Making the
Human Visible.

A path called authored use, and what it takes to make the promise of AI true rather than assumed.

By Alex Blumentals · Twin Ladder · 2026

1.The figure in the room

A figure I keep meeting in workplaces.

She still goes to the meetings, still files the reports, still gets paid. But when a real decision arrives — a hard call, a contested judgement, a moment when somebody ought to say this output is wrong and carry the consequence — she stutters. Not because she is weak. Because the conditions that form the standing to speak have been steadily denied to her.

This essay is about how to give those conditions back. It is about the practitioner who shows up — more visible, not less — because the work, the tools, and the standards around her have been deliberately built to develop her instead of replace her. I will call this authored use, and the rest of this piece is about what it takes to make it real.

The diagnosis is the easier work; the path forward is the harder one. We will get to the philosophy that names what is at stake — MacIntyre, Stiegler, Polanyi, the cartographer who cannot map — but the philosophy is here because it tells us what to build, not because it is the point. The point is the human becoming more visible. That is the destination.

2.The optimistic story, and what it costs to make it true

The promise everyone wants to make about AI is true: as the machines take on more routine cognitive work, the most valued human capabilities can become more visible, not less. Judgment, presence, taking a stand, the slow accretion of expertise into wisdom — these are the things only a person can bring, and they are the things that get harder to see when the volume of routine output is what gets counted.

I believe this story. I also believe it does not happen on its own.

The default deployment of AI in organisations makes the human less visible, not more. The metrics still reward throughput and speed. The AI optimises throughput faster than any person can. The practitioner who used to take five days to produce a careful judgement now takes five hours, ships the output, and is told her value is in the time saved. The thing she was actually trained to do — the careful judgement — has become invisible to the system that pays her, and what is invisible eventually goes away.

For the optimistic story to be true, four conditions have to be built deliberately. None of them are technological. All of them are organisational.

The measurement has to change. The organisation needs to measure judgement quality, internal goods, the development of the practitioner over years — not units shipped this quarter. If the metric does not change, the AI will optimise the old metric and the human will disappear into it.

The apprenticeship paths have to be protected. Juniors need to pass through stages that AI could shortcut, because skipping those stages produces no judgement. A practitioner who never had to do the work without the tool will never develop the standing to push back on the tool when the tool is wrong.

The authority has to stay with the practitioner. Her development, her data, her criterion of what counts as good — these belong to her. The twin reports to her, not the other way around. Authority does not migrate down the stack to the system that is supposed to support her.

The friction has to be preserved deliberately, in the right places. Simulations push back hard. Coaching stays human where relational depth matters. Some difficulty is the medium through which competence forms, and the systematic removal of all friction is a removal of the conditions that produce a practitioner. Not anti-machine, not anti-AI — anti-smoothing. The same tool can be deployed in a way that smooths away the formation of competence, or in a way that intensifies it. The question is which.

3.Why this isn't the calculator

The first serious objection has to be answered before anything else. Cognitive offloading is not new. Writing offloaded memory; Plato worried about it in the Phaedrus. The calculator offloaded arithmetic and most schools eventually adopted it. The printing press offloaded the labour of copying and the textual culture of late medieval Europe survived. Each was greeted with versions of the alarm now being raised about AI, and each turned out to be assimilable. What makes AI different — if anything?

The honest answer is not that AI is the first dangerous externalisation. The honest answer is that the previous externalisations stored artefacts, and this one externalises the act of judgement itself. Writing stored conclusions; the practitioner still had to form them. The calculator computed a known function; the practitioner still had to choose which function and interpret the result. The printing press copied a settled text; the practitioner still had to write and edit it. AI generates novel content whose criteria of evaluation are themselves part of what the practitioner is supposed to develop. It externalises not just the output but the standard by which the output is judged.

This is also why the complementarity literature is not the counter-evidence it is sometimes taken to be. Brynjolfsson's Turing Trap, Acemoglu and Restrepo's Power and Progress, Dell'Acqua on the "jagged frontier," Noy and Zhang on writing-task augmentation — these studies measure what happens with the tool, on bounded tasks, over short horizons. They show, correctly, that AI can complement and augment, particularly for novices. The offloading literature — Kosmyna's MIT EEG study, Lee's CHI 2025 survey of 319 knowledge workers, Gerlich's correlational study of 666 UK participants — measures something different. It measures what happens to the practitioner's capacity over time, and what persists when the tool is taken away.

Both can be true. The technology is not the question. The pattern of use across a working life is.

Brynjolfsson's point is in fact our ally. He warns against AI designed for imitation of human capacities and argues for AI designed for augmentation of them. That distinction is exactly the choice this essay is trying to make available. Authored use is the operational specification of what augmentation actually requires.

