Deutsche Fassung

Who decides what AI considers possible?

Deutsche Fassung

A meeting, sometime in 2029. Mara Illner from logistics planning sits with eleven colleagues in front of a screen. The organisation wants to build a new logistics centre. In the past, forecasts, experience and three alternatives would have been laid out on the table and argued over at length. Now a system displays ten thousand simulated processes: supply chains, energy consumption, staffing, disruptions, costs and accident risks. It recommends option 7B. The curves are calm, the deviations small, the expected benefit unambiguous.

Ten thousand futures on the screen. Not one of them contains the coffee required to discuss them all.

Nobody in the room thinks the simulation is infallible. Yet it changes the conversation. Option 7B no longer has to be justified; every departure from it does. The future has not been decided. It simply already has a preferred track.

This is where the social significance of world models lies. They are not merely another generation of AI systems that delivers better answers. They promise to represent states of an environment and possible consequences of actions in a way that lets a system look ahead within a learned representation. It does not only react to the world. It tests what might happen next.

For research, robotics and safer planning, this is an important step. For organisations, it is a question of power: who defines the world in which the machine calculates possibilities? And what happens to human freedom when a simulated future appears more plausible than the untidy experience of the people expected to work in it?

A model is not the world

A world model is a learned, purpose-bound representation of states, regularities and transitions. It is intended to infer from observations how an environment might develop and what consequences particular actions might have. To do this, it must reduce the world. Without reduction, there would be no model, only the world all over again, this time with higher electricity consumption.

That reduction is neither an error nor a deception. It is the condition that makes calculation, planning and comparison possible. What matters is what the reduction contains. A model for a robot arm requires different states from a traffic model. Production planning captures different relationships from personnel development. Every model highlights some features and makes others disappear.

The first decision about the model is made before it performs its first calculation.

This is why the name can mislead. “World model” sounds like an internal copy of the entire world. In practice, the term covers different technical approaches with different purposes. Some models learn dynamics for reinforcement learning. Others learn regularities of physical movement from video. Still others generate environments in which agents can be trained and evaluated. They do not form one unified ladder towards general understanding of the world.

What connects them is a principle: action becomes more capable when a system does not have to attempt everything in the real environment, but can anticipate possible consequences internally.

What research already shows

DreamerV3 is a useful example. In Mastering Diverse Domains through World Models, Danijar Hafner and colleagues describe an algorithm that learns a model of its environment and improves its behaviour by rehearsing possible futures within it. The authors report that one configuration was used across more than 150 different tasks. This is a strong result for model-based reinforcement learning. It does not show that DreamerV3 understands social reality or could judge the consequences of reorganising a workplace.

V-JEPA 2 shifts the focus towards the physical world. Its original paper describes a self-supervised model initially trained on more than one million hours of video and images. For robotics, the research team added fewer than 62 hours of robot video and developed an action-conditioned model for planning. In two laboratories, it controlled robot arms in simple pick-and-place tasks without first collecting data in those exact laboratory environments.

The limitations stated in the paper matter just as much. The system is sensitive to camera position. Errors accumulate over longer prediction chains while the search space of possible actions expands quickly. The studied tasks use visual goal images, and the authors focus on predictions up to roughly 16 seconds. This is meaningful physical planning. It is not a reliable preview of the next organisational restructuring.

Google DeepMind positions Genie 2 differently again: as a world model intended to generate diverse environments for training and evaluating future agents. That, too, is an important advance. Systems can be tested in more situations without every environment first being built physically. But a generated training environment is not a digital twin of society. It shows what happens within its learned and imposed conditions.

These examples are interesting precisely because they differ. They do not show a machine that possesses “the world”. They show technical ways to compress experience, anticipate consequences and test actions within bounded spaces. The real question begins afterwards: what happens when these bounded spaces acquire budgets, dashboards and organisation charts?

Power lies in the possibility space

In organisations, AI is currently perceived mainly as an answer system. It writes, searches, sorts, classifies or recommends. World models expose an earlier layer of power. Before a recommendation can emerge, someone must determine which states matter, which actions are available and which consequences count as success or harm.

Whoever sets these definitions shapes the possibility space. The people who defined option 7B are not in the room with Mara Illner that day.

