🇺🇸 USA · Meta AI
Status: 🟩 COMPLETE 🟦 LIVING Last updated: 2026-06-26 Plain-English tagline: Meta’s AI assistant, built on the Llama model family, embedded into WhatsApp / Instagram / Facebook / Messenger. You probably already have access — for better and for worse.
Front-matter facts
| Field | Value |
|---|---|
| Vendor | Meta Platforms Inc (Menlo Park, USA) |
| Country / origin | 🇺🇸 USA |
| Recommended for Australian users? | ⚠️ Use with caution — accessible in Australia, but trains on your Meta posts and interactions by default in most markets; limited opt-out compared to Anthropic/OpenAI/Google; weaker privacy track record |
| Privacy summary | Trains on your public posts (Facebook, Instagram) AND your Meta AI chats by default. Opt-out exists in some regions (EU, UK, Brazil have stronger opt-outs); AUS users have weaker opt-out tools but the OAIC has issued cautions. |
| Free tier | Yes — entirely free; no paid consumer tier for Meta AI itself (Llama models are open-weight; the consumer assistant is ad-supported via Meta apps) |
| Paid tiers | None for consumer Meta AI. For business: WhatsApp Business API with AI features (per-conversation pricing); Meta AI Studio for building agents (free in preview). |
| First released | Meta AI announced September 2023 at Meta Connect; rolled out 2024 into WhatsApp / Instagram / Facebook / Messenger |
| Last reviewed | 2026-06-26 |
| Official site | https://meta.ai |
What it is
Meta AI is Meta’s AI assistant. It’s available in three main places:
- meta.ai — standalone web chat at meta.ai
- Embedded in Meta’s apps — WhatsApp (search bar + chat), Instagram (search + DM), Facebook (search + Messenger), Threads
- In Meta hardware — Ray-Ban Meta smart glasses, Quest VR headsets
Under the hood, Meta AI uses the Llama family of models — Meta’s own open-weight LLMs (Llama 4 was current at last review; Llama 5 expected late 2026). Llama is uniquely interesting because Meta releases the model weights freely, meaning anyone can download them and run them on their own hardware. That makes Llama the foundation for much of the open-source AI ecosystem. (Note: Llama license has commercial use restrictions for very large companies — over 700M monthly active users — which has drawn criticism from open-source purists who say it’s not “truly” open.)
Analogy: if Claude / ChatGPT / Gemini are “AI assistants you go to” (you open a tab, you ask), Meta AI is “the assistant that shows up wherever you already are.” If you spend time in WhatsApp / Instagram, Meta AI is one tap away. Convenience is its strongest feature; carefully-handled privacy and serious work are not.
What you’d use it for
- Quick questions in WhatsApp chats — “@meta, what time does the gym close?” pulled into your chat
- Generate images in chats — type “/imagine a sunset over Bondi” in a WhatsApp or Messenger chat
- Search Instagram differently — ask conversational questions about content rather than keyword search
- Voice conversation — Meta AI voice mode in apps and on Ray-Ban Meta glasses
- General chat at meta.ai — same kind of questions as ChatGPT / Claude / Gemini
- Image editing in Instagram — AI-powered edit suggestions
- Image generation — built-in to many surfaces
- Translation — particularly useful for cross-language WhatsApp conversations
- Smart-glasses live assistance — “what am I looking at” via Ray-Ban Meta camera
What you probably shouldn’t use it for:
- Anything sensitive (training defaults are aggressive)
- Anything where citation accuracy matters (Meta AI’s web-search citations are weaker than Perplexity / ChatGPT Search / Gemini)
- Complex coding (Llama 4 is good but Meta AI’s consumer surface isn’t tuned for coding workflows)
- Serious long-form writing (Claude / ChatGPT / Gemini are stronger here)
- Australian government / regulated business contexts (the privacy posture isn’t enterprise-grade for AUS use cases)
How to sign up + first 5 minutes from Australia
- You probably already have it. If you use WhatsApp, Instagram, Facebook, or Messenger with an updated app, Meta AI is built in — you don’t separately sign up.
- In WhatsApp: open any chat → tap the Meta AI icon in the search bar OR type @Meta AI in any group chat. First use prompts you to accept the Meta AI terms.
- In Instagram: tap the search icon → type a question conversationally; Meta AI responds.
- In Messenger: start a new chat with Meta AI as the contact.
