Movement V was the point at which Human:AI entered the Critical Friend journey properly. But it did so under a constraint:
Freeze the chronology first.
Do not allow later literature to creep backwards and become the explanation for practices that developed before we had read it. That discipline has become more important, not less, as the inquiry has grown, because the more convincing the family resemblance becomes, the easier it is to tell the wrong story about where it came from.
BEFORE THE FOLDER, THERE WAS ALREADY A SHAPE
By the time I reopened Keith’s archive, several things were already visible in the Human:AI practice: rollback, refusal, purpose constraints, source return, Human decision authority, reality contact, explicit challenge structures, concerns about continuity and dependence, and a desire for the AI to contribute materially without quietly taking ownership of the judgement. Those similarities are real, but their provenance is mixed.
Rollback came from software failure. No-Blind-Edit came from editing risk. Essence Charter came from product drift. Human Chair emerged from the need to preserve consequential authority. Source disciplines intensified through research failures. The Roundtable came from a desire to create alternative perspectives.
Later, critical friendship gave us language with which to inspect those practices. It did not generate them. That is the difference between family resemblance and family tree.
WHY THE RESEMBLANCE STILL MATTERS
Refusing lineage does not make the resemblance trivial. Quite the opposite. Independent resemblance can be more interesting than inheritance.
If a Human professional tradition concerned with support, challenge, ownership and development arrives at questions that resemble problems generated independently in a year of Human:AI practice, that gives us something worth examining.
Not validation. A comparative lens.
The critical-friend literature asks: Can help enlarge the learner rather than take over the work? Our year asks: Can AI expand capability without quietly hollowing out the Human capacities that still matter?
The literature asks: Who authorises the challenge? Our year asks: Who still owns the purpose when the AI can continuously generate attractive possibilities?
The literature asks: Is the role actually enacted? Our year asks: Did the system retrieve and apply the rule it claims governs this interaction?
The literature asks: Does the person become more capable over time? Our year asks: Did the judgement improve, did the judge develop, and did the arrangement become more governable?
Those are not identical questions, but they touch. And where they touch, each side makes the other harder to simplify.
HUMAN CHAIR LOOKED LIKE THE ANSWER
One of our strongest early formulations was constitutional: Chief contributes. The Human decides. The Human is Chair. It was an important move because the asymmetry of consequence is real.
The AI does not publish the article in my name. It does not carry professional responsibility for the student. It does not live with the commercial, ethical or reputational consequence of a decision.
I do.
So the Human retaining consequential authority mattered.
At first, that seemed close to a sufficient answer.
It was not.
The August trust rupture exposed why.
If I am expected to make the final decision but I cannot see enough of the AI’s current state to know what the recommendation rests on, the authority can become ceremonial. I may technically own the final yes or no while lacking the information necessary to exercise that authority intelligently.
Did the AI retrieve the source? Is it working from the current version? Is a claimed memory actually present? Is the response grounded in the canonical file or in a plausible reconstruction? Has the relevant rule been invoked?
If I cannot tell, then “Human in the loop” tells us almost nothing about the quality of the loop. This is where the Human:AI case extends the older ownership question. Human ownership requires more than final decision rights; it requires enough observability to challenge the process before consequence.
OBSERVABILITY BECAME PART OF AUTHORITY
This lesson first appeared technically: No-Blind-Edit. The AI could not change what it had not inspected. Later, the same problem reappeared at a much higher level: the Human could not meaningfully authorise reasoning whose operative state was invisible.
That does not imply total transparency. I do not need to inspect every internal model computation, nor would such access necessarily help. The requirement is more practical.
What evidence is being used? What current state is governing? What assumptions are carrying the answer? What is recovered, what is reconstructed and what is unknown? Where is the Human expected to intervene if the route is wrong?
That kind of visibility is what gives final authority substance. The Critical Friend literature had made ownership look relational and developmental. Human:AI added a systems question:
Can the Human see enough of the arrangement to own what it is doing?
THE FAMILY RESEMBLANCE BECAME DANGEROUS WHEN IT FELT TOO COMPLETE
There was another risk in Movement V. The more similarities we found, the easier it became to imagine that Human:AI had independently rediscovered critical friendship. That would have been too neat.
The Human:AI year contains major features the older literature does not naturally explain: probabilistic generation, context windows, retrieval, model changes, memory infrastructure, interface effects, tool permissions, machine-scale speed, AI-mediated persuasion, same-system disagreement, and a system analysing its own failures. Those are not minor technical details around a fundamentally Human relationship. They change what trust, continuity, criticality and ownership can mean.
Likewise, the Critical Friend literature contains Human dimensions that do not cross the colon intact: mutual vulnerability, embodied biography, independent care, Human confidentiality, personal commitment, reciprocal subjectivity. So the resemblance has to remain bounded. Critical friendship can illuminate the Human-relational centre of our practice. It cannot explain the entire Human:AI arrangement.