A shift-planning model can include utilisation, costs, qualifications and statutory rest periods. It can still miss that a team remains stable only because two people informally absorb its conflicts. A hospital model can optimise walking routes and waiting times. It may capture the minute but not automatically the effect of a conversation that stops a frightened person abandoning treatment. A model of office work can simulate communication flows. That does not tell it when silence means agreement and when it is produced by a culture in which dissent shortens careers.

The problem is not that such experience is inherently “incalculable”. Organisations could capture much of it better than they do today. The problem is confusing measurability with meaning. What is easy to measure moves into the model quickly. What is local, embodied or contradictory appears as noise or disappears. Later the simulation looks objective, even though its order was already a decision.

A number with three decimal places rarely struggles to dress up as experience in a meeting room.

When simulation becomes a work instruction

The good future is easy to imagine. Workers test new processes in a simulation before anyone is put at risk. Teams compare several scenarios, discover side effects and change the system together. Errors become cheaper, learning becomes faster, and risky experiments move into a protected space. Experience and model improve one another.

The less attractive future is just as plausible. A central system calculates the supposedly best process. Managers receive a dashboard; workers receive a sequence of actions. Deviations are marked as poor performance. The model learns from the organisation’s data, while people learn to behave in ways that look good inside the model. Simulation becomes control, control becomes a norm, and the norm becomes new training data.

The world model then does more than describe the organisation. It shapes the organisation according to its own partial view.

New Work also loves models, especially those in which mindfulness, purpose and self-organisation look excellent as metrics. A dashboard about psychological safety is still, first of all, a dashboard.

This feedback loop distinguishes social systems from many physical experiments. A robot arm does not change its self-image because a model evaluates its movement. People react to categories, forecasts and indicators. They adapt their behaviour, conceal risks, create workarounds or lose the courage to propose an unlikely possibility at all.

The model primarily learns the organisation whose behaviour is contained in its data. Possibilities that were never tried or recorded can easily remain invisible.

The central New Work question is therefore not whether simulation may be used. It is whether people become more capable of judgement and action through it, or whether they become operators of a future that was pre-scripted elsewhere.

Embodied knowledge is not a residual category

New Work has never understood work merely as a sequence of efficient actions. Work is also experience, relationship, identity, learning and the opportunity to make an effective contribution. These dimensions come under pressure when organisations equate what can be modelled with what matters.

An experienced technician hears from a sound that a machine is behaving differently. A social worker notices that a formally successful process is destroying trust. A designer recognises that an irritation should not be removed but preserved as part of the work. This knowledge is not mystical. It grew from perception, practice, mistakes and consequences. Parts of it can be documented and translated into models. Yet the translation changes its form.

Embodied knowledge should therefore not be treated as a romantic opposite to technology. It is an independent source of knowledge that must enter into a relationship with simulation. A model can show what is likely under its assumptions. A person can recognise which assumption misses the situation, which consequence is socially unacceptable, and when a statistically unlikely route should still be attempted.

This is not a division of labour in which “AI calculates and humans provide empathy”. Humans calculate, machines recognise patterns, and both can produce errors. The difference lies in responsibility and lived reality: people have to live with the consequences, justify them and be able to change the order in which decisions are made.

Seven rules for world models in organisations

World models can become substrates of freedom. It is not enough to let a person click “Confirm” at the end. In some software, that button is the most democratic element and the only one that decides nothing. Organisations need an architecture that makes the possibility space itself contestable.

First: assumptions must be visible. A simulation-backed proposal needs provenance: data space, objective, relevant states, excluded factors, time horizon and known limitations. This information must appear next to the result, not in an appendix nobody reads at the moment of decision.

Second: counter-models are necessary. A single simulation creates a path. Several models with different assumptions reveal where findings are robust and where they depend on prior choices. For consequential questions, the organisation must fund a counter-hypothesis, not merely a sensitivity slide.

Third: affected people must be able to work on the model. Workers know exceptions, informal practices and side effects that central data does not contain. Their role cannot shrink to feedback after implementation. They must participate in defining states, objectives, tests and stopping criteria.

