, and—potentially—fuel advertising. The result is a sprawling dossier that extends far beyond the consent of the original data owner.

The User‑Interface Trap: Dark Patterns and Over‑Sharing

Muse’s onboarding flow is deliberately designed to encourage maximum data sharing. Options to limit access are buried beneath cheerful prompts and visual cues that make granting permissions feel like a natural part of the setup. Even when users think they are opting in selectively, the app can still harvest data through indirect channels such as full‑disk access on macOS or background sync services.

Reports from early adopters reveal incidents where the assistant accessed millions of text messages, disclosed home addresses, or even arranged in‑person visits based on data the user didn’t realize they had shared. In many cases, the root cause was a misconfigured permission rather than a malicious exploit, but the damage was already done—the assistant had already compiled the information.

Legal and Ethical Gaps: Consent That Doesn’t Exist

Privacy scholars point out that the consent model underpinning these AI agents is fundamentally broken. Users are asked to agree to sprawling terms of service that they rarely read, while the companies retain the right to process third‑party data without explicit permission from those individuals. This mirrors the “click‑wrap” consent practices that have long enabled the surveillance‑advertising economy.

Experts argue that placing the burden of privacy on ordinary consumers is unfair, especially when the technology is complex and the stakes are high. The reality is that even if you avoid installing an AI assistant, anyone you communicate with who does install it can expose your private conversations, location data, and personal history to the service.

What Can Users Do?

1. Audit Permissions Regularly – Review app settings on every device and revoke unnecessary access, especially to contacts, messages, and full‑disk permissions.
2. Limit Sharing in Conversations – Treat any platform that could be linked to an AI assistant as potentially public; avoid sharing sensitive details unless absolutely necessary.
3. Advocate for Stronger Regulation – Support policies that require explicit consent from all parties whose data is being processed, not just the primary user.
4. Consider Alternatives – Look for privacy‑first assistants that process data locally or offer transparent data‑handling policies.

The convenience of AI assistants is undeniable, but the hidden cost is a network of personal data that can be harvested without the knowledge—or consent—of anyone beyond the original user. Until the industry adopts clearer consent mechanisms and robust safeguards, the safest approach is to treat every interaction as potentially shareable and to keep a vigilant eye on the permissions you grant.