The Cocktail Party Problem

I present to you a tweet

Katie Notopoulos@katienotopoulos

The thing about using facial recognition to solve “the cocktail party problem” where you don’t recall someone’s name is that [barring the rare medical issues] this problem has already been solved via basic social graces. “I’m so sorry what’s your name again?” Boom problem solved

nxthompson @nxthompson

Meta’s AI glasses are one of the company’s most successful products in years. They offer genuine utility, with a direct way for users to interact with its AI. But people have long been wary of smart glasses—especially when face recognition is involved. Here’s @boztank on Meta's

10:49 PM · Jul 14, 2026 · 18.1K Views

10 Replies · 22 Reposts · 328 Likes

Yep! Agreed! But also, I hadn’t thought until seeing this tweet the degree to which this “cocktail party problem” holds true for other celebrated uses for AI.

Now caveat whenever I’m presenting a new harebrained Grand Unified Theory: the only universally true Grand Unified Theory is that all Grand Unified Theories are at least a little bit fictional1. So too will it be with this.

But think about what the glasses are trying to solve in this context (beyond the fact that it’s already a solved problem as Katie points out above.) You are talking to someone who remembers you, and you either remember them but don’t remember their name, or you don’t remember them at all. When that comes to light, you feel embarrassed, and if you’re being charitable to the validity of the scenario as something to “solve”, then I suppose also the other person might feel embarrassed or hurt or offended. Having your glasses remind you prevents that situation, and prevents any bad feelings on your side or the others.

And from a pure I guess consequentialist utilitarian side, sure! If the only thing that matters is the actual feelings existing in the world or not, then yay we’ve prevented bad feelings, seems good.

Why would someone feel embarrassed or hurt about not remembering (on either side)? Because it is, potentially, in the most uncharitable case, implying that the person is not worthy of being remembered.

What is the state of the world in this case? It’s that you actually did forget. If the implication of forgetting is the person is not worthy of being remembered, then that it is still true, even if your face phone reminds you of it.

This might be needlessly pedantic, but in a very real sense the presentation of “I remembered” when in reality it’s “my glasses remembered for me” is a lie! A potentially-low-stakes lie, but a lie nonetheless. I suppose again, consequentialistly maybe that “a lie was told” < “people don’t feel bad.” But I am not a pure consequentialist, and so I find it icky.

To somewhat immediately contradict myself, I do have some people’s birthdays on my calendar, and use that to remind myself to wish them a happy birthday. That is similar to the remembering faces and names bit, in that I am using technology to artificially remind myself of something I would otherwise forget. In a sense it is as much a lie as confidently saying the person’s name at the cocktail party.

But in another sense it’s not: when I wish someone happy birthday, is what I’m saying by that act primarily “I remembered your birthday” or “I thought your birthday was worth committing to memory”? Or is it primarily “I thought your birthday was worth remarking on” and “I wanted to let you know I was thinking of you”? I think it’s mostly the latter! In fact, the remembering or not remembering is incidental to the actual thing, which is the wishing of the happy birthday. The looking at the calendar and making a choice to reach out is what matters, IMO. The thing behind the thing is “I wanted and decided to wish you happy birthday”, which is still true even if the reminder is artificial.

I would consider a lie to be asking my new robot AI agent to send out an automatic happy birthday message without my involvement, or even having it draft one for me (or, non-AI, just have a calendar automation that does it without any AI involvement at all.) That then is presenting “I cared enough to wish you happy birthday”, which is in fact false! Extra specially the case if it was something like “Go through my emails or Facebook, find all the birthdays, and then do an automated message.” At least in the first case I’m still making a deliberative choice of which birthdays to add to the calendar, which I guess is something, just barely.

Again, this is all low-stakes for the most part (though the worse I get at synchronous communication anxiety, the happy birthday messages do actually become key ways of at least staying in some form of touch with those I care about!!) But I do think it resonates across a lot of AI use cases: how do I use an artificial language tool to present a state of the world and a version of myself that is fundamentally untrue.

Not all AI tools are like this! Despite my qualms about AI coding, at least in that case it’s primarily does the code actually exist and do what you want that matters. When the code is created, depending on how you talk about it you may or may not be projecting a false reality where you did the creating instead of the robot, but that is secondary to the existence and success or failure of the artifact itself. And it’s also different from even other AI glasses tools like translation or captioning; you are not presenting a false world where you can now understand French (and the tools and their latency don’t even allow for that even if you tried!)

But a lot of ways AI is pitched for interpersonal uses, especially consumer interpersonal use, do seem to fall into this category. In fact, as I was in the process of drafting this, I did literally see someone who I otherwise mostly like say that a “useful” version of using AI agents for commerce would be for it to go through your contacts and buy and send birthday presents for people automatically. No!!!!! Gross!!!! Cocktail party lies!!!!! And of course as always, the far and away primary consumer-level product market fit for AI is clearly: cheating on homework. Which again is using AI tools to present a fictitious version of yourself (as one who knows things and can do them.)

Anyway, doesn’t actually matter, but you have full permission if you ever see me at a cocktail party to ask “I’m sorry, have we met?” And perhaps I’ll feel a little offended, but secure in the idea that I’m at least interacting with a real person.