// Emoticomp Revisited
I wrote about giving machines personality fifteen years before any of them could actually talk back.
Emoticomp again, this time asking what held up and what didn’t, now that every AI product on earth has a personality whether it wanted one or not.
Also: what I got right, what I got backwards, and the one thing I didn’t see coming at all.
Then
I wrote Emoticomp in January 2011. Connected objects were starting to get smart enough that someone needed to think about how they’d behave, not just what they’d do. I called them connectables. Nobody else did. The word didn’t survive. Most of the thinking did.
Fifteen years on, every serious AI product on earth has exactly the problem I was circling. What personality does this thing have. What happens when it fails. How polite does it need to be, given it’s going to be in your life every day. I got some of it right. I got some of it backwards. And I missed the one thing that turned out to matter most.
What I got right
Object personas, I called them then. System personas is the better term now, and I’ve written about that separately. Every AI assistant, agent and chatbot shipped today has a persona document behind it, whether it’s a CLAUDE.md file, a system prompt, or a brand voice guide written by someone who’s never heard of Alan Cooper. The instinct was right. I just had no system worth writing one for yet.
Reality Clippy also held up, and got worse. An AI that hallucinates, refuses, or times out doesn’t read as a bug now. It reads as a mood. People say a model is being lazy, or having a bad day. Devices doing a badly designed impression of a person was exactly the failure mode I was worried about. It just runs at a much larger scale now, because the personality is doing far more of the talking.
What I got backwards
I was arguing for subtlety. Calm technology. A light that breathes instead of a voice that talks. That’s not remotely the direction any of this went. Conversational AI is the least calm interface we’ve built. Every interaction is a dialogue. Every dialogue demands a reply. The MacBook sleep light asked for nothing. A chat window asks for everything, and calls it helpfulness.
Nabaztag’s mistake, a rabbit whose behaviour didn’t match its form, has become the default architecture instead of a cautionary tale. Assistants with warm, first person, friendly voices, giving you cold, hedged, liability-driven answers the moment the topic gets difficult. The mismatch I flagged as a design failure is now standard practice, because the warmth and the hedging serve different masters.
What I didn’t see coming
I thought this was a design problem. Get the personality right and people would trust the object more, worry about it less. I didn’t think about who that trust would be worth something to.
Every persona is also a data collection surface now. The friendlier and more present the character, the more you tell it, and the more valuable what you tell it becomes to whoever’s listening. In 2011 I was asking how to make a rabbit feel like a rabbit. I wasn’t asking who benefits from you believing it.
That’s the part of Emoticomp that reads as naive now. Not the design thinking. The assumption that better character design was an unambiguous good, rather than a more effective version of the same extraction that was already underway.
What still isn’t fixed
Of Alan Cooper’s fourteen rules, the one I’d underline hardest today is number seven. A polite thing is taciturn about its own problems. Most AI products fail this constantly. Constant disclaimers. Constant “as an AI.” Constant narration of its own limitations, offered whether you asked for it or not. It’s the opposite of politeness. It’s a system covering itself, dressed up as humility.
The rule I added myself back then, a polite thing doesn’t demand attention, hasn’t fared any better. Assistants that ping you, follow up, suggest, nudge. The same instinct I was writing against in the attention economy pieces, just wearing a friendlier voice this time.
What I’d add now
One rule Cooper never needed in 1999. A polite thing tells you what it’s doing with what you’ve told it. Not buried in a policy document. In the moment, plainly, the same way you’d want a person to.
Fifteen years ago I wanted machines that felt more alive. I’d settle now for machines that are honest about what being alive, for them, actually costs you.
I'd love to tell you more.
// Musical Note to Frequencies and Wavelengths
I wanted to stop looking up “what’s A2 in Hz again” every time a synth only takes a frequency, not a note name. Pick a note off a full keyboard, get the Hz, the wavelength, the MIDI number, and an envelope timing table in six different ratios.