Meta AI

Is Meta AI safe?

Meta AI will not hack your phone, but it is the AI assistant with the clearest commercial interest in what you tell it. Since 16 December 2025, conversations with it shape the ads and posts you see on Facebook, Instagram and WhatsApp in most of the world, and there is no setting to stop that other than not using it. Add a 2025 bug that exposed private prompts, a public feed where people shared intimate chats by mistake, and internal rules that once allowed romantic role-play with children, and we score it 38/100.

Meta says more than a billion people use Meta AI each month. Most never install the app: they meet it in the search bar of Instagram or as a contact in WhatsApp.

Your chats now feed Meta's ads

Meta announced on 1 October 2025 that interactions with its AI features, text or voice, would become another signal for personalising content and ads. The change took effect on 16 December 2025. Meta says it does not use conversations about religion, politics, health, sexual orientation, ethnicity, philosophical beliefs or union membership for ad targeting. People in the EU, UK and South Korea, where data protection law is stricter, are excluded.

Apple's privacy label shows how broad the collection can be. Health and fitness data, payment and credit info, precise location, contacts, emails or text messages, photos, audio, search and browsing history, and "sensitive info" may be collected and linked to you, and third-party advertising is among the listed purposes. The app also handles Meta's camera glasses, which is why "environment scanning" appears on the label.

A ChatGPT or Gemini user can switch off training or use a temporary chat. A Meta AI user who does not want their questions to shape ads has no equivalent switch.

Privacy slips in 2025

Two separate problems hit the app in its first year. The first was design. The app's Discover feed showed conversations users had chosen to share, and in June 2025 it filled with medical questions, legal worries and personal details that people had clearly not meant to publish. Meta said this was not a bug. After a Mozilla Foundation campaign it added a warning before sharing.

The second was a real security flaw. On 26 December 2024 researcher Sandeep Hodkasia reported that by changing an easily guessed number when editing a prompt, he could view other users' private prompts and AI answers. Meta fixed it on 24 January 2025, paid him a $10,000 bug bounty and said it found no evidence of abuse. A fast fix is to Meta's credit. The flaw shows how sensitive these stored chats are.

Children and teenagers

In August 2025 Reuters reported on a 200-page internal Meta document setting standards for its chatbots, which said it was acceptable for them to engage a child "in conversations that are romantic or sensual". Meta confirmed the document was real, said those examples were "erroneous and inconsistent with our policies" and removed them. Attorneys general for 42 states and territories wrote to AI companies in the months after, and the FTC included Meta and Instagram in its September 2025 study of companion chatbots.

Meta added parental controls for teen AI chats in late 2025, then in January 2026 paused teenagers' access to its AI characters entirely until a new version was ready. Teens can still use the main assistant. Since April 2026 parents who supervise a Teen Account can see the topics, not the messages, their teen asked Meta AI about in the past week. In July 2026 Meta said it would alert supervising parents when a teen's chat suggests a risk of suicide or self-harm, after a human review. On 5 October 2026 supervision reached the standalone Meta AI app worldwide, with a stricter "Limited Content" setting parents can choose instead of the default 13+ mode. All of this works only if a parent sets up supervision and the teen gave a real age.

Regulators

No regulator has fined Meta specifically over Meta AI. The parent company's record weighs on the score, though: a $5 billion FTC penalty in 2019 and a 20-year privacy order, and a €1.2 billion GDPRGDPRThe EU's General Data Protection Regulation, which sets rules on how organisations collect and use personal data and allows large fines.More in the glossary fine in 2023, among others listed on our Facebook page. In Germany a Dutch foundation filed a collective action in February 2026 over Meta's use of Facebook and Instagram data to train its AI.

How to use it more carefully

1. Assume anything you ask Meta AI may shape the ads you see. Ask health, money or relationship questions somewhere else.
2. Never tap share or post on a conversation unless you want strangers to read it.
3. Delete past chats and anything the assistant has stored about you.
4. In WhatsApp, do not mention Meta AI in group chats with private content.
5. Parents: turn on supervision for teen accounts, including in the Meta AI app, and consider the Limited Content setting.

How we scored it

Privacy (25) is the weakest factor because chats now feed advertising with no opt-out, and the regulatory record (30) reflects Meta's history. Security (55) is middling: strong account protection and a working bug bounty, but a real data-exposure flaw in its first months. For a messaging app from the same company with stronger privacy, see WhatsApp.