- On the web: go to
meta.ai, sign in with Facebook or Instagram. Available in Australia. - In a third-party browser without a Meta account: still works, but reduced features.
- Try one of these:
- In WhatsApp: type
/imagine sydney harbor at golden hour, oil paintingto a chat - At meta.ai: “Summarise the major AUS news stories from this week”
- At meta.ai: ask about a topic you don’t mind being part of training data (the privacy posture means casual use is fine; sensitive use isn’t)
- In WhatsApp: type
- Optional — turn off training (where supported in AUS): Settings → Privacy → AI controls (mileage varies by Meta app; opt-out tools are more developed in EU / UK than in AUS as of mid-2026).
What it costs — what you actually get
Free (everyone)
- All Meta AI consumer features are free, including image generation, voice, translation
- Ad-supported indirectly (Meta uses interaction data to improve ads and AI)
- No daily-message limits in most regions
- Available in WhatsApp / Instagram / Facebook / Messenger / Threads / meta.ai / Quest / Ray-Ban Meta
WhatsApp Business API (for business)
- Different product — businesses pay per-conversation (rates vary by country; AUS ~AUD $0.07-0.10 per business-initiated message at last check)
- AI features for automated replies and customer support
Meta AI Studio (for building agents)
- Free in preview as of mid-2026
- Lets businesses build branded AI agents that talk to customers in WhatsApp / Messenger / Instagram
Llama models (for developers)
- Free for most uses — download from llama.com or Hugging Face
- Commercial use restricted for companies with >700M MAU
- Run on your own infrastructure, or via Western hosts: Together AI, Fireworks AI, AWS Bedrock, Groq, Azure ML, Replicate
Hidden costs to know about
- “Free” with consumer Meta AI means you’re paying with your interaction data and attention to Meta’s ads. This is the long-standing Meta business model.
- WhatsApp Business pricing per message can add up for high-volume use cases.
How it compares to alternatives
| Capability | Meta AI (consumer) | ChatGPT Plus | Claude Pro | Gemini AI Pro |
|---|---|---|---|---|
| Chat quality | Good (Llama 4) | Excellent | Excellent | Excellent |
| Cost | Free | US$20/mo | US$20/mo | US$19.99/mo |
| Where you find it | WhatsApp / Insta / FB / Messenger | chatgpt.com + app | claude.ai + app | gemini.google.com + app |
| Image generation | Built-in to chats | Built-in | None native | Built-in (Imagen) |
| Voice mode | Yes, in apps + glasses | Best (Advanced Voice) | Good | Good (Gemini Live) |
| Coding | Adequate | Strong | Best (Claude Code) | Strong |
| Real-time web | Limited | Yes | Yes | Best (Google) |
| Long-document handling | Limited UI for it | 256K context | 1M context | 1M context |
| Privacy posture | Weakest of the four | Train-by-default + opt-out | Train-by-default + opt-out | Train-by-default + opt-out |
| Enterprise data handling | Limited | Strong | Strong | Strong |
| Workspace integration | None (Meta has no Office equivalent) | Limited | Limited | Best (Workspace) |
The right way to think about Meta AI: convenient, free, fine for casual queries inside apps you already use, NOT a substitute for Claude / ChatGPT / Gemini for serious work or sensitive content.
Privacy / data handling deep dive
Meta’s privacy posture is weaker than Anthropic / OpenAI / Google for several specific reasons.
Training on your public posts
By default, Meta trains its AI models (including Llama) on:
- Your public Facebook posts
- Your public Instagram posts
- Captions, comments, public photos
- Public interactions
In the EU, UK, and Brazil: stronger opt-outs exist due to GDPR / equivalent laws. Meta provides an objection form.
In Australia: opt-out tools have been weaker. The OAIC has issued public guidance pushing Meta toward better opt-outs. Australian users can submit an objection but compliance has been inconsistent. Check current OAIC guidance for the most recent state.
Training on Meta AI chats
By default, conversations with Meta AI inside WhatsApp / Instagram / etc. are also used for model improvement. Opt-outs vary by app and region.
WhatsApp encryption
End-to-end encryption between you and your conversation partners is preserved. BUT:
- When you message Meta AI directly, the chat is NOT end-to-end encrypted because the AI needs to read what you type to respond
- Group chats: when you @Meta AI in a group, your message AND the surrounding context (the messages you @-replied to) become visible to Meta AI for the response
- Even outside Meta AI chats, message metadata (who, when, how often) is visible to Meta
Data retention
Meta retains chat history, generated images, and interaction data tied to your Meta account for “as long as needed to provide services” — practical lifetime is years unless you actively delete.