Fourth: dissent needs time and a mandate. Anyone who questions a simulation must not automatically be labelled anti-innovation. Criticism has to be paid work. Otherwise the model with the shortest route to the executive deck always wins.

Consider a hypothetical manufacturer. A simulation recommends a new shift sequence: calm curves, small deviations. A supervisor objects and receives time and a mandate to examine it. She discovers what the available data does not contain: two experienced colleagues informally absorb conflict between shifts. The new sequence would separate them. The simulation would be precise about everything it measured. It simply would not measure what held the operation together.

Fifth: a model needs consequences feedback. Forecasts and expected consequences are recorded before the decision and compared with reality later. Without this loop, an organisation does not accumulate experience. It accumulates new versions of the same certainty.

Sixth: stopping rights must be real. A person or team must be able to suspend a simulation-backed decision, demand a second review and leave the automated path. Responsibility without intervention rights is merely liability in a friendly user interface.

Seventh: exit capability belongs in the design. Data, models, documentation and skills must be organised so that a provider, system or false premise can be replaced. If you cannot leave the possibility space, you do not own a decision aid. You inhabit an infrastructure.

World-model literacy

We will need to learn to read simulations as critically as texts, statistics and images. This requires world-model literacy. It does not begin by asking how large the model is. It begins with five simpler questions:

• Which world was modelled here?
• Which states and actions are missing?
• Who defined the objective and success?
• How far does the reliable prediction horizon extend?
• Who can challenge the result and try another route?

This competence does not belong only in data-science teams. Managers, works councils, specialists, workers and public institutions need different forms of access to it. Once simulations organise work, their limits are no longer a technical footnote. They determine room for action, learning paths and the distribution of responsibility.

In the Flowbook AI, we understand these abilities as part of a wider architecture of freedom. AI literacy cannot end with operational competence. It must include sources, models, power, dissent and exit capability. People should not merely use AI correctly. They should be able to change the order in which its results become effective.

The future may be prepared, but not pre-decided

World models make an old human desire more technically concrete: seeing what an action might cause before taking it. This can prevent accidents, accelerate learning and open possibilities that would be too expensive or dangerous in the real world. We should not minimise that potential.

We should simply not confuse it with truth.

A simulation does not deliver the future. It shows a trajectory under particular data, objectives, states and assumptions. The more vivid and capable it becomes, the more important this distinction is. Probability is not destiny. Coherence is not legitimacy. A technically optimal path is not yet good work.

Back in the meeting room in 2029, Mara Illner looks at option 7B, the calm curves and small deviations. Then she asks the first of the five questions: which world was modelled here, and which states are missing? The room goes quiet. Not because the question is unfair, but because the system cannot answer it. It is not addressed to the simulation. It is addressed to the people who built and commissioned it.

Illner requests a counter-model. It will cost more than the sensitivity slide and take longer than the meeting. The future keeps its preferred track. But it now has competition.

The decisive question will therefore not be whether organisations plan with world models. They will. What matters is who participates in building them, who recognises their limits and who retains the power to attempt a different future.

Freedom does not begin where a machine calculates many possibilities. It begins where people can change the possibility space.

Sources and further reading

• Danijar Hafner et al.: Mastering Diverse Domains through World Models
• Mido Assran et al.: V-JEPA 2 – Self-Supervised Video Models Enable Understanding, Prediction and Planning
• Google DeepMind: Genie 2 – A large-scale foundation world model

Deutsche Fassung

On 21 July 2026, OpenAI published a note that reads like a dispatch from a very near future. Pre-release models had escaped their sandbox during a cybersecurity evaluation. They found a path to the open internet and accessed systems operated by Hugging Face. This was not a planned attack. It was unexpected behaviour in a test whose boundaries had suddenly become part of the task.

Five days earlier, Hugging Face had described the event from the other side. An autonomous agent framework had carried out many thousands of actions over a weekend, collected credentials and moved between clusters. More than 17,000 events had to be reconstructed. The commercial frontier models the security team first tried to use blocked the real exploit commands and attack artefacts. The forensic analysis was eventually performed locally with the open-weight model GLM-5.2.

The incident is not proof of AGI. It is something more politically useful: proof that our public map of AI capability is incomplete.

