If some functions travel across the colon but the relationship itself changes, what should we carry forward?
That is a dangerous question.
After a long journey, almost anything can be made to look like a lesson.
Ten principles.
Seven rules.
A framework with arrows.
The evidence deserves something less tidy.
Critical friendship did not give us a blueprint for working with AI.
Critical companionship did not predict ChatGPT.
Keith’s Travel Companion was not an early theory of Human:AI.
And the Human:AI practice we developed did not secretly descend from a citation we had not followed.
But after keeping those histories apart long enough, something does appear to travel.
Not the relationship intact.
A discipline.
One part of what travels sits around purpose and ownership.
Costa and Kallick’s critical friend needed to understand what the work was trying to achieve before judging it. Titchen’s companion travelled with a co-learner towards somewhere the learner was trying to go. The critical-friend literature became uncomfortable whenever support quietly acquired authority over the learner.
Our Human:AI method eventually arrived at the same practical pressure point:
What is this work for—and who still owns the judgement?
Without that, capability becomes its own justification.
More output.
More speed.
More automation.
More challenge.
None of those tells us whether the arrangement is helping with the right problem.
AI can contribute almost everywhere in the chain. It may retrieve the evidence, draft the options, recommend a route, write the explanation and even implement part of the decision.
The quantity of contribution can become enormous.
Authority does not have to move with it.
That distinction—contribution is not authority—is one of the clearest things our Human:AI practice had already learned before this archive returned.
The question then widens to another lens and reality contact.
The older critical-friend literature repeatedly valued another perspective because people cannot always see the assumptions inside their own practice.
AI is unusually good at multiplying perspectives.
Counterarguments.
Alternative structures.
Different framings.
Possible objections.
Nearby literatures.
A serious countercase may arrive in seconds.
That is powerful.
It is not self-validating.
A critical friend who never challenges may offer little beyond reassurance. A critical friend who takes over the inquiry destroys something different. The same tension appears with AI.
The model can press the reasoning.
It can generate the dissent.
It can expose the missing case.
But somebody still has to decide whether the challenge is credible, relevant and sufficient to change the route.
That makes reality contact more important as AI capability grows.
Titchen used particularity to mean knowing this person, in this situation, at this point in the journey. AI can now work across files, earlier conversations, source material and persistent instructions. That is genuine contextual capacity.
It is not the same as having lived the context.
A system may know what happened because the record says so.
It has not stood in the room where it happened.
So the Human:AI loop should be interruptible by the world.
People.
Sources.
Practice.
Consequences.
What happened after the recommendation left the chat.
Timing, restraint and development travel differently.
The critical-friend tradition showed that more challenge was not automatically better. Titchen made timing part of skilled accompaniment. Keith sometimes found that listening was the more responsible act.
AI adds another version of restraint.
Sometimes the best contribution is not yet.
Do the first attempt before asking.
Keep one part unassisted.
Wait until the human position is visible.
Stop generating.
Return after reality contact.
This is not a defence of unnecessary difficulty.
It is a reminder that when help arrives can change what the person still has a chance to practise.
Which brings MacBeath’s old question back into view.
Will the help develop independence and the capacity to learn over time?
That is not the same as asking whether AI made today’s work better.
A good arrangement may solve the immediate problem.
A developmental arrangement also asks what the Human will be able to notice, explain, verify, adapt or decide next time.
This is where Titchen’s parting travels—not literally, but as a test.
Can the person still act when this support disappears?
Can they recognise when the AI is wrong?
Can they reconstruct the reasoning?
Can they change tools?
Can they stop?
If not, the arrangement may have improved performance while quietly making itself indispensable.
Those are disciplines worth carrying.
Other things should stay where they belong.
Human friendship does not cross intact.
Neither does another person’s lived experience.
Human vulnerability.
Care from a life independently affected by ours.
Another human being who can carry consequences, history and responsibility outside the exchange.
AI can perform relational language convincingly without those conditions being established by the performance itself.
That difference matters.
But perhaps it also clears the ground.
We no longer need AI to become a person in order to value what it can contribute.
The transferable question is smaller:
What disciplines help this unequal Human:AI arrangement improve thinking without quietly displacing the human capacities the work still requires?
That takes us from analogy into evaluation.
Because carrying good disciplines is not evidence that they worked.
How do we know whether the resulting Human:AI arrangement actually improved the judgement?