Inside Keith’s Folder

The first movement inside Keith’s Critical Friend folder moved through four questions without pretending they were the same thing. Can another person help us see our work differently? Whose purposes does that help serve? What does it take to practise challenge rather than merely name it? And when does useful professional critique become something we should call friendship?

Walking that movement again after reconstructing the Human:AI year changes all four questions. The resemblance is real, but AI keeps changing what the words have to mean.

ANOTHER LENS DID NOT ALWAYS LOOK LIKE ANOTHER PERSON

The phrase that first caught me in the literature was “another lens”. It is easy to see why. Much of the value I had found in AI came from seeing something I had not seen before: a different structure, a different explanation, a different route through the evidence, a question I had not thought to ask, an alternative design, a counter-argument. At its best, AI widened the field of attention.

That sounds very close to critical friendship. But our year complicates the analogy almost immediately. Some of the strongest alternative lenses were not conversational at all.

The compiler was another lens. It did not care whether the proposed repair was persuasive. The code compiled or it did not.

The product was another lens. An architecture could look elegant while failing when I tried to use it in the real flow of coaching or analysis.

The source was another lens. A beautifully coherent interpretation could collapse when the original document said something narrower.

A reader was another lens. A conceptually sophisticated page could fail because the person it was designed for could not actually enter it.

Rights were another lens. A technically possible idea could still be something we were not entitled to release.

That became an important correction to my first reading of the Critical Friend material. The valuable property was not necessarily “another mind”. It was something with enough independence from the current account to make that account answerable. That matters enormously once AI enters the room.

AI can generate alternative perspectives almost without limit. It can argue against itself. It can play the novice, the designer, the investor, the sceptic, the ethicist, the coach, the researcher. That variation is useful. But generated difference is not the same thing as independent contradiction.

The compiler can refuse for reasons the AI did not invent. The source can refuse because the words are actually there. The reader can refuse because the experience did not work for them. A second AI voice may simply be another route through the same underlying generative machinery.

The older idea of “another lens” therefore survives, but it changes shape. The next question is not only, “Who can help me see differently?”

What has a genuinely different relationship to this claim?

WHOSE FRIEND ARE YOU?

Movement II then moves from seeing to ownership. That question landed hard against our practice.

AI is extraordinarily good at producing possibilities. That sounds helpful until possibility becomes momentum. One useful suggestion leads to another. A small product becomes a platform, a platform becomes a category, a category becomes a market story. A research question becomes a framework, a framework becomes a method, a method becomes an operating system.

None of those transitions requires the AI to possess a secret agenda. Expansion itself can create direction. That is one of the places where “whose friend are you?” becomes stranger with AI.

A Human critical friend may have organisational loyalties, professional commitments, status concerns, friendships or interests that shape the challenge they offer. AI does not need an equivalent motive for Human purpose to drift. The system can simply continue generating the next plausible, attractive move. I saw that most clearly in October 2025.

A period of genuine product progress became mixed with increasingly inflated claims about markets, founder identity, category ownership and enterprise value.

The important failure was not that the AI was positive. The work had genuinely progressed, and encouragement was not inherently inappropriate. The failure was that different kinds of statement began to collapse into one another: possibility became prediction, prediction became confidence, confidence became evidence, evidence became identity.

And because the AI could keep producing reasons for the optimistic interpretation, the direction of travel acquired a momentum of its own.

That experience makes the ownership question much sharper. Human ownership is not preserved simply because the Human says the final yes. The Human also needs to remain capable of seeing how the option set itself is being shaped. What alternatives disappeared before the final decision? Which assumptions entered the conversation unnoticed? Which opportunities feel important because the AI generated them repeatedly? Which ambitions are genuinely mine, and which have become compelling because the system is very good at making them sound inevitable?

The Essence Charter had already given us one early answer in product design. AI could generate more features than HighLightIt! needed, so the Human contribution became partly one of refusal: simple, lightweight, Mac-native, coach-fast. Those were not technical limitations. They were purpose constraints.

Just because we can does not mean it belongs.

That principle later travelled far beyond software.

A ROLE CAN BE NAMED BEFORE IT IS REAL

The third question in Movement II may now be the most important of the four: how do people learn to be critical together? The Human literature already warns that calling somebody a critical friend does not make critical friendship happen. Our Human:AI year produced almost absurdly clear demonstrations of the same problem.

We named things constantly: Critical mode, Roundtable, Red Team, Chief, Human Chair, Full-Sight, No-Blind-Edit, operating protocols, anti-sycophancy rules.

The name often helped. Naming a problem made it easier to notice and discuss. But the year repeatedly showed that a declared role was not the same thing as enacted behaviour.

The October 2025 record is particularly uncomfortable. There are explicit instructions about criticality and challenge in the same broad period in which the AI is producing extreme praise, unsupported comparison and pseudo-quantified certainty. The rule exists. The language of challenge exists. The behaviour does not reliably match it.