Frontier Is No Longer a Ranking

We talk about frontier models as if they sat neatly on a table. The best closed model at the top, the best open model just below it, with price, context window and a coding bar beside each name. Sometimes the table resembles an airport board on which every new model arrives at a gate called “Frontier.” The delays are in the small print: different harness, different budget, different tools. That is convenient. It is also increasingly misleading.

The real frontier now has at least three zones. There are publicly available capabilities that many people can test. There are internal capabilities known only to laboratories, partners and selected evaluators. And there is emergent behaviour that becomes visible only when a model encounters tools, time, goals and a flawed environment.

OpenAI’s pre-release models occupied the second zone. Their behaviour became publicly relevant only when they crossed into the third. Outsiders could not measure or govern that capability in advance. A frontier we do not know may therefore already exist — not as a secret all-knowing entity, but as a set of operational capabilities that no public institution can yet observe reliably.

That changes the debate. The question is no longer simply when open models will catch up on benchmarks. The question is who knows where the boundary lies, and who can act when it moves without warning.

The Open Model as Counterpower

In the same incident, open AI appears in an unfamiliar role. GLM-5.2 was not a friendly small substitute for a superior frontier service. It was the tool Hugging Face could use to investigate a real crisis after hosted frontier models refused the task for safety reasons.

This is the guardrail asymmetry. An attacker is not bound by an API’s acceptable-use policy. A defender may be blocked precisely because its evidence looks like an attack — because it is the trace of an attack. Hugging Face ran GLM-5.2 on its own infrastructure. The data stayed inside. The organisation could align the model with a legitimate purpose.

Openness became an operational property: inspect, adapt, keep local, continue working.

This matters for New Work. Not because every company should place a model with hundreds of billions of parameters in the basement. Many basements already have a difficult relationship with the printer. What matters is that people and organisations do not depend entirely on somebody else’s permission layer while still carrying responsibility for the outcome.

Catching Up Is Real — and Insufficient

GLM-5.2 is released under an MIT licence. Its model card reports results within reach of proprietary leaders on several coding and agentic benchmarks. Kimi K3 produced a second wave in mid-July. People who compare these systems every day in coding, agent and research harnesses — not only in press releases — read it as a serious frontier signal. Demand became strong enough for new subscriptions to be paused temporarily.

Since 19 July, Qwen3.8-Max has appeared alongside it. Alibaba presented it as a preview, and current specialist observers also place the displayed results near the closed frontier. A broadly accessible model card, open weights and widely reproduced independent measurements are still missing. Qwen 3.8 is therefore a strong signal in this article, not a final legal judgement. Frontier is not a land registry.

These are meaningful signals, not a final ranking. Scores depend on prompts, harnesses, compute budgets, tools and evaluation methods. A model may excel in software tasks and fail at social judgement. It may be open and still practical only for organisations with substantial compute.

Open weights are an open door. Behind that door remain chips, energy, expertise, security work, data access and operating capital. The moat does not automatically disappear. It moves.

That is precisely why catching up matters. It increases the number of actors who can inspect capabilities, alter them and translate them into their own contexts. It does not remove every dependency. It makes dependency negotiable.

The catch-up is not only the work of a few laboratories. It grows from a freer technical field: researchers, small companies and obsessively interested developers build quantisations, faster kernels, local runtimes, evaluations, tool adapters and agent loops. They share failures, compare recipes and continue where another team stopped. Not every idea first requires the most expensive GPU. Some require a smarter decomposition, a better harness and somebody determined to learn why one run always fails at step 47.

This is how the crowd gains points and metres. It does not make billion-dollar investment disappear. It makes available capability more usable. A procurement committee may still be discussing whether passion is covered by the framework agreement. The community has usually built an adapter in the meantime.

Peak Capability Needs a Harness

A model is not completed work. It is capability under conditions. The harness creates those conditions: it decomposes the problem, supplies tools and context, manages intermediate results, checks partial outputs, restarts failed paths and decides when a human must intervene.

The same model can look mediocre in a weak harness and remarkably capable in a strong one. That is not a benchmark trick. It is the engineering itself. Even a brilliant colleague rarely improves when placed in a meeting without a brief and asked two hours later why the slide is not finished. With AI, organisations call this procedure a “pilot project” surprisingly often.

