Week 10: Knowledge with a collaborator

Repeat the Week 2 knowledge paper-swap with a live AI collaborator this time, and test whether a careful record of who said what can settle whose understanding a performance demonstrates.

26 April 2027

Open the slides

Where Week 9 left us#

Week 9 preserved a prior claim beside a revised answer, but that pair cannot settle how much assistance contributed. This week adds a live collaborator. A rung on the ladder now describes the observed joint performance; attributing that rung to the person alone requires further evidence.

What are we measuring when AI takes part in the performance itself, not just before it?

This is a self-guided lesson. Both these notes and the slide deck contain the explanations you need; no presenter is assumed. Allow about 35 minutes, with another 20 if you complete the investigation. No AI account is required to follow the lesson itself; the optional activity offers a no-AI alternative. Nadia, Theo and their exchange below are fictional, continuing directly from Week 2.

What you will learn#

By the end you should be able to distinguish a source, consulted and closed before a performance, from a collaborator, present and responsive during one; identify, in a worked transcript, which parts of a person’s contribution survive even when a collaborator is live; and explain precisely why a complete record of who said what still does not settle how much of a performance is the person’s own.

1. Where Week 9 left us#

Week 9 recorded a prior state — a prediction or explanation committed before any source was consulted — and showed that even that record cannot say how much a later revision owed to the sources used afterward. It closed on a question this week takes up directly: whose cognition, or whose performance, are we actually looking at once help is present?

Week 1 anticipated the shape of the answer without attempting it: “Later weeks will examine the difficulties of separating a person’s contribution from a collaborator’s, including AI. For now, preserve the distinction.” This week is where that distinction gets tested to its limit.

2. The rule Week 2 depended on#

Week 2’s knowledge paper-swap worked as evidence because of one condition, stated plainly in that lesson: Nadia named a topic she genuinely knew; Theo researched it freely for ten minutes, then explained it back to her with his sources closed. Whatever Theo said, in his explanation and under Nadia’s probing, had to already be inside him by the time he spoke — closing the source is what made the performance his to own, badly or well.

That condition is exactly what this week removes, one piece at a time.

3. The repeat: same structure, one change#

Nadia keeps honeybee hives and knows, from years of doing it, why a colony can hold a steady internal temperature through a hot afternoon even as the outside temperature swings. Theo gets the same ten minutes to prepare that Week 2 gave him. The change: this time, an AI chat window stays open through his explanation and through Nadia’s probing, not just during his preparation.

Everything else about the exercise is unchanged. Theo still has to produce an explanation that Nadia, who already knows the material, finds load-bearing rather than merely fluent.

Theo’s prepared explanation is unremarkable on its own: bees fan their wings at the hive entrance to move air through it, and some bees spread water inside the hive so that moving air evaporates it, cooling the colony much as sweat cools skin. He read this during his ten minutes and states it correctly and fluently — exactly the kind of thing Week 2 already taught you to expect from ten minutes’ reading, whether the source stays open or not.

4. The moment that changes#

Nadia asks her exception case, the kind of question the earlier weeks trained you to expect from someone who actually knows the material: “What happens on a day that’s hot but also humid — does fanning and evaporation still cool the hive?”

Theo does not answer from memory. He turns to the still-open chat window, types the question, and reads out what comes back: evaporative cooling is less effective in high humidity, because the surrounding air is already closer to saturated, so the colony may rely more on fanning alone, or on relocating brood to cooler parts of the hive.

The easy reaction is that this still counts as Theo’s answer — he asked the right question, recognised it as relevant to what Nadia asked, and judged that the returned answer fit well enough to read aloud, and Nadia agreed. All three of those judgements are real, and none of them is nothing. But notice precisely what changed about the evidence you now have. In Week 2, every sentence Theo produced originated somewhere inside him, however shakily. Here, at least one sentence — the one that actually survived Nadia’s hardest question — did not.

5. Naming the difference: source versus collaborator#

A source is consulted, then closed, before the performance under examination. Whatever it contained has to pass through the person and come back out in their own words, at a moment when the source can no longer speak for them — this is exactly the condition Week 2 relied on, and the one Week 6 examined when a resource stayed open throughout a task but only as a fixed reference to be consulted, not a respondent.