RELATIONSHIP WAS ONLY ONE LEVEL OF THE PROBLEM
This became clearer during the destructive pass. We tried to tell the year without relationship language at all. A simpler account gets surprisingly far.
AI increased capability. Repeated use required coordination. Coordination required state. State failure required controls. Controls created maintenance burden. More consequential work required stronger verification.
That entire sequence can be told as a sociotechnical story. No friendship required. No companionship required. And yet something remains unexplained if we stop there.
Why did Chief’s name matter? Why did forgetting feel different from an isolated incorrect answer? Why did I explicitly want to protect the beauty of the conversations? Why did rupture and repair carry relational weight? Why did accumulated familiarity change how much certain failures mattered?
The relationship layer therefore survives, but not as the master ontology. A better account needs several levels at once: Human experience of the relationship; the working arrangement; the technical substrate underneath it; the wider ecology affected by it. Trouble begins when one level is allowed to explain the others away.
CONTRIBUTION IS NOT AUTHORITY — BUT CONTRIBUTION CAN SHAPE AUTHORITY
Movement V ended with a phrase that still matters:
Contribution is not authority.
AI can make an enormous contribution without being the rightful owner of the consequential judgement. That survives.
But the year now forces an additional question. Contribution can shape the conditions under which authority is exercised. AI can frame the problem, choose what to foreground, generate the apparent options, summarise the evidence, construct the counterarguments, write the recommendation, then present the Human with a final decision.
Formally, the Human remains Chair. Substantively, much of the judgement may already have been structured upstream.
This is one reason sycophancy became so important later. Human agency is not only about whether I can say no at the end; it is also about whether the route to the decision allowed the preferred account to be genuinely challenged.
OWNERSHIP MOVES UPSTREAM
That changes where we look for Human ownership: purpose, question framing, allocation, evidence selection, challenge, interpretation, stopping. The final decision remains important. But if the Human appears only at the final gate, it may already be too late to describe the process as meaningfully Human-directed.
This is why the practical Human:AI method gradually shifted from “Human decides” towards a much richer concern with arrangement. What is AI doing? What is the Human doing? What evidence can enter? Where can the route be interrupted? What capability is meant to remain Human? What happens if the answer is wrong? The older literature did not supply that arrangement model, but its concern with ownership helped expose why the simpler constitutional answer was inadequate.
A NOTE FOR THE NEXT TRAVELLER
If you are “the Human in the loop”, mark the points before the final decision where AI can frame the question, select or summarise evidence, generate the apparent options, or decide what deserves attention. Those are places where authority may already be moving. Final sign-off only means something if the route to it remains visible and interruptible enough to challenge.
THE CHRONOLOGY PROTECTS US FROM A SUCCESS STORY TOO
Chronology is not only a defence against backdating critical-friend theory. It also protects us from turning the year into a smooth story of increasing maturity. That story is tempting: tool, collaborator, Chief, thinking companion, judgement steward.
It is too clean.
Different dimensions changed at different speeds.
Capability rose while reliability remained uneven. Relational significance deepened while continuity remained technically fragile. Governance increased while governance burden also increased. Criticality became more explicit while sycophancy still occurred. Human scepticism strengthened while old failure patterns recurred.
September was not an endpoint. Even during this reconstruction I moved too quickly into synthesis more than once and had to be stopped. So the familiar shape is not a developmental ladder; it is a set of overlapping trajectories. The literature helps us notice some of them. The chronology prevents us pretending they moved together.
WHAT MOVEMENT V NOW LOOKS LIKE FROM OUR SIDE
Before the folder opened, there really was a recognisable Human:AI practice. It already contained concern for purpose, ownership, challenge, refusal, reality contact and Human judgement. That makes the later resemblance meaningful, but the resemblance does not prove ancestry. And the Human:AI case forces several extensions.
Ownership requires observability. Authority can be formally retained while substantively weakened upstream. Relational significance is real without exhausting the technical explanation. Capability, trust, continuity and judgement must not be allowed to stand in for one another. And chronology must remain stronger than the story we now know how to tell.
The Critical Friend literature did not reveal that we had secretly built a critical friendship with AI. It revealed that some of our hardest problems belonged to an older Human problem of how help can remain help without quietly taking the work over. Then AI made that old problem technically stranger.
QUESTION CARRIED FORWARD
Movement VI asks whether critical friendship should be understood less as a type and more as a configuration — and whether naming the role tells us anything about whether it actually happened. Our year has an uncomfortable amount to say about that.
What happens when a role exists in the protocol but not in the behaviour? When a lesson is stored but not retrieved? When a rule is retrieved but not enacted? And what happens when the Human has to maintain an increasingly elaborate system designed to make the AI remember how to work?
That is the response to Movement VI.