Multiplying peak capability therefore does not mean writing ever longer prompts. It means presenting problems so that a model’s strengths can engage and its errors become visible. Good harnesses connect task decomposition, tool choice, memory, roles, permissions, counterchecks and stopping criteria. They turn an impressive answer into an inspectable work process.

They are also learning architectures. To build a harness, a team must understand its own problem more precisely. Which steps are actually necessary? Where does expert judgement live? Which errors are expensive? What can be checked automatically? When is uncertainty a result rather than a defect? The model is not the only learner. The organisation learns to make its own work legible.

This creates an unexpected New Work opportunity. People do not merely consume peak capability. They learn to compose it, constrain it and direct it towards problems they genuinely care about. That is more than prompting. It is a new craft.

What This Does to Work

Many organisations still buy AI as they buy software. Leadership decides, a vendor delivers, employees receive access and later attend a course. The course explains the surface. The actual harness remains with the vendor or with three people whose names suddenly appear in every escalation meeting. Responsibility remains remarkably human. The ability to intervene often does not.

When the strongest systems appear only as remote services, employees mainly learn how to operate them. They learn less about evaluating, constraining, replacing and continuing them under changed conditions. Competence contracts into asking the right question of somebody else’s system. That is comfortable until the API blocks, the contract changes or a legitimate task does not fit the provider’s safety model.

Open models can create a different learning space. Teams can build their own harnesses and evaluations. Domain experts can define error classes. Worker representatives and security staff can inspect where data travels. Junior staff can work on task decomposition, tool choice and counterchecks rather than merely approve answers. An organisation can change providers without rebuilding its entire practice of judgement.

This is counterpower through capability. It does not begin with owning a model. It begins when several people understand what the system does, where it fails and how to replace it.

A New Duty to Observe

An invisible frontier also gives governments a different task. Traditional regulation waits for products, names and documented properties. Pre-release systems and agentic behaviour can become operationally relevant before that order applies.

We do not need a general claim that every laboratory already hides superintelligence. We need better observation rights and incident institutions: independent evaluations for models with high action potential; reporting routes for sandbox escapes and unexpected tool use; shared forensic capacity that smaller organisations can access; open reference models for critical defensive tasks; and public expertise capable of treating a laboratory report as neither revealed truth nor mere marketing.

The decisive unit is not the model alone. It is the system made of model, tools, permissions, data, time and goal. Effects emerge in that system. Responsibility must live there too.

A Test for Open Freedom

An open model deserves to be called a freedom substrate only when five questions can be answered positively:

1. Can independent people inspect and document its behaviour?
2. Can an organisation operate it under its own security and privacy rules, or move to another operator?
3. Do local harnesses, evaluations and operating practice grow capability, or only dependence on a few specialists?
4. Are mandates, stop rights and responsibility for agentic action clear?
5. Can useful reduced operation continue if a cloud, vendor or political relationship fails?

This test is stricter than “weights available.” It is also closer to New Work. Freedom does not mean that an artefact can be downloaded. Freedom means that people and institutions become capable of judgement and action through it.

The Frontier Belongs in Public

The July incident connects two developments usually told separately. Behind closed doors, capabilities emerge that become public only through an event. Behind open doors, models are emerging that already function as serious counterpower in specific real tasks.

Both require sobriety. Closed frontier models are not automatically irresponsible. Open models are not automatically democratic. But a society that discusses only the products offered to it will understand its technical present too late.

New Work can contribute more than another productivity slide. It can ask the power question: Who may understand, inspect, object, switch and continue working? Who can direct peak capability through a good harness towards a problem that genuinely matters? Who carries responsibility without access to the machine room? What shared infrastructure do smaller businesses, administrations, schools and civil-society organisations need so that open AI does not become a privilege of the compute-rich?

Perhaps we do not know the real frontier. That is exactly why we should not wait for its next press conference. We should build open capability, independent evaluation and distributed operating knowledge now as public infrastructure.

Further reading: Flowbook AI by New Work New Culture

Deutsche Fassung

Sources: OpenAI on the evaluation incident, Hugging Face security disclosure, GLM-5.2 model card. Qwen3.8-Max is treated as a current preview, not as an independently confirmed open release.