4.What's being lost when it goes wrong, named

It helps to have words for what the default path destroys, because without those words it becomes impossible to defend the alternative.

Three vocabularies converge on a single loss. Alasdair MacIntyre distinguishes internal goods — the developed eye, the calibrated hand, the years-long feel for a domain — from external goods like money or status. Internal goods cannot be transferred without being practised; they live in the body of the practitioner and in the relationships of the working community. The disease MacIntyre diagnoses is the systematic substitution of external goods for internal ones in measurement and legitimacy. The stuttering practitioner's predicament is this disease in lived form: she carries internal goods in abundance, in a measurement system that recognises only external ones, and is therefore illegible.

Bernard Stiegler tells the same story economically. The nineteenth century externalised savoir-faire — craft — into machinery. The twentieth externalised savoir-vivre — knowing how to live — into mass media. The twenty-first is externalising savoir-théoriser — the capacity for reflective thought — into AI systems. Each phase produces a worker who is biologically present and contractually employed and de-skilled at a deeper level. The endpoint is not a worker. It is a substrate.

Michael Polanyi gives the epistemic detail. There is a kind of knowing — tacit knowledge — that exceeds what can be articulated, and it is precisely this kind of knowing that the AI workflow most efficiently dissolves, because tacit knowledge cannot be made legible to the prompt-and-output frame. What the senior carries cannot be uploaded.

The combatant, in the long traditions of just war and the laws of arms, is recognised as a political subject who can refuse, who can surrender, who can hold ground against an order they judge wrong. The figure is borrowed here as a metaphor for political-deliberative agency, not as a claim about literal combat. What gets lost when the default path of AI deployment runs unchecked is what I have come to call the agentic death of the combatant: the slow hollowing of the deliberative subject who held standing to contest, refuse, push back.

These names matter not as decoration but as defence. Productivity language is not a defence; it is the medium through which the loss is administered. Risk language is not a defence; it merely manages the externalities of the loss for those who can afford liability. The defence has to be humanistic before it can be policy and design, because the criteria by which any policy or design choice is to be judged depend on what one believes is being lost and what one believes ought to be preserved.

5.The empirical signature is consistent with the philosophical claim

I am not making a metaphysical argument. The pattern is now visible in the empirical literature, in three independent studies with different designs, different populations, and different instruments — and they point in the same direction.

A study from the MIT Media Lab (Kosmyna et al., 2025) used EEG to measure neural connectivity, memory retention, and sense of authorship across three groups working with LLMs, search engines, and unaided writing. The LLM group showed weaker neural connectivity, lower memory retention, and a reduced sense of authorship over their own work, with effects observable in a follow-up subset after the tools were removed.

A study from Microsoft Research and Carnegie Mellon (Lee et al., 2025) surveyed 319 knowledge workers across 936 real tasks and found that higher self-reported confidence in generative AI was associated with less critical thinking, while higher confidence in one's own skills was associated with more.

A 2025 study in Societies by Michael Gerlich found a significant negative correlation between AI tool use and critical thinking across 666 UK participants, mediated by cognitive offloading and strongest in younger cohorts.

The three studies are not a settled body of evidence — one is a preprint, the sample sizes vary, the methods diverge — but they point in the same direction, and the direction is the one the philosophical framework predicted. What atrophies first is not output capability but the standing from which to evaluate output. The senior who can no longer reconstruct her own reasoning, the junior who never learned to form judgements without the tool, the team that ships faster but no longer remembers why it ships what it ships.

The engineering side of the empirical signature points the same way from a different angle. A 2026 Princeton paper by Rabanser, Kapoor, Narayanan and colleagues — Towards a Science of AI Agent Reliability — evaluated fourteen frontier models across eighteen months of releases on two agentic benchmarks. On GAIA (general agentic tasks) reliability improved at roughly half the rate of accuracy; on τ-bench (customer service) at roughly one-seventh the rate. The capability curve is racing ahead of the reliability curve, and the gap is widest in exactly the kind of work — customer-facing, judgement-laden, low-tolerance-for-recovery — where organisations are automating first. The practitioner whose standing has thinned is precisely the person an organisation needs to catch what the agent gets wrong, and the rate at which that catching is becoming necessary is not slowing.

This is the empirical signature of what philosophy has long been able to name. The cartographer atrophies; the combatant loses standing. Savoir-théoriser externalises into the model. Internal goods become invisible because the only legible signal is the speed and volume of output. The qualified life of cognitive practice — the slow accretion of judgement, the personal authorship of one's own conclusions, the standing from which to refuse a bad output — is being thinned at the same moment the technology around it is becoming less, not more, dependable in the work it is being asked to do.

