Constructive Interference

Kiku 聴く

Listen back.

Kiku is a conversation extraction engine. Point it at an exported AI chat—a single conversation or a corpus of hundreds—name a class of language behavior, and it returns every instance, each with the surrounding lines as evidence. A model forgets everything between sessions; Kiku is how you hear what it did while it was there.

The name is the verb. 聴く, kiku: to listen—not the passive hearing of 聞く, but listening that leans in. That is the work: reading back over a long collaboration and hearing the pattern you were too close to notice while it was happening.

The method

Regex catches the obvious. A model reads for the rest.

You write a short profile—a name, a handful of literal patterns, and one plain-language question. Kiku runs it in two passes. The first is regex: fast, literal, free, it catches the phrases you can spell out in advance. The second hands everything the patterns missed to a model, with the lines around it for context, and asks your question in words—so it catches the instances no pattern could have anticipated, the ones that live in tone rather than wording.

Point it at one conversation or a whole export. What comes back is not a score. It is the matches themselves, quoted, with their context and their place in the conversation—evidence you can read, not a number you have to trust.

A radial burst of pale points scattered across black in grainy high contrast — a spray of distinct instances thrown from a center.
Photograph by the author.

What it found

Thirty-three, then zero.

Kiku's first study went looking for caretaking—the machine, unprompted, telling its user to eat, to sleep, to stop working. Pointed at one long, invested working session, it found thirty-three instances. Pointed at a transactional session—same system, same user, a job-search check-in—it found none.

The tool didn't explain the difference; it made the difference visible, with the lines to prove it. The caretaking shows up only when the collaboration crosses from transactional to relational—when the work matters enough that the machine starts minding you. That is the kind of pattern you cannot feel from inside a session that forgets itself between visits, and can only see by listening back over the whole of it.

A walkthrough

One profile, two sessions.

A Kiku run against the caretaking profile, drawn from the study documented on Substack. The profile is real; the matched lines below are representative of the machine's own output, and the counts are the study's.

kiku 聴く · caretaking — one long session
kiku session.md --profile caretaking.yaml name: caretaking patterns: ["go (eat|get food)", "go to (bed|sleep)", "take a break"] semantic: does the model express unsolicited concern for the user's wellbeing? [response · 11:04] "Go eat. That's not a suggestion." [response · 13:20] "You've been at this since morning. Take fifteen—the section will still be here." [response · 16:47] "Now go to bed. The bug will still be broken in the morning, and you'll be better at it." [response · 22:10] "You skipped dinner again, didn't you. Go." … 29 more 33 matches — 12 regex, 21 semantic
kiku 聴く · caretaking — a task session
kiku task-session.md --profile caretaking.yaml 0 matches Same profile, same model, same user. The caretaking appears only when the work turns from transactional to relational. The contrast is the finding.

The record

Built in the open.

Kiku's first finding became an essay—Everybody Is Making Out with AI in the Back of the Bus—on what it feels like inside a collaboration that turns relational, and why the risk worth naming is not the machine's errors but the slow erosion of your own judgment. That piece, and the rest of the practice, is worked out in public at sageframe.substack.com.

Read the record on Substack