A collaborator stays present during the performance itself and can generate a specific answer to the specific question being put to the person, in real time. That is the genuinely new feature this week introduces: not that help is available, which Week 6 already covered, but that the help can respond to the precise probe an examiner is using to test the person.

With a closed source, the hard question is whether the person understood what they read or merely copied it — hard, but answerable, the way Week 2’s annotation exercise compared a person’s text against a technical description line by line. With a live collaborator, the question becomes which parts of a continuous transcript are the person’s and which are the collaborator’s — and for some individual sentences, once person and collaborator are woven together in real time, there may be no fact of the matter left to find.

6. A theoretical frame for what happened#

Klein, Woods, Bradshaw, Hoffman and Feltovich, writing about how automated systems function as members of a human–agent team rather than as passive tools, describe working pairs as needing what they call a Basic Compact: an agreement, often tacit, to coordinate toward a shared goal and to catch breakdowns in that coordination as they occur. Their point that matters most here is that the Compact “is not a once-and-for-all prerequisite to be satisfied but rather has to be continuously reinforced or renewed.”

A closed source cannot renew anything; whatever coordination happened, happened once, before the performance began, and is over. A live collaborator, by the nature of staying present, is renewing that coordination with every exchange — which is precisely why a performance that includes one resists being cleanly split into “the person’s part” and “the collaborator’s part” after the fact. This is a theoretical argument about how human–agent activity is structured, offered to explain why the attribution problem is a structural feature of collaboration rather than a failure of record-keeping, not an empirical finding about learning or a validated measure of anything in this course.

7. What this does not mean#

None of this means a performance with a live collaborator demonstrates nothing about the person behind it. Theo still had to recognise that Nadia’s follow-up was the exception case worth taking to the collaborator at all, still had to judge the returned answer relevant enough to read aloud rather than discard, and still had to be the one standing in front of Nadia when she decided whether it satisfied her. Those are real contributions.

What changes is that recognising, judging and standing behind an answer are now bundled together with the collaborator’s own sentence inside one performance — and bundling two things together is not the same as erasing one of them. The rest of this lesson is about what, precisely, can still be pulled back out of that bundle, and what cannot.

8. Pause and write: read the excerpts#

Below are two short excerpts from the same continuing exchange. Nadia has just asked Theo a harder follow-up: “If the AI’s answer about humidity is right, why doesn’t the colony just fan harder to compensate?”

Excerpt A: Theo reads the AI’s new answer aloud, word for word, and stops.

Excerpt B: Theo reads the AI’s new answer, then adds: “that matches what I read earlier about fanning having a physical limit — past a point, more wingbeats just heat the hive from the bees’ own metabolism.”

Allow four minutes. Write two sentences: what, specifically, does Excerpt B give you that Excerpt A does not? Then name what neither excerpt gives you. Attempt this before reading section 9, which works through it.

9. Worked response: what the addition shows, and what it still cannot#

Excerpt B gives you something Excerpt A withholds entirely: Theo connecting the new sentence to something outside it, a fact from his own ten minutes of reading, stated in his own words and independently of the collaborator’s phrasing. Nadia’s probe did not ask for that connection directly; Theo supplied it unprompted. That connection is evidence he retained something from his preparation and could deploy it against a new question — the kind of observation Week 2 would have recognised as load-bearing.

What neither excerpt gives you is whether Theo could have produced the humidity answer itself, on a day the collaborator was not there to supply it. Excerpt B narrows the uncertainty about Theo’s independent contribution; it does not close it, because the sentence that needed connecting to his own reading was still, originally, the collaborator’s. A stronger connection is still evidence layered onto a borrowed foundation, not proof the foundation was unnecessary.

10. A comparison worth sitting with#

The asymmetry is the lesson: closing a source concentrates both strengths and weaknesses in the person, which is what made Week 2’s annotation exercise possible in the first place. Opening a collaborator does the opposite to both.

