The Assistant in the Room

At Meta Connect on September 27, 2023, the important AI announcement was not just another assistant that could answer questions. It was where Meta intended the answers to appear: inside WhatsApp, Messenger and Instagram, with Ray-Ban Meta smart glasses and Quest 3 named as further destinations. An assistant would enter places people already used to talk to one another.

Rebecca Bellan’s contemporaneous TechCrunch report described Meta AI as a US beta, with the hardware integrations still coming. That distinction matters. A product presentation can establish a design direction without proving that every promised surface is ready, or that people will welcome the company into the discussion.

A shortcut with a speaking voice

Meta’s own example was disarmingly ordinary: friends choosing a hiking trail in Santa Cruz. The assistant could surface options directly in the group chat. Afterward, it could generate a digital badge commemorating the outing.

There is a genuine convenience here. Nobody needs to leave the conversation, search elsewhere and return with a handful of links. A shared question gets a shared answer in the place where the decision is being made. Helping people coordinate something they actually want to do is a better use of computing than manufacturing another reason to keep scrolling.

But the shortcut changes the arrangement. The platform is no longer merely carrying the group’s suggestions. Its assistant supplies suggestions of its own.

According to Bellan’s report, Meta’s generative-AI executive Ahmad Al-Dahle described an “orchestrator” behind the assistant that would identify intent and route a prompt to an appropriate extension. Web search through Bing was the first extension. Meta’s announcement likewise described real-time information through its Bing partnership, alongside image generation.

That is not one all-knowing conversational mind. It is an assembled service: interpret a request, select a capability, retrieve or generate material, and present an answer. A friendly tone can make those operations feel continuous. It does not make their different failure modes disappear. Search access is not a guarantee that the resulting answer is correct.

Familiarity is not accountability

The same launch also introduced a cast of AI characters, including characters embodied by celebrities and influencers. Meta explicitly said it wanted interaction to feel like talking to familiar people. Separately, it outlined plans for creators to build AIs that would extend their virtual presence, with creator sanction and direct control.

Those are different relationships. Asking an assistant for trail suggestions, chatting with a fictional character and addressing a creator’s authorized stand-in should not require the user to untangle who is speaking after the conversation has begun.

The coherenceist test is not whether an interface feels connected. It is whether the connection preserves the participants’ ability to understand and shape it. In a group chat, convenience for the person invoking an assistant is not automatically agreement from everyone else about its role. That is a design question, not evidence that Meta’s launch tools secretly read every message or lacked any controls; these sources do not establish either claim.

A useful assistant should make its status legible, distinguish retrieved information from generated suggestions, and leave people room to continue without it. These are standards for the relationship, not a claim that the September announcement demonstrated them all.

The launch deserves credit for locating AI in a concrete activity: friends trying to make a plan. It also exposes the larger ambition. Meta was offering not just tools for producing messages, but new parties to message and new voices within existing exchanges.

The hiking group still chooses the trail. The company providing the room has started offering directions.