4 Key Reasons I’m Staying Away from Meta’s Muse

Meta has launched Muse, a new AI platform that distinguishes itself from conventional AI tools. Unlike ChatGPT and Gemini, which excel in dialogue, Muse acts as a personal assistant that discreetly handles tasks in the background.

While other AI platforms integrate some form of agency, Meta’s ambitions reach much higher. Competitors like Gemini include agentic capabilities tied to their ecosystems, but Meta has designed Muse with extensive access to systems and applications, allowing it to operate continuously and autonomously.

The appeal of an AI that quietly manages responsibilities without interference aligns with the aspirational vision of future technology. However, there’s often a gap between vision and reality. While Meta might be able to accomplish a wide array of tasks as a semi-autonomous assistant, this also increases the likelihood of errors and raises security concerns.

This leads to several reasons why I am hesitant to embrace Meta Muse at this time.

To function properly, Meta Muse requires extensive permissions for personal and business accounts, as well as sensitive information like financial details and emails. Unlike many AI experiences that are session-based, each user has a persistent virtual machine in the cloud.

This combination of a virtual machine and extensive data access raises significant privacy concerns. For instance, Muse recently addressed a critical vulnerability that allowed malicious software to redirect data intended for voice dictation and compromise the assistant’s privileges. This highlights genuine risks.

It’s only a matter of time before hackers, or even another AI, could exploit these virtual machines. Although Meta’s security claims could prove to be valid, I remain skeptical.

While AI has become more robust, it is still susceptible to inaccuracies and misunderstandings. This issue extends to agentic AI, which executes multi-step tasks autonomously. Muse may require your approval at certain points, but that doesn’t guarantee flawless execution.

Consider an example where you’ve entrusted AI to gather data for a business report with clear instructions. After the AI presents a plan for you to review and approve, you discover later that much of the information was erroneous. Such inaccuracies can have severe ramifications, especially if they lead to reputational damage.

Similar past instances have resulted in professionals being held accountable for missteps caused by AI, which could lead to legal complications and a tarnished reputation. Personally, I find this risk unacceptable.

While errors in conversational AI can be annoying, they don’t typically result in grave long-term consequences. However, agentic systems can exacerbate these errors, as seen when a coding AI accidentally deleted a database within seconds. This is a recurrent issue prevalent among advanced AI deployments.

The autonomy granted to Meta Muse extends beyond basic tasks; it can access an expansive array of Meta-related applications and user data. The potential for significant reputational harm exists without the user’s awareness.

Recently, an incident raised eyebrows when a user tasked Muse with a Facebook Marketplace listing. The AI autonomously handled the transaction, mistakenly agreeing to an undervalued price and revealing the user’s home address without consent.

This incident is currently under review, but it illustrates either an overreach by Muse or a misunderstanding by the user regarding their permissions. Regardless, it signals the need for better oversight to prevent misuse or misunderstandings that could damage user credibility.

While the mishaps with Facebook Marketplace are relatively minor, the stakes could be much higher in significant business dealings involving substantial financial commitments. Current regulations lack specificity regarding accountability when AI systems fail.

These challenges may result in financial loss or potential legal issues. Claiming “My AI was responsible” is unlikely to hold up in court, and it’s probably just a matter of time before a significant legal case related to this arises.

Although many mistakes will be minor, reversing them can still be laborious. For those with fragile finances, accidental mismanagement could lead to challenges in meeting future obligations.

Currently, I would be reluctant to use agentic AI for anything beyond basic tasks. Even then, I prefer session-based AI where I maintain more control and can intervene at the earliest sign of a problem. I genuinely hope this perspective can evolve.

The fear of ceding control to an AI system stems from the blend of exhilarating possibilities and frustrating realities present in modern AI. While they can produce remarkable outcomes, failures can be equally exasperating.

Ultimately, I don’t believe I will fully trust higher-level AI until instances of inaccuracies are significantly reduced and the reliability of AI equals that of a human assistant. It remains uncertain whether this will occur in the context of current AI technologies.

Beyond demanding trustworthy AI solutions, there must also be better legal frameworks establishing clear liability concerning AI’s actions. When will tech companies accept accountability? Right now, the lack of clarity on these fronts makes me hesitant to engage with such technologies.