Alien Intelligence
The same lease. The same meeting. Two very different ideas about what matters.
You stand in an empty apartment while the broker looks for the key to the laundry room. From the kitchen window you can see the station. You walk to the smaller bedroom and discover that it is quiet enough to sleep in. By the time a forty-page lease arrives, you have already pictured a Tuesday there.
Give the lease to an AI and it starts in a different place. It can trace the early termination clause to an addendum, put the two passages side by side, and show you what you would owe if you left before the year was up. It has no Tuesday in that apartment. From the document alone, it has no way to know why the quiet bedroom matters.
The human can mistake a good apartment for a good lease. The AI can mistake a good lease for a good apartment. Each can bring the other evidence it would otherwise miss.
Yuval Noah Harari, the author of Nexus, has used “alien intelligence” to describe an intelligence made here on Earth that operates differently from ours. I want to take that phrase to the kitchen table. Say you have an AI in front of you. Where does its attention go that yours doesn’t? And what will it overlook unless you tell it to look?
The listener with no turn
Picture a long meeting. Someone says the supplier can deliver on Thursday. An hour later someone else says Thursday works if the power is connected by Wednesday. Most people leave remembering Thursday. The useful note is the condition, who owns it, and where in the recording you can hear it.
Listening in a conversation is already two jobs. In an experiment on turn taking, Sara Bögels and her colleagues found signs that people began planning their responses while the incoming question was still underway. That overlap helps an exchange move. It’s also why being responsible for answering is a different assignment from being responsible for the record.
A recording system can stay with the meeting without needing a turn. Afterward it can revisit the same stretch of audio or search the transcript for every mention of Thursday. But a machine can capture the qualifying words and still miss that the electrician sounded doubtful. It can also get the words wrong: Allison Koenecke and colleagues found that the version of Whisper they studied sometimes inserted phrases absent from the recording. A timestamp gives you somewhere to check. It doesn’t make the transcript true.
This is what I mean by a 180-degree difference. It’s a sketch of where the two kinds of attention point, not a score for intelligence. One participant is trying to live inside the conversation. A system given the job of making a record can keep returning to what was said. Put those jobs together and you get something neither offers alone: a room that can move forward and a record it can go back to.
When the arrows line up
Suppose I hand over four hundred pages about my father in the war. I ask for a list of inconsistent dates, spellings of the same name, and places where the order of events changes. I want page numbers and passages so I can make the calls myself. It’s a useful way to read a long manuscript, even though a model that can accept the whole thing can still miss a passage. In “Lost in the Middle,” Nelson Liu and colleagues found that the models they tested often used relevant evidence less reliably when it appeared in the middle of a long input. For work like this I’d check coverage section by section and ask to see the evidence.
Now give the AI a different instruction: Write the book. My reader gets the same kind of object that a human author could have made: a book with sentences, chapters, and my father in it. What changed was the role of the artifact. The first job gives the writer more to work with. The second can stand where the writer’s work would have stood. A book that helps me recall my father and a book that merely sounds like someone remembering a father can look remarkably alike on a shelf.
That’s where I expect to hear the word slop most often: when the outputs point in roughly the same direction and can substitute for each other. A drawing does it too: two very different processes can put a picture in the same empty frame. Some of the objection is about competition. Much of it is about the work: a paragraph can be fluent and still be generic, false, or indifferent to what a reader came for. In a short story experiment, Anil Doshi and Oliver Hauser found that access to AI story ideas improved how individual stories were rated on average while making the stories more similar to one another. A reader, a writer, and an editor looking at an entire magazine might have reason to give the same intervention different grades.
Some differences are closer to a right angle. At a park you might notice that the boardwalk would be a good place for a wedding. A vision system can be asked to mark the lamp fixtures visible in a photo, one by one. It still needs other views to inventory the grounds. Neither answer is the other answer, and each can tell you what to look for next.
What are you handing over?
When ChatGPT opened to the public in November 2022, the familiar relationship was a question and an answer. I use four words to describe how that relationship can change. They’re names for what we ask of an AI, even when much of the underlying system stays the same.
- Assistant
- Draft my email about the shuttle. I still address it and send it.
- Agent
- Send that email to the property manager. One authorized action has an end.
- Delegate
- Keep track of my move. Follow up on missing documents, watch the deadlines, and bring me decisions only I can make. The assignment survives the conversation.
- Talent
- I seek out this particular AI songwriter because I would choose its songs from a lineup. Its distinction is the work people come for, beyond its ability to take an assignment.
There’s a difference between how a system decides its next action and who is responsible for the work on Tuesday morning. A delegate needs time, a way to notice changes, a record of what happened, and a way to reach you when its authority runs out. I’d let it chase a missing attachment on its own. An appointment it can’t confirm, or a change to what I’m agreeing to pay, should come back to me. If I still have to check whether it remembered to follow up, I’m still carrying the move.
I think talent is the next frontier. We’ll know it when audiences seek out the work of a particular AI by name. A chart that shows what the system can do won’t tell us whether someone plays its song again on the drive home.
The part you bring
I’d give the AI the lease and tell it why I want the apartment. I’d ask it to point to language that could undo those reasons, and to list what the document can’t answer. Then I’d walk home from the station. The AI can send me to page 27. It can also send me down a street I wouldn’t otherwise have thought to walk.
The useful question about an alien intelligence is where it can direct your attention next. Sometimes it will be a sentence you would have skipped. Sometimes it will be the distance between that sentence and the life you thought you were signing up for.
3 thoughts on “Alien Intelligence”