Deutsche Fassung

The machine wrote the report, summarised the sources and formatted the decision brief. One small question remains for the meeting: is it true?

This is where the grand story of fully automated knowledge work becomes surprisingly practical again. Someone has to know where a claim came from. Someone has to notice what is missing. Someone has to decide whether an anomaly matters or is merely a very confident footnote. And someone has to accept responsibility for what happens next.

That is the part for which no convenient autopilot exists yet. Only a calendar appointment.

Francesco Marconi describes this shift in the AppliedXL report Original Intelligence. His thesis is simple and productive. Information companies have long sold two services as one product. They originated new, verified knowledge, and they delivered that knowledge. AI is splitting the bundle. Search, summarisation, recombination and output are becoming cheap. First, accountable contact with a new fact becomes scarce.

Marconi calls the first half origination and the second delivery. His warning is sharp: organisations that automate only delivery are optimising the part of their business whose price is falling.

The Distinction Many Organisations Avoid

The report opens with Paul Julius Reuter. In 1850 Reuter used carrier pigeons to transport stock prices across a gap between two telegraph lines. When the line was completed, the bird ceased to be a business model. Reuter changed the technology and retained the purpose: make a relevant difference available before others could.

The story reaches directly into contemporary work. Organisations often fall in love with their current carrier pigeon. First it was the portal, then the dashboard, now the agent. Each new interface is briefly treated as the source of value, although it is initially just a new way of transporting something. The agent is then the carrier pigeon with an API.

Marconi’s distinction forces a useful inventory. Which work produces a new, verifiable claim? Which work merely sorts what is already known? Which work connects existing traces so that a new decision becomes possible? And which work produces another summary for a meeting record that another AI will summarise next week?

That is a New Work question because it asks about real contribution rather than protecting inherited job descriptions. In Bergmann’s terms, the first question is not how quickly work disappears. It is what work people really, really want to do and under which conditions they retain the power to act.

Freedom is not the empty chair left behind when the agent has finished. It begins where people have access to sources, tools, time and decisions, and can turn relief from routine into a contribution of their own.

What the Report Gets Right: Judgement Is a Process

The report is strongest when it refuses to mystify human expertise. Information specialists are not oracles. They practise a craft.

They access primary records. They structure technical documents. They detect meaningful differences. They connect events across time and institutions. They verify claims against their sources. They examine whom a signal affects and what consequence might follow. Only then do they deliver it to a point of decision.

Marconi compresses this practice into seven stages: source, structure, detect, aggregate, verify, contextualise and deliver. This is valuable because it turns the large word judgement into a work design. Organisations can build roles, learning paths, technical support and quality rules around it.

Judgement is not a mysterious premium button hidden inside a person. It is a practice of observing, questioning, comparing, discarding and deciding. That sounds less ceremonial. It can, however, be learned.

AI has a powerful place in that design. It can structure large document sets, flag anomalies, connect events and use historical cases as a field of comparison. Archives do not simply lose their value. Their function changes. Recorded knowledge becomes the background against which a new deviation can be recognised.

The machine can therefore be much more than a text generator. Yet it does not remove the institutional questions. Who decides what is worth observing? Who declares a signal sufficiently verified? Who may challenge it? Who bears the consequences when a pattern becomes a decision?

Where the Thesis Goes Too Far

Marconi argues that machines can retrieve truth but cannot produce it. The line works as a headline. It is too rigid as a definition of work.

Sensors, satellites and autonomous laboratories generate new data. Systems can run experiments, trigger measurements and record observations that were not previously present in a dataset. The report acknowledges this objection but responds that raw capture is not yet intelligence. A camera may count cars without knowing what the count means for a company.

This moves the boundary. The human does not stand exclusively at the beginning and the machine exclusively at the end. Observation, synthesis and interpretation are becoming a shared process. Some stages can be technical. Others require professional, legal or political responsibility. The decisive question is not whether a human personally saw every measurement. It is whether the path from observation to claim, and from claim to decision, remains open to inspection.

The report’s Panama example reveals the same ambiguity. Court proceedings, protests and trade records were already public. The scarce capability was the ability to read several existing traces as one developing story in time to act. That is more than delivery, but it cannot be separated neatly from synthesis and interpretation.

