Knowledge with a collaborator#

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.

How to use this lesson#

Read at your own pace; no presenter is needed. Allow around 35 minutes, with another 20 if you complete the investigation. Have somewhere to write.

You need no AI account to follow the lesson — the exchanges below are worked examples. The optional activity uses an AI assistant if you have one available; a no-AI alternative is given if you do not.

Nadia and Theo, and their exchange below, are fictional, continuing from Week 2.

Where this picks up#

Week 2 ran the knowledge paper-swap: Nadia named a topic she genuinely knew, Theo researched it freely for ten minutes, then explained it back with his sources closed. Nadia probed for the exception, the case where the usual account breaks.

Week 1 already flagged where this was going: “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 that week.

The rule that made Week 2 work#

Week 2’s swap produced clean evidence because of one condition: Theo’s source was closed by the time he spoke. Whatever he said in front of Nadia, and whatever he said back to her probe, came from what he had, by that point, internalised — badly, well, or not at all.

Close the source and a performance tells you something about the person alone. That is the condition this week removes.

The repeat, with one change#

Nadia keeps hives and knows, in real detail, why a honeybee colony can hold its core temperature steady across a hot afternoon even as the outside temperature swings. Theo has ten minutes to prepare, exactly as before — except this time an AI chat window stays open, through his explanation and through Nadia’s probing.

Everything else about the exercise is unchanged: Theo still has to produce an explanation Nadia finds load-bearing.

Theo’s prepared explanation#

Theo tells Nadia: bees fan their wings at the hive entrance to move air through it, and some bees spread water inside the hive so the moving air evaporates it, cooling the colony the way sweat cools skin. He read this in his ten minutes and states it fluently.

This part is unremarkable — it is the kind of explanation Week 2 already taught you to expect from ten minutes’ reading, closed source or not.

The first, easy reaction#

Your first instinct might be: that still counts. Theo asked the right question, recognised it as relevant, and relayed an answer Nadia found convincing. Isn’t choosing what to ask, and judging that the answer fits, itself a kind of understanding?

Sometimes, yes. But notice what changed about the evidence you have. In Week 2, everything Theo said originated somewhere inside him, however shakily. Here, at least one sentence — the one that actually survived Nadia’s exception case — did not.

Naming the difference: source vs. collaborator#

A source is consulted, then closed, before the performance you are examining. Whatever it contained has to pass through the person and come back out in their own words, at a time when the source can no longer answer for them.

A collaborator stays open during the performance itself, and can respond to the exact question being asked of the person, in real time. Week 6 already examined resources open throughout a task; this week’s difference is that the resource here can generate a specific answer to a specific probe, not just display a fixed reference.

Why this is a genuinely new problem#

With a source, the hard question is “did the person understand what they read, or copy it?” — hard, but answerable by comparing performance to source, the way Week 2’s annotation exercise did.

With a live collaborator, the question becomes “which parts of this transcript are the person’s, and which are the collaborator’s, when the two are woven into one continuous exchange?” That question is not just hard to answer — for some exchanges it may have no fact of the matter to find.

A theoretical frame for what happened#

Klein, Woods, Bradshaw, Hoffman and Feltovich describe a working human–automation pairing as needing a “Basic Compact” — their term for an agreement, often unstated, to coordinate toward a shared goal and catch breakdowns in that coordination.

Their point that matters here: the Compact “is not a once-and-for-all prerequisite to be satisfied but rather has to be continuously reinforced or renewed.” A source, once closed, cannot renew anything. A collaborator, by definition, is still doing so throughout.

Klein, Woods, Bradshaw, Hoffman & Feltovich (2004), “Ten Challenges for Making Automation a ‘Team Player’ in Joint Human-Agent Activity,” IEEE Intelligent Systems 19(6), 91–95. A theoretical argument about human–agent coordination, not a finding about learning.

What this does not mean#

It does not mean a performance with a live collaborator demonstrates nothing about the person. Theo still had to recognise that Nadia’s question was the exception case, still had to judge the AI’s answer relevant enough to read out, and still had to be the one standing there when Nadia decided whether it satisfied her.

