The Gap Is the Work

Deutsche Fassung

Why precision is AI’s most boring superpower—and selection remains human work

We were promised the ultimate cognitive machine, finally setting the human spirit free. In many organisations, it now helps fill another management dashboard. The human spirit may join later, once the metrics have been approved.

I work as an artist with video projection, light and rooms that were never built for any of it and never filed the appropriate request. I also work in organisations introducing AI. To me, the two belong together. Both confront the same question: What happens when a tool capable of generating almost anything meets systems that would prefer to prescribe everything?

The machine can do a great deal. But on its own, it does not know what may be left out.

It can shorten, delete and summarise. Yet omission is more than less text. It is a decision about what matters, whom it matters to and which gap an audience can be trusted to complete. That unspectacular gap is where something begins that we have called art for centuries: not filling, but leaving room.

Two Empty Hands, an Entire Theatre

Rowan Atkinson stood on a stage and did almost nothing. Two empty hands. No prop, no cut, no effect. His face looked as if humanity had disappointed him personally. His hands reached for a drum that was not there, struck it—and the entire theatre heard every beat.

The trick was not that Atkinson delivered very little. He delivered a precise structure of tension, pause and absence. The audience had to enter the scene itself. It saw the drum because the drum was missing.

Jazz musicians often say that the notes you do not play matter as much as the ones you do. Atkinson played no note and electrified a room. The art did not supply the finished image. It marked the place where an image could form in somebody else’s mind.

The AI debate, meanwhile, likes to treat creativity as a volume problem: more data, more parameters, more variations before lunch. A thousand drum tracks in seconds. The buffet is full, nobody can taste anything, but everyone admires the logistics.

The bottleneck was never the equipment. It still is not. The bottleneck is understanding people—and knowing when, for once, not to bury them under output.

Precision Is Necessary. It Is Not Sufficient.

Organisations have good reasons not to treat AI like an open jam session. Decisions must be traceable. Data must be protected, rules followed and errors accounted for. Anyone preparing an official decision, checking medical information or supporting a personnel process does not need a charming detour. They need reliable foundations and a clearly named person who carries responsibility.

The problem begins when that legitimate logic becomes the only logic.

The machine is then expected to be more precise than the human, faster, cheaper—and above all predictable. Every output is standardised, every deviation becomes a defect, every pause an efficiency loss. A harness is pulled so tightly around the model that what emerges is a very expensive paperclip. The paperclip is celebrated in a meeting. Somebody says “game changer”. Somebody applauds. Somewhere, an idea dies, but the dashboard has no field for that.

The harness itself is not the problem. It is the constitution of the system. It determines where a model may vary, where sources are mandatory, who can stop the process and who ultimately carries responsibility. The cultural choice is therefore not control or freedom. It is this: Which forms of control protect people—and which suffocate the judgement on which responsible decisions continue to depend?

Where a false claim can cause harm, the corridor must be narrow. Where language, images, hypotheses or new perspectives are being explored, it may be wider. Force both into the same process template and you get either risky arbitrariness or correct interchangeability.

Productive misunderstanding needs boundaries. Within those boundaries, it also needs room.

Error Is Material, Not a Verdict

In my artistic practice, almost all the most beautiful moments began as mistakes.

A projector was tilted by one degree. The image ran across an edge, and the edge began to speak. A corrupted file swallowed a colour channel; the result was better than anything I had planned. It took me a week to reproduce the error deliberately. During a rehearsal, a projection failed, and in the darkness of the room it became clear that darkness was the protagonist. Nobody had thought to ask it.

Error is the only colleague who never does what is expected. That is why it can be indispensable. It is also why it still does not receive a permanent contract.

Not every glitch is a gift. Most defects are simply defects. They cost time, money and nerves; sometimes the only lesson is where the cable was lying. A creative mistake does not emerge from the defect alone. It emerges because somebody looks, selects, discards, keeps working and accepts responsibility for the result.

Selection is the work.

That is the connection to AI. A model does not automatically produce meaning simply because it produces something surprising. Variation is raw material. Judgement gives it form.

Hallucination Is Not Creativity

A language model is not a train running on rails. It calculates possibilities. Depending on the model, settings and task, it can produce highly expected or more unusual continuations. That space contains precise phrases, crooked images, useful detours—and false claims.

These things must not be confused.

In a factual context, a hallucination is not a creative spark. It is an error. Elegant phrasing does not redeem it. In a clearly bounded artistic experiment, however, the same generative openness may produce material that a person would not have reached alone. The untruth is not being celebrated. An unexpected form is being recognised, removed from its risk context and deliberately developed further.

