Meta says its Muse AI agent can do things for you. I put it to the test

Cheng Xin/Getty Images

It’s been a busy afternoon for Pip, the AI agent I’ve been using in Meta’s Muse app.

The agent, which Meta launched on September 8, booked a Friday date night and made a packing schedule for my upcoming move. It also emailed a friend about a trip next month and made a planning document with places we want to visit.

A Pixar-like avatar typed away on a laptop at the top of the screen to let me know my AI helper, whose name I chose, was hard at work.

Chatbots like ChatGPT have changed the way people find information and get answers to questions. Now, tech giants are attempting to develop AI assistants that can handle complex tasks, with the goal being that a user can give a broad command like “order my groceries for the week” and go about their day.

Muse is currently the most popular free app in the iPhone’s App Store and has been downloaded more than 2.5 million times since its launch, according to market intelligence firm Sensor Tower . It is Meta’s latest and most promising attempt to rival popular chatbots like OpenAI’s ChatGPT and Google’s Gemini.

Brendan McDermid/Reuters

Meta just picked a side in a big debate over the future of AI

Meta CEO Mark Zuckerberg is expected to share more about his vision for Muse and the company’s AI plans during its Connect conference on Wednesday.

Perhaps most importantly, there’s the question of whether people will trust Meta with even more personal details: Muse’s rise in the App Stores comes after the company’s smart glasses stirred controversy and privacy concerns.

“While feedback (around Muse) has been positive, privacy and trust remain key considerations for broader adoption,” Bank of America analysts wrote in a research note Tuesday.

Muse has more guidance for users compared to ChatGPT or Gemini, which can feel like a blank canvas.

The Muse app has five tabs: the main chat for communicating with the agent, a customizable feed of your interests and priorities, an Ideas tab with prompt examples, a list of goals to track based on previous conversations, and a section for files and media.

For my upcoming move to a new apartment, I told Muse what I still needed to pack, how much time I had before moving and that I wanted to pack in short spurts.

The AI agent made a list of what to pack each day this week, along with how much time I should spend on packing each day. It also gave general tips like setting aside a specific bag for the first night in my new apartment with all my essentials.

Samuel Boivin/NurPhoto/Getty Images

Like other AI agents, it’s good at planning-oriented tasks that usually require some combination of Googling, communicating with friends or family and booking reservations.

After connecting my Google apps to Muse, I instructed the agent to email a friend about what she had in mind for our trip to Cape May, New Jersey, next month. When she responded, I had Muse package her restaurant recommendations into a polished Google Doc, along with similar alternative options.

One thing that impressed me was Muse’s ability to understand and communicate roadblocks in tasks. When I had trouble purchasing a bottle of shampoo, it scanned the website to find out why. (It turns out the store it was trying to buy from was a beauty supply store that requires a professional license number).

Muse uses a virtual computer in the cloud for tasks like these, meaning it’s actually browsing the web to accomplish its assigned job. You can see it navigating websites and jump in at any time.

Still, other AI agents like ChatGPT and Gemini can handle similar tasks and sometimes provided more detailed responses.

Muse also served as a reminder that AI isn’t ready to handle everything. When I asked for a list of places to go for date night, it pulled up two options in my neighborhood that had been closed for years.

In one instance, Muse seemed to exaggerate its capabilities.

In its list of prompt ideas, Muse said it could haggle on Facebook Marketplace: “When the right one lands I message the seller, talk the price down against comparable listings, and lock in a time and place,” the app’s Ideas tab said.

But when I tried this, Muse said it couldn’t message sellers directly. It would instead vet the listings and draft a message for me to send to negotiate a lower price.

A Meta spokesperson said the description in the prompt ideas list is accurate and that Muse can negotiate with sellers under parameters set by the user.

Meta’s Mark Zuckerberg lays out new philosophy on AI

Shopping is also often easier to do manually, largely because of restrictions that prevent bots from crawling websites. Amazon prevents Muse from browsing its online store, although Meta’s agent can still retrieve product links through search engines.

While Muse was able to browse Target’s website, I still had to enter personal details manually because the retailer blocked autonomous clicks at checkout. (Muse uses Stripe’s Link digital wallet service for purchases, which it says generates a one-time use card to prevent your credit card number from being shared.)

Meta and its peers are betting their future on AI agents at a time when consumers are increasingly skeptical of the technology.