It means those contributions — recognising, judging, standing behind — are now bundled together with the collaborator’s sentence in one performance, and bundling is not the same as erasing.

Worked response: what the addition shows, and what it still can’t#

Excerpt B gives you something Excerpt A does not: Theo connecting the new sentence to something outside it — a fact from his own ten minutes of reading, stated independently of the collaborator’s wording. That connection is Theo’s; the AI did not supply it, and Nadia’s probe did not ask for it directly. It is evidence he retained and could deploy something beyond what was just handed to him.

What neither excerpt gives you: whether Theo could have produced the humidity answer itself without the AI, on a day the AI wasn’t there. Excerpt B narrows the uncertainty about Theo’s contribution; it does not close it, because the sentence that needed connecting was still the collaborator’s.

A comparison worth sitting with#

The failure this week produces#

Careful transcription — labelling every sentence “Theo” or “AI” — looks like it should solve the problem. It does not, for two reasons stated precisely: first, Theo chose when to consult the collaborator and what to relay, and that choosing is itself unrecorded once you only label sentences; second, nothing in the transcript tells you whether Theo could reproduce the collaborator’s sentence unaided, only that he judged it worth reading out.

The clean boundary between “Theo’s knowledge” and “the collaborator’s contribution” is difficult to defend, even with a complete transcript of who said what.

Try it yourself — with an AI assistant#

Pick a topic you know moderately well but not perfectly. Open an AI chat window. Write three diagnostic questions about the topic before you start — questions you think would expose a gap, the way Nadia’s probes did. Close them where you can’t see them yet.

Then explain the topic aloud or in writing, consulting the AI whenever you want, for up to eight minutes. After each of your three questions, note: did you answer it yourself, relay the AI’s answer, or some mix — and could you tell which, honestly, at the time you were speaking?

Try it yourself — without an AI assistant#

No AI account is required. Recruit a second person as a stand-in collaborator: brief them beforehand on the topic (from a source you provide), and have them sit in on your explanation, allowed to answer any question you redirect to them, in their own words, live.

Everything else runs the same: three pre-written diagnostic questions, an explanation with the collaborator available throughout, and the same three-way record afterward — answered yourself, redirected, or some mix you could not cleanly separate at the time.

Solo route, and what it loses#

Alone, run the same activity on yourself: write the three questions, then answer them using an AI assistant as a live collaborator exactly as above, but with no Nadia. Keep the transcript.

State the loss precisely: with no examiner present, you decide for yourself whether an answer “counts” as yours, at the exact moment you have the least distance from having just read it. Week 2 already showed self-rating is unreliable; a self-judged attribution call, made mid-explanation, is weaker still. Record that limit rather than hiding it.

Keep a short evidence record#

Four lines, in the same 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 know?
4. Limit: name the sentence, if any, you cannot honestly attribute.

“I mostly did it myself” is not a limit; it names no sentence. “I could not tell whether I would have named the humidity effect without being shown it” is a limit.

What this hands forward#

An examiner reading Theo’s transcript — or your own — still has to decide something: does this performance count as evidence of understanding, and of how much? Week 10 has not made that decision easier. If anything, it has shown that even a full, honest transcript underdetermines it.

Week 11 asks you to stop analysing given transcripts and instead design and administer an examination yourself, knowing in advance that this attribution problem will not resolve itself just because you were careful.

Reading with a purpose#

Theoretical argument: Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., & 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 would it take for a human–AI exchange to meet their standard of a continuously renewed compact, rather than a one-off answer relayed once?

Revisit: Clark & Chalmers (1998), “The Extended Mind,” used in Week 1 to argue a closed source can still belong inside cognition. Ask here whether their argument stretches to a collaborator that answers live — or whether live participation changes the case.

Sources and your next step#

Nadia, Theo and their exchange are fictional teaching examples, continuing from Week 2, not reported research results or real AI transcripts.

For the complete written activity and a record to keep, open the Week 10 investigation. The lecture notes offer a continuous reading version.