6.Authored use, specified

Now the positive part. Before specifying the conditions, it helps to name what AI is actually going to do in organisations over the next three years, in plain terms — because the conditions only make sense as the answer to whether the fifth statement below becomes true or only aspirational.

1. AI will become a normal part of every employee's development journey. Personalised learning assistants will help people set goals, practise skills, receive feedback and track progress over time.
2. Coaching will become hybrid. Human coaches will focus on cultural nuance, ethical judgement and relational complexity, while AI will support preparation, reflection, practice and check-ins.
3. AI will help organisations analyse data on learning behaviour, change readiness, and impact.
4. Leaders will use AI simulations to rehearse conflict, performance conversations, crisis decisions and strategic dilemmas before applying them in real organisational contexts.
5. The most valued human capabilities will become more visible, not less. As AI takes on more routine cognitive work, it becomes time for the human to show up at work — bringing their uniqueness to work.

The first four are likely whatever an organisation does. They are the operating reality of the next three years. The fifth is the open question — it becomes true if, and only if, the first four are deployed under specific structural conditions. Those conditions are the rest of this section.

Authored use is the third position between AI refusal and AI embrace-without-conditions. It is not "no AI in this workflow." It is "AI in this way and not that way." The conditions can be named. There are seven, and each is testable.

1. The practitioner owns the criterion. What counts as good is decided by the working community of practice, not by the model's default or the prompt template's expected output.
   Can the practitioner say, in the specific case, what she would have judged as good independently of what the model produced?
2. Friction is preserved deliberately. The cognitive strain that develops judgement is not smoothed away in early-stage practice, even where the tool could shortcut it.
   In the apprenticeship phase, are juniors required to do the work without the tool before being allowed to use it?
3. Community before model. The practitioner is embedded in a community of practice with shared standards, not facing the model alone.
   Does a contested AI output get reviewed by a peer, or does it ship?
4. Apprenticeship paths are protected. Juniors pass through the stages that produce judgement even when AI could shortcut them.
   Is there a stage in the career path that an AI-augmented junior cannot skip?
5. Capacity-without-tool is monitored. The practitioner retains the ability to do the work without the tool, and that capacity is measured against atrophy.
   Are there periodic tool-free exercises, with results tracked over time?
6. Authored use is legible. What was thought, prompted, iterated, judged is documented in a way that lets a reviewer reconstruct the practitioner's contribution.
   When somebody dismisses "you used AI, so you didn't make it," can the practitioner answer with literacy about the specific authorship trail?
7. Internal goods are organisationally recognised. What the senior carries — the tacit eye, the calibrated judgement, the practice-community standing — is named, valued, and made measurable in ways that do not collapse it into productivity metrics.
   Does the organisation have a competence framework that names internal goods explicitly, or only output metrics?

These conditions are not a regulatory checklist. They are an organisational design proposition. The interesting work is in operationalising each — turning the test into a measurement, the measurement into a practice, the practice into a habit.

7.What this looks like in the learning-and-development domain

The seven conditions become five practices in the L&D domain — the surface where authored use is most tractable today, because the tools for personalised learning and rehearsal have matured fastest.

The personal learning twin. AI becomes a normal part of every employee's development journey. A personalised learning assistant helps the practitioner set goals, practise skills, receive feedback, and track progress across years rather than across a quarterly performance cycle. The developing version asks first: what would you say? what would you do? what do you see? — and only then provides the comparison. The replacement version offers shortcuts. The test is whether the practitioner becomes more capable without the tool over the period of using it with the tool. If yes, it is a learning twin. If no, it is a crutch.

Hybrid coaching. Human coaches remain — and should — but focus where they are irreplaceable: cultural nuance, ethical judgement, relational complexity, the slow trust that lets a coachee say the hard thing about themselves. The AI handles preparation, structured reflection, between-session practice, rehearsal, check-ins. The human hour of coaching gets its full weight back because the surrounding work has been absorbed by the twin. The coach owns the developmental arc. Authority does not migrate down the stack.

The data layer. AI analyses data on learning behaviour, change readiness, and impact across the organisation. Same data, two opposite uses: it can improve the conditions of practice — where are people getting stuck, where is feedback flowing, which teams are improving — or it can rank, sort, and discipline. The line between develops and replaces is whether the practitioner has authorship over her own data. Same technical setup. Entirely different politics.

Rehearsal for the real conversation. Leaders use AI simulations to rehearse conflict, performance conversations, crisis decisions, and strategic dilemmas before applying them in real organisational contexts. The simulation is the workshop floor on which the leader's hand learns. The hand learns from the hand by encountering resistance, Sennett said. The simulation must push back hard — find the weakness in the leader's framing, surface the discomfort she is unconsciously avoiding. A simulation that says "good job" produces no judgement. Difficulty calibrated, not removed.