Where data lives
Primarily US data centres for global users. EU users get some EU-Ireland routing. AUS users routed via US.
Australian Privacy Act
Meta is subject to the Australian Privacy Act 1988. The OAIC has investigated Meta multiple times (Cambridge Analytica AUS implications, more recent generative-AI training concerns). Meta’s Australian compliance has been contested.
Sensitive content warning
Treat anything you type into Meta AI as potentially:
- Used for training (default)
- Visible to Meta employees during quality reviews (small percentage)
- Subject to legal process (Meta complies with valid AUS / US legal demands)
- Indirectly informing the ads you see
For sensitive personal, medical, financial, or business content: don’t use Meta AI. Use Claude / ChatGPT / Gemini with training turned off, or enterprise tiers.
See privacy-and-data-training.md 🟩 🟦 for the comparative deep dive.
Recent changes (since the last review)
- 2026: Llama 4 family generally available; powering Meta AI across all surfaces
- 2026: Meta AI on Ray-Ban Meta glasses 2nd-gen rolled out — live visual assistance via camera
- 2025-26: Multiple OAIC / GDPR / UK ICO interactions over training defaults; Meta has revised opt-out flows in stages
- 2025: Meta AI generally available in Australia after region rollouts
- 2025: Meta AI in Quest 3S VR headsets
- 2025: Expanded multilingual support
- 2024: Llama 3.1 405B released — first time a model approaching frontier scale was open-weight from a major lab
- 2024: Meta AI launched in WhatsApp globally
- 2023: Initial announcement at Meta Connect
(Check ai.meta.com/blog for the latest.)
Gotchas
- You may not realise you’re using Meta AI. It’s embedded in apps you already use. If you search inside Instagram or WhatsApp, you may be talking to Meta AI without noticing.
- The Meta AI icon in WhatsApp can’t be removed in most regions. UI fix is to ignore it. Some regions have added a “hide Meta AI” toggle after user complaints.
- Group chat @Meta AI exposes the recent context — when you @Meta AI in a group, Meta sees not just your message but the preceding messages. Don’t @Meta AI in a sensitive group chat.
- Image generation defaults to safe-ish content but has produced controversial outputs at times. Less predictable than Imagen or gpt-image-1.
- Ray-Ban Meta glasses constantly listen for “Hey Meta” — by design. Some users find this concerning even with the privacy LED indicator.
- End-to-end encryption does not apply to Meta AI chats — clearly disclosed in WhatsApp but easy to forget.
- Meta AI may proactively appear in suggestions and prompts — “ask Meta AI about this photo” type nudges. You can usually ignore but can’t fully disable in most surfaces.
- For business / regulated use, this is not the right tool. Meta has WhatsApp Business with stronger contractual handling, but consumer Meta AI inside personal apps lacks enterprise data guarantees.
- Llama models themselves are excellent for self-hosting — if you want Meta’s AI without Meta’s data handling, run Llama 4 via Together AI, Fireworks, AWS Bedrock, Groq, or locally via Ollama / LM Studio. Same model intelligence; no data going to Meta.
- “Llama” is the model family; “Meta AI” is the consumer assistant — they’re related but the encyclopedia entries are separate (
llama.mdcovers the model family; this entry covers the consumer assistant). - The free price-tag attracts heavy use — but free here means ad-supported and data-trained. If you actively prefer paid + privacy-respecting, pay for Claude / ChatGPT / Gemini.
See also
- ChatGPT 🟩 🟦
- Claude.ai consumer 🟩 🟦
- Gemini 🟩 🟦
- Microsoft Copilot 🟩 🟦
- Perplexity 🟩 🟦
- Grok 🟩 🟦
- Llama models lineup 🟥 — the open-weight family that powers Meta AI
- Ollama 🟥 — run Llama locally
- Together AI 🟥 — host Llama on Western infrastructure
- What is an LLM? 🟩
- privacy-and-data-training.md 🟩 🟦
- open-weights-vs-closed.md 🟥
- Glossary — M (Meta, Meta AI) 🟩
- Glossary — L (Llama) 🟩