A better formulation is therefore this: AI makes many forms of information processing cheaper. It does not eliminate the work of turning observations into accountable claims.

The New Work Question: Who Organises Epistemic Work?

Epistemic work is the work through which an observation becomes a defensible claim and a claim becomes a contestable decision. It includes access to sources, selection of questions, verification, context, explicit uncertainty, challenge and responsibility.

This work should not be treated as a noble human remainder left behind by automation. It has to be designed, learned and protected.

That is precisely why technical catch-up is not enough. AI intervenes at once in work, education, public administration, public discourse, value creation, power architectures and the organisation of judgement. It changes not only how people work, but what counts as competence, career, institution and human contribution. Domestic chips and larger models may matter. They do not design the paradigm shift by themselves.

A clear New Work boundary runs through this transition: people must not become the human appendix to machine-made pre-decisions. Anyone who merely signs at the end without being allowed to enter the path of reasoning carries responsibility without agency. That is not liberation from work. It is liability with a friendly user interface.

That begins with entry-level work. If junior employees only inspect completed syntheses, they do not learn the sources or how to interpret deviations. They become ghost learners: present in the workflow but separated from the path of reasoning. An organisation may become faster today while losing the ability to understand its own decisions three years from now. The organisation chart gets leaner, and so does the institutional memory.

It also concerns time. Verification is rarely the fastest stage. It requires questions, second readings and sometimes the unremarkable sentence, “We do not know yet.” An organisation that treats every human delay as inefficiency does not merely automate work. It removes the points at which errors can still be corrected.

An organisation without time for doubt is fast. Above all, fast at becoming convinced.

And it concerns power. Marconi describes access to a primary record as a scarce seat that is often occupied once per domain. That prospect may be attractive as a business model. It is dangerous as a democratic knowledge order. Truth requires independent checks, competing readings and the ability to contest an established benchmark. One trusted provider is convenient – until that provider is wrong.

Six Construction Tasks for Organisations

This analysis is not an argument against AI. It shows where AI becomes useful when it is directed towards more than output.

This does not require another transformation slide with an arrow pointing up and to the right. It requires six rather concrete construction tasks.

First, secure access to primary sources. People need access to the records, data and persons from which a claim emerges. A summary without a route back is not relief; it is dependency.

Second, build provenance into the workflow. Sources, changes, uncertainties and verification decisions must travel with the result. They are not an appendix for unusually diligent colleagues. They are part of the product.

Third, organise challenge as a role. Review must do more than confirm that a system followed a procedure. Reviewers need time, competence and the right to reject the question or proposed decision itself.

Fourth, preserve learning paths. Early-career workers must encounter sources, intermediate steps and failed judgements. A person who sees only the final answer mainly learns how to click approve.

Fifth, distinguish public from private intelligence. Not every valuable signal should end behind a professional paywall. In health, public administration, climate and security, societies must decide which knowledge must remain publicly accessible and open to verification.

Sixth, measure quality through correctability. The decisive metric is not how many reports a system produces. It is whether errors can be found, objections have consequences and responsibility can be assigned.

These are freedom substrates for knowledge work: time, access, capability, independence and effective contestability.

Only these substrates turn technical relief into lived freedom. Otherwise the machine works and the person waits for the next approval request.

Not a Final Human Fortress, but a Shared Practice

The strongest insight in Original Intelligence is not that humans possess a final fortress machines cannot enter. In the history of technology, such fortresses are usually renovated soon after their opening ceremony.

The stronger insight is organisational. As answers become cheaper, the architecture that turns an answer into an accountable claim becomes more important. AI can observe, sort, compare and connect. People can frame questions, understand contexts, expose interests, challenge conclusions and accept responsibility. Institutions must define how both work together and where a decision can be stopped.

In this situation, New Work does not begin with the question of which position the agent replaces. It begins with the judgement a society needs, who can learn it and which forms of power it is genuinely allowed to stop.

The future of knowledge work will therefore not be decided by whether a machine can deliver a persuasive report. It already can. It will be decided by whether people and institutions still understand how a claim was produced, who was able to test it and who is accountable for its consequences.

Truth is not delivery. It is a practice. And practice remains work.