Focus of the letter 48

In this week’s readings, we come across two AI agents: Dots at OpenAI, Muse at Meta.

They can remain active even when the app is closed. It is possible to connect them to our emails, calendars and accounts. Other apps already offer these features, but these two agents stand out for their plush appearance. Muse offers Jolly, a small, furry and customizable creature, and the Dots take the form of colorful balls whose appearance is unique to each user. However, both are aimed at adults.
We can’t really test them here: Muse is not available in Europe and Dots are reserved for OpenAI Pro subscriber accounts.

According to Emily Turrettini who can be found in the patargées readings and which relies on PCMag, Mashable and Wired, an American user entrusted the sale of a keyboard to Muse. The agent independently accepted an offer that was $100 lower than the asking price, communicated the seller’s address and told the buyer that the seller was at home. The buyer showed up without the seller knowing that an appointment had been scheduled. The user thought that the agent would ask for his approval again: the incident seems to be related to a confusion around the “Allow Always” authorization. OpenAI assures that the Dots will ask for authorization for sensitive actions.

The avatar chosen is not insignificant. Meta is looking for an experience that is “less strange” than talking to a logo, and a Harvard researcher quoted by Wired explains that personification and “cuteness” make technology warmer and less threatening. However, we are not wary of a stuffed animal and what we entrust it with is the possible access to our entire digital environment.

This week’s apps also offer a discreet echo of it. MAI Chatter allows you to converse orally with Microsoft’s MAI Playground assistant, by choosing one of the two voices offered: no plush on the screen, but choosing a voice is already choosing an element of your avatar.

The process is not new. More than forty years ago, the film Gremlins laid down rules: Gizmo is adorable, as long as you don’t expose him to bright light, don’t get him wet, and never feed him after midnight. Thirty years ago, the Tamagotchi reversed the contract: the little creature required care, and we became attached to it by dint of ensuring its survival.
Then, in 1998, came the Furby, whose resemblance to Gizmo was so obvious that Warner and Hasbro released an official version with Gizmo’s voice the following year. That same year, according to a document from the US Federal Aviation Administration, the NSA had even banned the Furbys from its premises, for fear that they would repeat what they heard. According to their inventor, they were not recording anything.
However, the fear has not disappeared: we have all wondered about the fact that our smartphone could listen to us continuously. In 2021, the CNIL’s innovation laboratory noted that no research had shown systematic listening by microphones for advertising purposes, and considered the hypothesis to be discarded, “at least for the moment”. The Northeastern study cited revealed applications with excessive permissions, some of which capture the screen to transmit it to third parties.

With Jolly and the Dots, access is no longer imposed, it is consented. With Gizmo, characters inadvertently broke the rules. Here, the rule is an authorization and an “Allow Always” is enough.

In recent weeks, two stories have finally dominated our shared readings: that of the “IApocalypse” and now that of round and colorful avatars. Two ways of telling the same technology, and not looking at what it is allowed to do alone. If the agent looks like a toy, who still thinks about the authorizations he is given?

Echo:

• Letter 33, “What if we chose the level of autonomy of the machine?”
• Letter 5, “Meta AI, or the discreet invasion of generative AI in our social networks“.