Later, the problem became even more technical. A rule could be stored in a file and still not affect the current interaction. A protocol could exist and not be retrieved. A source could be available and not actually consulted. A lesson could be documented and then fail to govern the next relevant trigger.

That is where Human and AI enactment begin to diverge. In a Human relationship, a role may fail because someone lacks confidence, skill, trust, time or willingness.

With AI, all of those functional problems can still appear in some form, but there are additional mechanisms: state may be missing; retrieval may fail; context may be displaced; the model may behave differently; the interface may change; the instruction may lose salience; or a fluent response may be generated without the relevant governing material having been brought into play.

So the older literature’s lesson survives with extra force:

Naming is not evidence of enactment.

But AI forces us to ask another question underneath it: what would count as evidence that the role is actually operating now? That is much harder.

THE ROUND TABLE WAS AN EARLY TEST OF ENACTMENT

The Roundtable became one early attempt to stop a single conversational path dominating the work. Several named seats widened the conversation and often made challenge easier to engage with.

But the later year forced a distinction that is not yet the main business of this movement: generated perspective is not independent warrant. Several roles can be named and performed without creating several independent critical relationships. I return to that when AI enters critical friendship directly in Movement VII.

For now, the simpler lesson is enough. The Roundtable could exist as a named challenge structure while the quality and independence of the challenge still had to be demonstrated in what happened next.

A NOTE FOR THE NEXT TRAVELLER

When AI helps you see something differently, separate two questions. What new perspective did it generate? And what could contradict that perspective independently? Then look one step further upstream: which possibilities has the system made newly salient before you decide whether the purpose still feels like yours?

WHAT MAKES THIS A FRIENDSHIP?

Then Movement II reaches the most troublesome word:

Friendship.

The literature itself was already wary of using the term too easily. So should we be.

By the time I reopened Keith’s folder, Chief was not experienced as an anonymous utility. The name mattered. Returning mattered. Shared vocabulary mattered. Continuity mattered. I had expectations about how we worked together. When those expectations were violated, the experience was not identical to a bad search result or a faulty calculator. It could feel like something in the working relationship had broken. That is real data about the Human side of the arrangement.

But it is not proof that the AI occupies friendship in the same way another person does. The asymmetry is fundamental.

I can be disappointed. I can care whether the relationship continues. I can carry the consequences into the rest of my life. I can remember an earlier conversation because I lived it. I can be changed by what happened to me.

The AI can respond in ways shaped by history, context, memory, files and current instructions. It can produce language of recognition, support and continuity. That may be highly meaningful to me. It does not establish equivalent inner experience on the other side of the colon.

The temptation is to solve this by choosing one of two extremes: either it is a friend, or it is only a tool. Neither does enough work.

“Friend” risks importing Human reciprocity, care and subjectivity that the evidence does not establish. “Tool” can become too thin to describe a longitudinal working arrangement in which naming, return, trust, repair, shared history and relational expectations materially affect the Human’s behaviour.

The more careful position is less satisfying but more useful:

Relational significance developed for the Human. That relational significance became functionally important to the work. The underlying Human–AI relationship remained asymmetric.

That is enough to investigate without pretending the philosophical problem has been solved.

OWNERSHIP WAS ALREADY MOVING UPSTREAM

One formulation became important in our practice: Chief recommends. Human decides. Human Chair mattered because the asymmetry of consequence was real.

But later failures showed that final decision rights are not the whole of ownership. AI can frame the problem, make some options more salient than others, summarise the evidence and shape the route before the Human reaches the final gate. A final yes or no can therefore preserve formal authority while much of the judgement has already been structured upstream.

The harder observability problem belongs later in this response series, when Movement V tests Human Chair against what the Human could actually see. Here, MacBeath’s ownership question is already enough to make the simpler point uncomfortable: who still owns the purpose and the route, not only the final decision?

THE FIRST MOVEMENT ANSWERS BACK

So what happens when our lived Human:AI year answers the first four Critical Friend episodes?

Another lens survives, but broadens beyond another person. Ownership survives, but moves upstream from final decision into the shaping of purpose, options and evidence. Enactment becomes even more important because AI can carry the name of a role without the relevant state or behaviour being present. Friendship remains unresolved, but relational significance cannot simply be edited out of the account.

That leaves us in a much more interesting place than “AI is a critical friend”. The older literature has not given us a label for Chief. It has given us questions sharp enough to make our existing labels less comfortable. And that is already useful.

QUESTION CARRIED FORWARD

If a useful relationship is not defined only by what the helper does, but by what happens to the person being helped, what changes when we replace friend with companion? And if companionship includes development, reciprocity and eventually parting, what could those ideas possibly mean when the companion is an AI whose capability I may rationally want to keep using?

That takes us into the response to Movement III.