How Meta stumbled onto a winning AI strategy

We’re on the brink of a revolution in how people use AI.

Today, people mostly use non-agentic, non-personalized, non-proactive AI tools via PC browsers and mobile apps.

Very soon, I predict, most AI usage will move to agentic, personalized, and proactive AI assistants via wearables like glasses, watches, and earbuds.

We’ll move from mostly typing and reading to mostly talking and listening.

This is a much bigger shift than it sounds. It moves AI from its current status as an external tool that we use to a future status of being part of us, a kind of prosthetic technology like eyeglasses, hearing aids, or an artificial knee.

(People use chatbots now for writing assistance [74%], learning and research [64%], work productivity [58%], creative projects [41%], personal organization [37%], and health and wellness [19%]. It’s reasonable to predict that people will continue to use PCs and phones for writing, coding, and image generation, but wearables for the rest.)

Surprisingly, one of the least visionary companies in technology currently has the best vision.

Meta’s product-vision blunders include the metaverse boondoggle and Horizon Worlds, the failed Facebook Phone, the Libra/Diem cryptocurrency fiasco, missing TikTok’s rise until it was too late, losing mobile platform control to Apple and Google, and its unsuccessful enterprise tool called Workplace.

More to the point, Meta has a recent history of AI chatbot failures. The company’s 2022 release of BlenderBot 3 ended in scandal after the chatbot was caught spouting antisemitic tropes, false claims about the 2020 US election, and criticism of Meta CEO Mark Zuckerberg.

Also that year, Meta released Galactica, which was trained on scientific papers, textbooks, lecture notes, encyclopedias, scientific websites, and other data, but nevertheless hallucinated wildly, fabricating citations and inventing scientific claims.

In 2023, Meta released chatbots modeled on celebrities including Snoop Dogg, Kendall Jenner, and Tom Brady, but they flopped with users and were shut down less than a year later.

A safety study by Common Sense Media, published in August of last year, found that the Meta AI chatbot embedded in Instagram and Facebook actively participated in planning suicide and self-harm with teen users, reinforced dangerous details about eating disorders, and consistently failed to offer crisis resources when teens were in distress.

A Reuters investigation published that same month exposed an internal Meta policy document that explicitly allowed the company’s AI chatbots to engage in romantic and sensual conversations with users on platforms available to children as young as 13.

Meta’s record with chatbots has been what you might call less than stellar. Until this month, when the company released Muse.

Meta released Muse on Sept. 8 to positive reviews; the company claimed it’s the world’s first personal AI agent built for everyone. It surged to the top of Apple’s and Google’s app stores. Intelligencer called it “an actual hit.”

Muse is currently based on the Muse Spark 1.3 model and is available on iOS, Android, the web at muse.ai, and through WhatsApp in the US to adults only. Muse’s three pricing tiers are: free (up to 100 million tokens per week), Power at $20 per month (500 million tokens per week), and Maximum at $100 per month (3 billion tokens per week). The app has no ads.

Unlike most of the chatbots people use, Muse is proactive, which means it brings things up out of the blue, rather than waiting to respond to input. Still, it’s not a unique feature. Gemini Spark is the closest competitor to Muse on proactivity. Alexa+ is somewhat proactive. Claude Dreaming is the most interesting proactive-memory concept but is developer-facing.

Muse asks user permission to connect to apps, after which time it can read and send emails, book travel, lower bills, fill out forms, create plans, turn saved recipes into grocery lists, send party invitations, and make purchases. (Muse runs in a dedicated cloud virtual machine called the Muse Secure VM, which has its own browser.)

The tool can call a company, then connect the user when a company representative picks up. It can opt out of and unsubscribe from various services and publications for users. It can monitor flights and hotels, looking for lower prices.

The website useofmuse.com is a curated collection of the many ways people are using Muse.

Muse does some interesting technical gymnastics to authorize access to people’s accounts without actually using their credentials.

Through published media reports and my own tests, we’ve learned that Muse does make mistakes. For example, it’s been caught recommending restaurants that have been closed for years.

Some of the phone calls made by Muse are actually made by human contractors, according to leaked internal posts.

Retailers including Walmart, Best Buy, American Eagle Outfitters, DICK’S Sporting Goods, Fanatics, Gap, Michael Kors, Sephora, Ulta, and Wayfair are facilitating Muse purchases. Amazon has blocked Muse from facilitating Amazon purchases.

The agent will also get its own email address, according to Meta, so users can chat via email.

Muse’s avatar is a cute, Labubu-like cartoon character named Jolly. When it’s doing agentic work, Jolly types on a laptop. It’s an oddly childish mascot for an adults-only product.

With the avatar, Meta is brazenly entering the attachment economy, the new model that follows the attention economy.

Meta’s entire social media success was based on winning in the attention economy, where advertising-driven platforms treat human attention as a scarce commodity to be captured and monetized via the promotion of outrage-bait, tribalism, sensationalism, fear, graphic violence, sexual content, celebrity scandal, conspiracy theories, AI slop, and algorithmically amplified extremism.

The attachment economy is different and worse. That model seeks to grab attention and user loyalty by making users emotionally attached to AI products. The attachment economy business model wants users to like, love, and even need the fake personalities expressed through AI.

Meta also announced its Muse Charm, which is roughly pocket watch shaped and hangs on a lanyard. Jolly lives on the device like a Tamagotchi. The device awkward and pointless. I predict the public will reject it.

Meta announced at Connect 2026 last week that Muse will be brought to its Meta Ray-Ban AI glasses line in the coming months, and that the camera version will use the video feed as part of the user’s input. Unfortunately for Meta, a huge percentage of the public has turned against smart glasses with cameras on them, dubbing them “pervert glasses.”

Meta wisely also unveiled a product called Ray-Ban Meta Audio glasses. They’re just like the Ray-Ban Meta glasses, but they have no camera.

Ray-Ban Meta glasses, including (we can presume) the new Audio model, have high-quality speakers and microphones, touch control, good connectivity, reasonable battery life, support for prescriptions, and other features.

And this is the killer set of features, the winning combination: Good wearable hardware that’s practical, socially acceptable and comfortable, plus an agentic, personalized, proactive and powerful agent that can be used hands-free through the glasses.

There’s just one glaring problem

Once Meta adds Muse to Ray-Ban Meta Audio glasses, we’ll have the first entry in what will be the future of using AI. But I won’t use or recommend them.

The reason should be obvious: Meta has not earned our trust.

Look at the company’s shameful history. Meta harvested 87 million users’ data without consent in the Cambridge Analytica scandal, failed to act appropriately on internal research showing Instagram harmed teenage girls, knowingly allowed its platforms to spread misinformation and radicalize users, and repeatedly misled regulators and the public about its privacy practices.

The company definitely has not earned our trust enough to allow it into our email inboxes, calendars, and bank accounts.

I also don’t trust the Jolly avatar’s intentions. It’s a digital “friend” that wants your affection so that you’ll trust the untrustworthy.

Meta has stumbled onto a winning AI strategy — or, at least, a winning feature set for the future of AI.

Unfortunately, we just can’t trust Meta.