The human becomes more visible. As AI takes on more routine cognitive work, the most valued human capabilities become more visible, not less. But the redesign of measurement is not optional. If the organisation continues to measure throughput, the AI optimises throughput and the practitioner disappears into it. If the organisation measures judgement quality, internal goods, the legibility of authored use, and the development of the practitioner over years, the AI becomes a partner in producing what the organisation now wants to see. The technology is the same in both cases. The doctrine is the entire difference.

8.The twin ladder

This is where the metaphor that names the company also names the philosophy. There are two ladders to be climbed in parallel: the practitioner's competence, and the AI's capability. The mistake of the default deployment is to climb only the second, faster and faster, while the first is taken away. The mistake of refusal is to refuse the second entirely, leaving the first to climb in conditions the rest of the world has already moved past.

The third position — the position this whole body of work has been trying to make available — is to climb both ladders together. AI capability rises; practitioner competence rises with it, structurally tied to it, each developing the other. The twin is not a substitute. The twin is a climbing partner. And, like any serious climbing partner, the twin's job is to teach the practitioner to do what the practitioner can do, not to do it for her.

9.What this means on Monday

The work begins with the questions an organisation has not yet had a vocabulary to ask. Are we deploying AI in a way that makes our people more capable or less, measured over years? Do our juniors have a stage in the career path that AI cannot let them skip? Does our coaching system concentrate human time on what only humans can do, or does it dilute it across what the twin should be doing? Does our data give the practitioner authorship over her own development, or does it give the organisation surveillance over her behaviour? Do our simulations push back hard enough to build judgement, or do they reward the practitioner for showing up?

These are not academic questions. Each has an answer in the practice of the organisation today, and each answer either builds the practitioner-with-a-twin or builds the stuttering figure. There is no neutral deployment. There is only the deployment one is choosing, by acting and by not acting. The choice is being made now, every quarter, in every contract signed and every workflow redesigned.

It begins with refusing the false binary. The choice is not AI or authorship. It is which kind of AI use produces which kind of practitioner. Authored use names one answer to that question. There are others. What there is not, anywhere, is a neutral default.

This is the work that Twin Ladder exists to support. An open standard for measuring AI competence at the workforce layer, mapped to the EU AI Act, designed so the seven conditions named above can be turned into measurements, the measurements into practices, the practices into habits. It is the apparatus that lets the optimistic story become operational rather than aspirational.

The human becoming more visible is the destination. The path is in our hands.

§References and reading

1. Acemoglu, Daron, and Pascual Restrepo. Recent work on automation, AI, and labour markets, 2019–2024.
2. Arendt, Hannah. The Human Condition. University of Chicago Press, 1958.
3. Ashby, W. Ross. An Introduction to Cybernetics. Chapman & Hall, 1956. The Law of Requisite Variety — a regulator must be at least as varied as the system it regulates — underwrites why the practitioner's competence must rise alongside AI capability rather than be replaced by it.
4. Brynjolfsson, Erik. "The Turing Trap: The Promise and Peril of Human-Like Artificial Intelligence." Daedalus, 2022.
5. Dell'Acqua, Fabrizio, et al. "Navigating the Jagged Technological Frontier." Harvard Business School working paper, 2023.
6. Gerlich, Michael. "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking." Societies, 2025.
7. Kosmyna, N., et al. "Your Brain on ChatGPT: Cognitive Cost of Using LLMs in Essay Writing." MIT Media Lab preprint, 2025.
8. Lee, H.-P., et al. "The Impact of Generative AI on Critical Thinking." Proceedings of CHI 2025, Microsoft Research / Carnegie Mellon.
9. MacIntyre, Alasdair. After Virtue: A Study in Moral Theory. University of Notre Dame Press, 1981.
10. Noy, Shakked, and Whitney Zhang. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science, 2023.
11. Plato. Phaedrus.
12. Polanyi, Michael. The Tacit Dimension. University of Chicago Press, 1966.
13. Rabanser, Stephan, Sayash Kapoor, Peter Kirgis, Kangheng Liu, Saiteja Utpala, and Arvind Narayanan. "Towards a Science of AI Agent Reliability." Princeton, arXiv:2602.16666, February 2026. The comparison of fourteen frontier models over eighteen months of releases from which the GAIA (~1:2) and τ-bench (~1:7) capability-to-reliability scaling ratios are drawn.
14. Sennett, Richard. The Craftsman. Yale University Press, 2008.
15. Stiegler, Bernard. For a New Critique of Political Economy. Polity Press, 2010.