To the enthusiasts, this means: More data does not automatically create more creativity. A model that has seen a great deal still does not have a point of view.

To the pessimists, it means: The machine is not a deterministic automaton. Its variance can be useful when we give it the right space and do not confuse its output with truth.

Variance without judgement is noise. Precision without openness is bureaucracy. The interesting work lies between them.

The Seventh Answer

I can create that in-between space quite practically. For an artistic task, I may take a small local model, vary its settings and give it an intentionally open instruction. I am not looking for the first correct answer. I read several attempts.

Sometimes the seventh version omits the very word I considered essential. The model has not “refused” anything; it has no artistic intention in this process. Yet the omission changes the rhythm. Suddenly, the sentence occupies the room differently. I keep it.

Something I could not commission emerges from the collision between my intention and the tool’s deviation. Not because the machine has secretly become an artist, but because I can treat its result as material, examine it and transform it.

Turned upside down, this is the Atkinson structure: AI is brilliant at filling space. My work is deciding which spaces it should leave empty—and which of its fillings are so wonderfully crooked that they deserve to remain.

Play-Doh for the Smaller Models

Small local models can be particularly interesting tools for artistic and exploratory work. Large commercial frontier systems are powerful, polished and optimised for very many people at once. They answer well. An open process, however, does not always need the best answer. Sometimes it needs friction, idiosyncrasy and a form that has not yet been sealed.

A small model on your own machine can be more like Play-Doh than a precision instrument. You can alter parameters, compare versions, see limitations and keep the process in your own workshop. You may use the tool incorrectly—knowing that you are doing so—and discover a new form precisely through that misuse.

This does not make local models automatically better, freer or innocent. They carry the traces of their training data, architecture and licences. They require hardware, energy, maintenance and competence. On complex tasks, they often lose decisively to large systems. Denying that is advertising, not analysis.

The gain lies elsewhere: in malleability and exit capacity. Can I understand what my tool is doing? Can I alter, replace and continue operating it when a provider changes prices, rules or access? Sovereignty does not begin with the label “local”. It begins with real capacity to act.

Freedom needs substrates: tools we can control, open knowledge, protected learning time, room to decide and people able to continue a process. No subscription to our own thinking—but no romance of doing everything alone either.

Who Owns the Time That Was Saved?

The efficiency story says that AI saves time. That is true. It is the least interesting of all possible truths.

The interesting question comes next: Who owns the time that was saved?

If every free hour is immediately filled with more tasks, nothing was saved. The same hamster wheel was merely accelerated. If the time remains with people—for thinking, researching, learning, experimenting and failing at something worth failing at—efficiency becomes a substrate of freedom.

This is a question of power, not wellness. Who sets the goals? Who defines quality? Who may reject a proposal even though it was fast and cheap? Who is allowed to take the detour? And who decides whether a saved hour disappears as a labour cost or returns as time for thought?

AI therefore changes more than how we work. It shifts what counts as work. Generation becomes cheap. Selection, context, dissent, responsibility and the ability to hold a gap open become more valuable—until it is clear what belongs there, or whether it should remain empty.

New Work does not mean decorating an office with creativity zones. It means giving people real command over tools, time and decisions.

Four Decisions Instead of a Culture Slide

Organisations that take this seriously can begin very concretely:

1. Separate the test track from the workshop. Factual and legally consequential processes need tight guardrails, sources and approvals. Exploratory work needs its own clearly marked space for variation and detours.
2. Name selection and stop rights. Every AI-supported process must make clear who may discard results, leave questions open and end the automation—and who is responsible for allowing it to continue.
3. Protect the time dividend. Part of the time gained must remain with people as learning, research and thinking time. Otherwise AI merely finances a faster pace.
4. Build exit capacity. Local competence, open formats and replaceable tools are not technological romanticism. They prevent organisations from renting out their capacity for judgement to a platform.

These are not soft questions about creativity. They are decisions about ownership, power and the organisation of judgement.

The Gap Is the Work

Precision is a superpower. It is simply the most boring one because it does not yet create meaning. It helps us examine whether a statement is true, a rule has been followed or a result reproduced. It does not tell us why something matters, which risk we should take or which thought would be better left unsaid.

That requires judgement.

Atkinson did not need a thousand variations. He needed two empty hands and an audience he trusted to think along. The machine can generate a thousand variations. Our task remains the old one: knowing which nine hundred and ninety-nine not to play.

The gap is not what remains after the output. It is what somebody has accepted responsibility for.

The gap is the work.

Deutsche Fassung