Roughly two-thirds of American adults say AI is advancing too quickly, and more Americans predict AI will have a negative impact on society rather than positive, according to a Pew Research Center report from June.

Then there’s the privacy question and whether consumers will trust Meta to manage payment information and access apps on their behalf, especially after the Cambridge Analytica scandal . And it’s not just Meta; 41% of Americans said they have very little or no confidence in large technology companies in a Gallup poll published in July.

Hollie Adams/Bloomberg/Getty Images

‘Turn the cameras off:’ London’s growing privacy pushback against smart glasses

“AI agents are often framed as helpful personal assistants, but to live up to their potential they typically need an enormous amount of private data to function,” Miranda Bogen, director of the Center for Democracy & Technology’s AI Governance Lab, told CNN in an email. The fact that agents are designed to take action based on this information means the bar is even higher for safeguards, she added.

Meta said it built Muse to be “safe, secure” and “private” and that it runs on a dedicated cloud-based computer that can’t be accessed by other agents. It also said Muse cannot see a user’s passwords or payment methods.

AI companies claim that AI agents are well-suited for online shopping. But market research firm Gartner found that 64% of consumers said they think AI shopping agents will benefit corporations and brands more than consumers.

And despite Muse’s early App Store success, a survey from research firm Emarketer suggests ChatGPT and Gemini are the most widely used AI assistants in the United States by a long shot. When asked which AI tools they prefer or use the most, 33.7% of respondents said ChatGPT, while 18.2% said Gemini and 4.7% said Meta AI, the firm reported in July.

Still, Muse is Meta’s best chance yet at giving them serious competition. At least for now.

程鑫/Getty Images

对于我在 Meta 的 Muse 应用中使用的 AI 代理 Pip 来说，今天下午过得很忙碌。

Meta 于 9 月 8 日推出的这款智能助手，帮我预订了周五晚上的约会，还为我即将到来的搬家制定了打包计划。此外，它还给我的朋友发邮件，询问下个月的旅行计划，并生成了一份包含我们想去景点的旅行规划文档。

屏幕上方出现了一个类似皮克斯动画中的虚拟形象，它正在笔记本电脑上敲击键盘，告诉我我选择的人工智能助手正在努力工作。

像 ChatGPT 这样的聊天机器人改变了人们获取信息和解答问题的方式。如今，科技巨头们正试图开发能够处理复杂任务的人工智能助手，其目标是让用户只需发出“帮我订购一周的食材”这样的简单指令，就能继续他们的日常活动。

据市场调研公司Sensor Tower的数据显示，Muse目前是iPhone App Store中最受欢迎的免费应用，自发布以来下载量已超过250万次。它是Meta公司最新推出的、也是最有希望与OpenAI的ChatGPT和谷歌的Gemini等热门聊天机器人竞争的产品。

布伦丹·麦克德米德/路透社

Meta刚刚在一场关于人工智能未来的重大辩论中表明了自己的立场。

Meta 首席执行官马克·扎克伯格预计将在周三的 Connect 大会上分享更多关于他对 Muse 的愿景以及公司人工智能计划的信息。

或许最重要的是，人们是否会信任 Meta 并提供更多个人信息：Muse 在 App Store 中的崛起，正值该公司推出的智能眼镜引发争议和隐私担忧之后。

“尽管（关于 Muse 的）反馈是积极的，但隐私和信任仍然是更广泛采用的关键考虑因素，”美国银行分析师周二在一份研究报告中写道。

与 ChatGPT 或 Gemini 相比，Muse 为用户提供了更多指导，而 ChatGPT 或 Gemini 则给人一种空白画布的感觉。

Muse 应用有五个标签页：用于与经纪人沟通的主聊天、可自定义的兴趣和优先事项动态、包含提示示例的“想法”标签页、根据先前对话要跟踪的目标列表，以及文件和媒体部分。

为了即将搬到新公寓，我告诉 Muse 我还需要打包哪些东西，搬家前还有多少时间，以及我想分批打包。

人工智能助手列出了本周每天需要打包的物品清单，以及每天应该花多少时间打包。它还给出了一些实用建议，比如为入住新公寓的第一晚准备一个专门的包，装上所有必需品。

Samuel Boivin/NurPhoto/Getty Images

与其他人工智能代理一样，它擅长以计划为导向的任务，这些任务通常需要结合使用谷歌搜索、与朋友或家人沟通以及预订服务。

将我的 Google 应用连接到 Muse 后，我指示代理程序给一位朋友发邮件，询问她对下个月去新泽西州卡普梅旅行的想法。收到回复后，我让 Muse 将她推荐的餐厅整理成一份精美的 Google 文档，并附上一些类似的备选方案。

