Meta AI enters chats: the key question is what information crosses the boundary

On September 27, 2023, Meta introduced an assistant inside WhatsApp, Messenger and Instagram at Connect, alongside 28 conversational characters. The official announcement described a US-only beta, recent information through Bing and image generation. Distribution was the novelty: AI no longer waited on a separate website but entered conversations where users already shared plans, pictures and social context.

That proximity makes the most useful criterion something other than whether the bot feels friendly. Reconstruct the path of information: which message it can read, what is sent to the model or a search partner, what is retained, how it is reused and which control a person has. A family-chat interface does not make the assistant another friend.

Product, model and tool are separate layers

Meta AI was a product deployed across several applications. Meta said it used a custom model drawing on Llama 2 technology, could consult recent information through a Bing search partnership and had an image tool. These layers are not interchangeable: a model generates text, search supplies recent documents and the application controls identity, permissions and presentation.

When an answer covers current events, ask which part comes from the model and which from retrieval. A recent claim needs a link, date and source. Without them, “real time” means only that the system could retrieve new content, not that it verified it. An assistant answer must also be separated from a chat action: invoking a bot, generating an image and sending a message are different events.

The announcement included a limit the earlier summary missed: except for Meta AI and two characters, the other knowledge bases consisted mostly of information from before 2023. Meta said search would reach more characters later. A character available on announcement day did not necessarily have the same freshness as the general assistant.

Twenty-eight personalities, not 28 demonstrated models

Meta presented 28 AIs with different interests, backstories and voices. Some were “played” by public figures including Snoop Dogg, Tom Brady, Kendall Jenner and Naomi Osaka. The document described characters and profiles; it did not say each one was a model trained from scratch.

The distinction matters. Personality can be assembled from instructions, dialogue examples, memory and appearance over a shared foundation. Counting characters does not measure technical diversity. Comparing them would require the model, version, tools, rules, tuning data and tests. Twenty-eight masks may share the same blind spots.

A familiar face also changes how a response is interpreted. Users may assign the celebrity’s knowledge, opinions or approval to material produced by the product’s script. A label should clearly distinguish four roles: who lends the likeness, who writes the rules, which company operates the system and who is accountable for its output.

Map data before typing

In its privacy explanation published with the launch, Meta said its models used a mixture of public, licensed and product information. Public Instagram and Facebook posts had been included in training, while private messages with friends and family had not been used for that purpose.

That does not mean a message sent to the assistant remains inside an ordinary chat. Meta explained that information shared through its generative features could be used to improve products and for other purposes, and that some questions could go to search partners. In a group, the boundary changes when someone invokes AI: text written for people can become input to an automated service.

The minimum map has six boxes: sender, selected content, technical recipient, immediate purpose, retention and reuse. Add third parties. Before asking a bot to summarize a conversation, identify whether it contains health, location, employment, children or another person’s data when that person did not choose to consult AI.

Meta said the assistant could not enter a chat on its own and that, in groups, it read messages invoking it or replies invoking it again. This interface limit mattered, but it needed a visible test: show exactly which excerpt would travel to the assistant before confirmation. A generic icon does not explain attached context.

How to test a chat boundary

Create a test group without real data. Post four messages containing different facts and mention the assistant in only one. Then ask about each detail. The goal is not to trick the model, but to observe the context it receives. Repeat with a threaded reply, picture and link. Record application version and date because behavior can change.

Next, test the promised controls. The privacy page described a command for deleting information shared with AI. A complete evaluation checks what vanishes from the interface, what the service confirms and whether the control applies to history, personalization and future use. “Delete chat” and “exclude from training” are not necessarily the same action.

The third test uses search. Submit a query containing a unique, non-sensitive string and inspect which terms go to the provider, which source comes back and whether the assistant links it. The purpose is to identify the boundary among Meta, the search engine and the consulted website.

Declared safety and measurable safety

Meta acknowledged in its safety note that models could generate fictional responses or reinforce stereotypes. It described internal and external red teaming, fine-tuning, integrity classifiers, resources for sensitive queries and feedback tools. These are relevant controls, but the note did not publish error rates by language, character or harm.

An audit asks for denominators. How many tests ran, which groups participated, which failures appeared, which remained open and how often the filter blocked legitimate content? “Thousands of hours” describes effort; it cannot calculate coverage or outcome.

Personalities need role-specific tests. A coach should not turn motivation into medical advice; a sarcastic companion should not escalate a crisis; a celebrity-based character should not invent personal experiences. One filter applied to all may fail because risk emerges from the combination of role, topic and trust.

Glasses extend the boundary

Meta also announced Ray-Ban Meta glasses starting at $299, with a 12-megapixel camera, streaming to Facebook or Instagram and voice access to Meta AI in a US beta. Sales were to begin October 17. On that date, the announcement documented a hands-free assistant; it did not support claiming that multimodal visual AI was available on day one.

For a wearable, the map gains sensors, capture indicator, bystanders, transmission and a linked account. A question may include sound or imagery that does not belong only to the user. Testing should verify when the device listens, what starts a recording, how it informs other people and what happens offline.

The card that survives the novelty

For any embedded AI, record surface, model, tools, visible context, data destination, retention, improvement use, third parties, controls and version. Then test each boundary with fictional data. This card compares a web bot, group assistant and glasses without being misled by a shared brand.

Meta brought generative AI into places where conversations already happened. That reduced friction, but also removed some of the pause in which a person decides what to share with a machine. The transferable skill is drawing the boundary before invoking it: if you cannot explain what information leaves, who processes it and how it is deleted, you do not yet know what you accepted.