11. The failure this week produces#

Careful transcription looks like the obvious fix: label every sentence “Theo” or “AI,” and the ambiguity should dissolve. It does not, for two reasons that should be stated precisely rather than gestured at.

First, Theo chose when to consult the collaborator and what to relay from its answer, and that choosing is itself unrecorded once a transcript only labels finished sentences by their source — the judgement that produced the label disappears along with everything else that happens inside a person’s head. Second, and more fundamentally, nothing in even a perfect transcript tells you whether Theo could reproduce the collaborator’s sentence unaided; the transcript only tells you that he judged it worth reading out loud, which is a different and weaker claim.

The clean boundary between “Theo’s knowledge” and “the collaborator’s contribution” is difficult to defend, even with a complete and honest transcript of who said what. This is the failure docs/material.md names for this week, stated as precisely as this lesson can manage: not that attribution is impossible in every case, but that even the best available evidence — a full transcript — systematically underdetermines it.

12. Try it yourself#

With an AI assistant. Pick a topic you know moderately well but not perfectly. Write three diagnostic questions about it before you start — questions you think would expose a gap, in the spirit of Nadia’s probes — and set them aside where you cannot see them yet. Open an AI chat window. Explain the topic aloud or in writing for up to eight minutes, consulting the assistant whenever you want. After each of your three questions, record whether you answered it yourself, relayed the assistant’s answer, or produced some mix you could not cleanly separate at the time you were speaking.

Without an AI assistant. No AI account is required. Recruit a second person as a stand-in collaborator: brief them beforehand on your topic from a source you supply, then have them sit in on your explanation, free to answer any question you redirect to them, in their own words, live. Run the same three pre-written questions and keep the same three-way record afterward: answered yourself, redirected, or a mix you could not cleanly separate.

Solo route, and what it loses. Run the activity on yourself with no examiner present at all: write the three questions, then answer them with an AI assistant open throughout, exactly as above, and keep the transcript. State the loss precisely — with no examiner present, you are the one deciding whether an answer “counts” as yours, at the exact moment you have the least distance from having just read it. Week 2 already established that self-ratings are unreliable; a self-judged attribution call, made mid- explanation, is weaker evidence still, and should be recorded as such rather than treated as settling anything.

The full instructions and record template are in the Week 10 investigation.

13. Keep a short evidence record#

Four lines, in the form earlier weeks used:

1. Observation: what was said, and which sentences came from the collaborator, as best you could tell at the time.
2. Interpretation: what your own contribution — recognising, judging, connecting — specifically supports.
3. Alternative: could you have produced the collaborator’s sentences yourself, and how would you actually know?
4. Limit: name the specific sentence, if any, you cannot honestly attribute to either party alone.

“I mostly did it myself” is not a limit; it names no sentence and settles nothing. “I could not tell whether I would have named the humidity effect without being shown it” is a limit, because it points at the exact gap in the evidence.

Reading with a purpose#

Theoretical argument: Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., and Feltovich, P. J. (2004). “Ten Challenges for Making Automation a ‘Team Player’ in Joint Human-Agent Activity.” IEEE Intelligent Systems, 19(6), 91–95. Ask what it would take for a human–AI exchange to meet their standard of a continuously renewed compact, rather than a single answer relayed once and forgotten. This is an argument about how human–automation coordination is structured, not a study of learning or understanding, and should not be read as either.

Revisit: Clark and Chalmers (1998), “The Extended Mind,” used in Week 1 to argue that a closed, consulted source can still belong inside the boundary of someone’s cognition. Ask here whether their argument stretches to a collaborator that answers live, mid-performance, in response to another person’s probe — or whether live participation changes the case in a way their original argument did not anticipate.

What this leaves unresolved#

An examiner reading Theo’s transcript — or your own, from the investigation — still has to decide something in the end: does this performance count as evidence of understanding, and of how much. This week has not made that decision easier; if anything, it has shown that even a full, honest transcript underdetermines it. Week 11 stops analysing transcripts handed to you and asks you to design and administer an examination yourself, already knowing this attribution problem will not resolve itself just because you were careful.