令我印象深刻的一点是 Muse 能够理解并传达任务中的障碍。当我购买一瓶洗发水遇到困难时，它扫描了网站以找出原因。（结果发现它尝试购买的商店是一家美容用品商店，需要专业许可证号。）

Muse 使用云端的虚拟计算机来执行此类任务，这意味着它实际上是在浏览网页来完成分配的任务。您可以随时查看它浏览网页的过程并加入其中。

不过，像 ChatGPT 和 Gemini 这样的其他人工智能代理也能处理类似的任务，有时还能提供更详细的回复。

Muse 也提醒我，人工智能还无法处理所有事情。当我询问适合约会的场所时，它却推荐了我家附近两家已经关门多年的餐厅。

在某一场合，Muse 似乎夸大了自己的能力。

Muse 在其提示创意列表中表示，可以在 Facebook Marketplace 上讨价还价：“当遇到合适的商品时，我会给卖家发消息，根据类似的商品价格进行议价，并确定时间和地点，”该应用程序的“创意”选项卡中写道。

但当我尝试这样做时，Muse 提示无法直接向卖家发送消息。它会审核商品信息，然后自动生成一条消息供我发送，以便协商更低的价格。

Meta 的一位发言人表示，提示创意列表中的描述是准确的，Muse 可以在用户设定的参数范围内与卖家进行协商。

Meta公司的马克·扎克伯格阐述了人工智能的新理念

手动购物通常也更容易，这主要是因为网站受到诸多限制，阻止了机器人抓取网站内容。亚马逊禁止 Muse 浏览其在线商店，但 Meta 的代理仍然可以通过搜索引擎获取产品链接。

虽然 Muse 可以浏览 Target 的网站，但我仍然需要手动输入个人信息，因为该零售商在结账时阻止了自动点击。（Muse 使用 Stripe 的 Link 数字钱包服务进行购物，该服务声称会生成一次性使用的卡片，以防止您的信用卡号被泄露。）

Meta 及其同行将未来押注于人工智能代理，而此时消费者对这项技术却越来越持怀疑态度。

根据皮尤研究中心 6 月份的一份报告，大约三分之二的美国成年人认为人工智能发展过快，而且越来越多的美国人预测人工智能将对社会产生负面影响，而不是正面影响。

此外，还有隐私问题，消费者是否会信任Meta来管理支付信息并代表他们访问应用程序，尤其是在剑桥分析公司丑闻之后。而且，这不仅仅是Meta的问题；盖洛普7月份公布的一项民意调查显示，41%的美国人表示他们对大型科技公司几乎没有或完全没有信任。

Hollie Adams/Bloomberg/Getty Images

“关掉摄像头”：伦敦民众对智能眼镜的隐私保护抵制日益高涨

“人工智能代理通常被视为有用的私人助理，但要充分发挥其潜力，它们通常需要大量的私人数据才能运行，”民主与技术中心人工智能治理实验室主任米兰达·博根在给CNN的电子邮件中写道。她补充说，代理被设计成根据这些信息采取行动，这意味着安全保障的要求更高。

Meta公司表示，Muse的设计理念是“安全可靠”和“私密”，它运行在专用的云端计算机上，其他代理无法访问。该公司还表示，Muse无法查看用户的密码或支付方式。

人工智能公司声称人工智能助手非常适合在线购物。但市场研究公司 Gartner 的调查发现，64% 的消费者认为人工智能购物助手对企业和品牌的益处大于对消费者的益处。

尽管 Muse 在 App Store 上早期取得了成功，但市场研究公司 Emarketer 的一项调查显示，ChatGPT 和 Gemini 才是美国使用最广泛的 AI 助手，遥遥领先。该公司在 7 月份发布的报告显示，当被问及他们最喜欢或最常使用的 AI 工具时，33.7% 的受访者选择了 ChatGPT，18.2% 选择了 Gemini，4.7% 选择了 Meta AI。

不过，Muse 仍然是 Meta 目前为止最有希望与他们构成真正竞争的乐队。至少现在看来是这样。