Field of view: what AI sees, and who decides.

We started with a white square on a photograph of the sky, an exercise by Bruno Moreschi at Aalto University. What could be hidden in it? First something probable, then something that is not. People reach for the improbable easily. A model trained on data reaches for the probable every time.

Machine learning works from examples that people have labelled: data, training, a model stored as numbers, and an output. Several works at Ars Electronica 2026 make this visible. ShadowGAN was trained only on mountains, so it sees mountains everywhere, even in the shape of your body. Memo Akten's Learning to See turns cables and cloth on a table into waves, fire and flowers.

A model's field of view is its training data. When the data is skewed, so is the machine. In Gender Shades (2018), Joy Buolamwini and Timnit Gebru found commercial face analysis erred 0.8% of the time for lighter-skinned men and 34.7% for darker-skinned women. The system was equally confident in both cases.

Behind every labelled example is a person. The World Bank estimates that between 154 and 435 million people do online gig work, and data labelling for AI is a growing part of it. In Bruno Moreschi's conversations with crowd workers, two of them put it plainly:

"We are teachers of computers." Anand, crowd worker, India

"There are no intelligent machines, but machines with intelligent humans." Sonia, crowd worker, Brazil

Knowing the data, the work and the labour a tool is built on belongs to using it well.

When we think about who has control, we tend to think in this kind of dimension: the more the machine does, the less the person decides. Ben Shneiderman's work on human-centred AI (2020) draws two separate axes instead. One measures how much the system automates. The other measures how much control the person keeps. Good tools score high on both.

Seen this way, the aim is to raise both dials together. A well-designed tool takes on the repetitive steps and keeps the person in charge of the choices that carry meaning: what to make, for whom, in whose name, and when to stop. That combination is what makes a system reliable, safe and trustworthy.

For arts organisations the two dials make the conversation concrete. Forecasting ticket sales, drafting a press release, answering visitor questions and coordinating a tour all sit in different places on the grid, and each one can be moved by design.

AI has moved from following rules, to learning patterns, to generating content, and now to taking action with tools. Each step hands more of the process to the machine. Using a grant application as the example, four levels of handover:

An agent works in a loop. It plans, uses tools, observes what happens and changes its next step based on what it finds. Traditional automation repeats the same steps. That adaptiveness is useful, and it is also why the points of human approval need to be designed in from the start.

More autonomy is not automatically better. The right level depends on the job, and the final decision should stay in your own field of view.

Image generators return what Hito Steyerl calls the mean image: the most probable version of the world, with everything at the edges of the field left out.

For artists and the people who manage their work, this matters twice. The probable becomes cheap, and the human mark becomes rare. Luxury houses are already investing in hand-drawn lines and visible brushstrokes. Epstein, Hertzmann and colleagues (2023) describe generative AI as a new medium with its own affordances, as photography once was. What an artist may gain is speed, range and a sparring partner. What may be lost, one delegated step at a time, is agency, a personal visual language, and authorship.

Control gets harder to see once personal agents act for us. You ask your agent for good paper for a large print. It searches, compares and is pitched to by sellers. It buys, and you never saw the persuasion. Gabriel and colleagues (2024) treat this as a central ethical question for advanced AI assistants: who shapes the agent's choices, and how would the person it serves ever know?

The students worked in groups on real scenarios from their field: a music agency using AI for press kits and booking, a gallery letting an agent run its social media and ads, a festival forecasting ticket sales, an orchestra discovering its recordings trained a music model, an artist using image generation to speed up commissions. Each group built one poster in three rounds.

1. Control map. List every step from first idea to audience and place it on the grid of human control and automation. Mark where AI is already in use.
2. Agency line. Draw a line through the map: which steps would you never hand over, and why? For each tool, what data goes in, and whose work was it trained on?
3. Agent charter. An agent now runs your promotion and buying. Who will try to influence it? Write three rules it must always follow to protect you and your audience.

There were no right answers. The value was in making each line visible and explaining it to the room.

1. Keep your own judgement.
2. Never trust the output 100%.
3. Know the data, the work and the labour the tool is built on.

The same exercise works in any organisation. Field One uses control maps and agency lines with teams adopting AI, so the decision about what to automate is made on purpose, step by step, with the people who do the work. Humane Technology by Design →