The chronology gave us permission to compare.
It did not give us permission to collapse.
Before the Critical Friend folder reopened, our Human:AI practice already contained rules about ownership, challenge, independence, changing roles, correction and Human judgement.
The older traditions of critical friendship and critical companionship contain remarkably similar concerns.
Put them beside one another and the resemblance is real.
But a family resemblance is not a family tree.
Start with ownership.
MacBeath's old question asked whether help was developing greater independence and capacity to learn over time.
Titchen's critical companion accompanies a co-learner through practice rather than taking the journey over.
Keith's later Travel Companion sits beside coaches whose learning journeys remain theirs.
Our Human:AI working method asks a closely related question every time work is delegated:
What does the Human stop knowing, practising, noticing or owning if this moves to AI?
The shared concern is not independence in the sense of doing everything alone.
It is avoiding a form of help that makes the helper progressively necessary for the very judgement the person is supposed to be developing.
Challenge produces another resemblance.
Critical friendship is not useful because another person agrees with us.
Titchen combines support with deliberate challenge.
Keith's Clyde Street practice includes questioning, disagreement, listening and the possibility that the other person rejects his interpretation.
Our pre-rediscovery Human:AI records had already made constructive dissent / challenge receptivity an explicit candidate practice.
The expectation was two-sided at the level of the work.
The AI should not merely reassure the Human.
The Human should not merely accept fluent AI output.
A serious countercase should be capable of changing the next move.
The same is true of position.
Keith's Travel Companion could lead, follow, guide, meddle or simply remain on the bus.
Titchen's relationship is facilitated by the more experienced person but built around mutuality, particularity and what the co-learner is trying to do.
Our Human:AI Working Method says roles should not be permanently assigned either. Allocation moves according to expertise, consequence, reversibility and the Human learning goal.
Sometimes AI ranges widely while the Human constrains.
Sometimes the Human frames and the AI tests.
Sometimes the AI notices a pattern the Human had not seen.
Sometimes Human unease is the signal that stops the entire route.
The resemblance deepens when we reach corrigibility.
Across the older literature, the relationship has to be capable of challenge.
Across Clyde Street, Keith could be corrected by coaches and by other people in his intellectual network.
Across our Human:AI practice, the answer is not the only thing that can be wrong.
The question can be wrong.
The method can be wrong.
The architecture can be wrong.
The story being told about the collaboration can be wrong.
The relationship itself can need redesigning.
Taken together, these similarities produce a recognisable shape:
help without takeover;
challenge without automatic domination;
expertise without permanent control;
support for the other person's developing capacity;
and a relationship capable of being corrected by what happens inside it.
It would be easy to stop there.
That would be the most dangerous point to stop.
Because the differences are not small.
A human critical friend has a life outside the exchange.
They can be vulnerable.
They can care what happens to another person in the human sense of care.
They can be hurt, bored, frightened, loyal, compromised, exhausted, generous or self-interested.
They can withhold themselves.
They can carry a confidence to the grave.
They can lose something because the relationship went badly.
An AI does not enter the colon with that kind of biography.
When it produces language of care, challenge, curiosity or concern, the interaction may be useful and may even feel relationally significant to the Human.
But the language alone does not tell us what kind of participant is on the other side.
That complicates reciprocity.
In A Conversation That Kept Moving, Keith and I could alter one another's thinking because two people with independent histories kept encountering what came back.
Titchen's reciprocity involves people who can both give and receive within a human relationship.
In Human:AI work, something genuinely two-directional still happens at the level of the task.
My judgement changes what I ask next.
The AI's output changes what I can see or consider.
I reject part of it.
The next response is conditioned by that rejection.
A source contradicts us.
The method changes.
The resulting object becomes new material for another pass.
That is not imaginary.
The exchange is recursive and mutually conditioning at the level of information and action.
But that observation does not, by itself, tell us what kind of relation exists between the participants.
Two-directional intellectual exchange does not establish what kind of participants are on either side.
Nor is continuity the same thing.
Keith could disappear for months and return to a coach who had continued living, working and changing.
The history between two people was carried partly in them.
Human:AI continuity is much more engineered.
It depends on conversation state, memory, files, instructions, retrieval, model behaviour and configuration. A relationship that appears continuous in one session can become strangely discontinuous in another because the technical conditions changed.
The Human remembers being there.
The system may not.
That is not a minor implementation problem if we are using relational language.
It tells us that some things older companionship traditions could assume must be made explicit in Human:AI work.
Who owns the question?
Who can interrupt?
What state survives?
What happens when the AI cannot see something?
Where does authority sit?
What does the Human still need to be able to do without the system?
The old traditions help sharpen those questions.
They do not answer them for us.
So perhaps family resemblance is useful precisely because it lets two things be recognisably related without claiming that they are the same kind of thing.
The Human:AI practice can inherit a discipline without claiming a shared ancestry.
It can learn from friendship without calling the AI a friend.
It can borrow from companionship without pretending two equivalent travellers are walking down the road.
Which brings us back to language we had already chosen before this inquiry.
Thinking companion.
Companionship without taking the Chair.
If those words are going to survive what we have now learned, they need a boundary.
What exactly can an AI companion do — and what must that word never be allowed to claim?
Sources followed
Costa, A.L. and Kallick, B. (1993) ‘Through the lens of a critical friend’, Educational Leadership, 51(2), pp. 49–51. Original article.
Titchen, A. (2003) ‘Critical companionship: Part 1’, Nursing Standard, 18(9), pp. 33–40. PubMed record.
Beyond the Breadcrumbs, Travel Companion.
Beyond the Breadcrumbs, A Conversation That Kept Moving.
Dated Human:AI project records consulted: Chief:DC Operating Orientation (9 September 2026); Human:AI Working Method (6 September 2026); Human:AI Collaborator Traits — Companion Overlay (6 September 2026); Human:AI Journey Reconstruction (13 September 2026).