Ask most people which AI assistant they use and they’ll name ChatGPT, Claude or Gemini. Ask which one they touched most recently and the answer is often different: the little circle in the WhatsApp search bar, or an @ mention dropped into an Instagram group chat to settle an argument about dinner. By Meta’s own count, Meta AI has climbed toward a billion monthly users, largely because nobody had to sign up for anything.
That reach is the whole pitch. No new app, no separate tab, no subscription. Reach and usefulness aren’t the same thing, though. The Meta AI assistant is an uneven product: genuinely handy in a few places, oddly shallow in others, and wrapped in privacy defaults worth reading before you paste anything sensitive into it.
Where the Meta AI assistant actually lives
Meta’s strategy has been to drop the assistant wherever a conversation is already happening rather than build a destination people have to remember to visit. It makes for a slightly confusing spread of entry points:
- WhatsApp: the search field at the top of the chat list opens a private thread. Inside group chats, typing @Meta AI pulls the assistant in.
- Instagram: DMs and the search bar.
- Messenger and Facebook: the same @ mention trick works in a thread, plus a dedicated chat in each app.
- The standalone app for iOS and Android, which arrived in April 2025 and is where voice mode and the newer experiments tend to land first.
- Ray-Ban Meta and Oakley Meta glasses, where a tap on the frame lets you ask about whatever the camera is pointing at.
- meta.ai in a desktop browser, which remains the least cluttered way to use it.
The @ mention is the part people underestimate. Six people trying to pick a restaurant can type “@Meta AI we’re six people, £40 a head, one vegetarian, Saturday at seven, somewhere near Shoreditch” and get three options plus a rain plan in about eight seconds. That isn’t deep reasoning. It’s the right answer at the right moment, in a place everyone already is, and it’s why the product spreads without ads.
One caveat: conversations you have with Meta AI inside WhatsApp don’t get the same end-to-end encryption treatment as your chats with actual people. Meta has been reasonably clear about this in its help pages, and it’s the kind of detail people skim past.
Under the hood: Llama, and what it changes for you
Meta AI runs on Meta’s own Llama models. The current generation is an open-weight frontier multimodal mixture-of-experts model from Meta, which is a mouthful that translates into three practical things.
First, it reads images, not just text. Photograph a handwritten recipe, a parking sign in Portuguese, or a fridge shelf and ask what you can make. Second, the context window is long enough to hold a chunky document or a long group thread. Third, because it’s a mixture-of-experts design, only a slice of the model fires for any given query. That’s what makes it cheap enough for Meta to hand out free to a billion people, and cheap inference is why the assistant can afford to answer “what’s this plant” ten times a day without anyone paying a penny.
What it does well
Quick answers with zero ceremony
There’s no system prompt, no model picker, no token counting. Ask a question, get a short answer. For the 80% of daily requests that are small (convert this measurement, rewrite this text so it doesn’t sound passive-aggressive, what’s the difference between two phone plans), the lack of friction is the feature.
Image generation and edits that are genuinely fun
The Imagine feature handles text-to-image and, more interestingly, conversational edits: turn this photo into a watercolour, put a tiny hat on the dog, restyle the background. Results vary, hands still go wrong, and the safety filters refuse more than you’d expect. For quick visual play it’s a joy. If image work is the actual job, a dedicated tool such as Canva’s Magic Media for text-to-image and video gives you far more control over aspect ratios, brand assets and exports.
Hands-free help through the glasses
This is the least hyped and most useful surface. Standing in a supermarket aisle asking what to substitute for buttermilk, or pointing at a monument on holiday and getting the history without opening a browser, feels materially different from typing. The latency is noticeable but tolerable, and it works best in short bursts rather than long conversations.
Where it falls short
Depth is the big one. Ask for a 1,500-word briefing on a niche topic and you’ll get six bullet points and a chipper follow-up question. The assistant is tuned for brevity and conversational tone, which is right for WhatsApp and wrong for research.
Source transparency is thinner than the competition. It can pull from the web, and it will sometimes name a site, but the citation discipline you get from a research-first tool isn’t there. For anything you’d repeat in a meeting, verify independently.
Then there’s availability. Rollout in the EU and UK has been slow and stop-start because of data protection objections, and features appear in some countries months before others. Voice quality, glasses functions and app features all vary by region. If a feature you read about doesn’t exist on your phone, that’s probably why.
How to get better answers out of it
Meta AI responds well to constraints, badly to vague wishes. The single biggest upgrade is adding the details you’d tell a friend who already knows your life. If you want a broader method, this five-step workflow for using AI tools with real numbers is a solid template to steal from. In the meantime, three swaps that work immediately:
- Instead of “write a birthday message”, try: “Three sentences for my sister, 34, who just finished her first marathon. Warm but not gushy. She’ll appreciate a joke about the knee brace.”
- Instead of “summarise this chat”, try: “Pull out every decision we made about the venue, plus anything still unresolved, as two lists.”
- Instead of “make me an image of a beach”, try: “Overcast morning on a pebble beach, 35mm film look, low camera angle, one seagull, no people.”
A useful mental model: treat it like a capable stranger who knows nothing about you unless you say so. It also has a memory of sorts now, leaning on details from your profile and past chats, but it’s inconsistent. Don’t assume it remembers; repeat what matters.
The privacy trade-off, stated plainly
Meta trained earlier models on public posts and comments from adult users, and the company has acknowledged drawing on publicly shared content for AI features. Regulators in the UK and EU pushed back hard enough that Meta paused training and later added opt-out routes. What you get depends on where you live and when you check.
Practical hygiene, then. Open the assistant’s settings and look for the controls on using your activity to improve the models, and for the option to delete your Meta AI history. Turn off personalisation features you don’t want drawing on your profile. And keep work documents, client data and anything under NDA out of consumer chat windows entirely, because a free assistant embedded in a social app is not a confidential channel.
At work, it becomes a different product
Businesses mostly meet Meta AI through WhatsApp Business tools, where it can answer customer questions, qualify leads and hand off to a human. That’s a real use case with real money attached, and it’s also where the consumer product stops being adequate. There’s no audit trail, no admin console worth the name, no data residency options, and no way to constrain what the assistant says about your brand.
Bridging that gap is the same work as any production deployment: permissions, logging, evaluation and a way to roll back when the model confidently invents a refund policy. The architecture behind those systems is covered well in this breakdown of building secure, governed AI agents from prototype to production. And if the goal is an assistant that actually knows your customers, products and internal shorthand, the harder problem is context, which is exactly what building a context layer and a company brain sets out to solve.
Meta AI versus the assistant you already pay for
Pick by where your work already happens, not by benchmark charts. If your team coordinates in WhatsApp groups and your customer conversations happen in Instagram DMs, the in-thread convenience is hard to beat and no competitor can copy it. If you spend your day in long documents, code editors or research tasks, a dedicated assistant will beat it on depth, citations and control every time.
Most people will end up doing both, and that’s fine. Use Meta AI for the two-line question you’d otherwise type into a search engine and forget, for the group-chat logistics, and for the photo you want restyled for a laugh. Keep the real work, the long drafts, the sensitive files, the things you’d be embarrassed to get wrong, somewhere with a keyboard, a bigger model and an undo button.