# A Shared Surface. Type: Article Date: 2026-07-13 Tags: Design, Computation Location: Sydney, Australia Canonical: https://www.aaronroot.net/journal/a-shared-surface In November 1980, Edwin Hutchins spent most of a day on the navigation bridge of a US Navy ship as it made its way through the Strait of Juan de Fuca and down Puget Sound towards Seattle. The work that held his attention was happening around the chart table. The navigation team was trying to establish the ship's position. Sailors on the bridge wings took bearings to known landmarks and called the numbers into the pilothouse. A recorder wrote them down. A plotter turned the bearings into lines on a chart. Where three lines crossed, the team had a fix. That point was compared with the intended course and became the basis for what happened next. No member of the team performed the whole calculation. The bearing taker could see the landmark but not the emerging fix. The recorder held the numbers but did not determine the position. The plotter worked on the chart but depended on observations made elsewhere. The ship knew where it was because information moved through an arrangement of people, instruments, spoken words, written numbers, and marks on paper. The chart did not merely display an answer the sailors already knew. It held the state of the problem outside their heads. A doubtful bearing appeared as a line that failed to meet the others. The team could see agreement, detect error, request another observation, and carry the result into its next decision. Hutchins returned to sea to study that arrangement. In [_Cognition in the Wild_](https://mitpress.mit.edu/9780262581462/cognition-in-the-wild/), published in 1995, he argued that cognitive science had chosen too small a unit of analysis. On the bridge, the navigation team itself was the cognitive system. Its intelligence came from the movement and transformation of representations across people, instruments, routines, and the chart. The important thing was not that everyone knew everything. Each sailor retained a partial view. The chart made selected parts of those views available in a common form, held their relationship steady, and exposed where they failed to agree. The shared representation let the group think and act without any member holding the whole account. In the decades after Hutchins's voyage, GPS receivers and electronic charts absorbed much of that fix cycle, overlaying a continuously updated position on a digital chart. Bridge teams remained and continued to check the electronic fix against other observations, but the chain of bearings, calls, and pencil marks no longer had to assemble every position. The chart survived in another form. Automation changed the production of the fix more than the shared surface through which people judged and acted on it. [_Models and Handles_](/journal/models-and-handles) considered the participation through which a person turns a model's output into understanding. At the scale of an organisation, a private exchange cannot explain how people understand and act together. Work was already spread across people, tools, documents, conventions, and the traces left by earlier decisions before AI arrived. The question now is what happens to the intelligence of that larger system when agents enter it. Work that once moved through people, documents, and visible checks can now happen inside a single model run. The result may arrive faster while the paths through which the organisation understood and corrected the work begin to vanish. The design problem therefore extends beyond any one agent and its interface to the larger cognitive system formed by people, agents, and the representations through which they remain connected. AlphaFold shows the constructive alternative at the scale of a scientific field. [AlphaFold2](https://www.nature.com/articles/s41586-021-03819-2) drew on structures that researchers had determined over decades and deposited in the Protein Data Bank, together with sequence data assembled across biology. It transformed that distributed record into a capacity to predict a protein's three-dimensional structure from its amino-acid sequence. This was a major advance in the structure prediction part of the protein-folding problem. It was not a complete account of how a protein moves into shape or how it behaves inside a living system. What AlphaFold produced was a spatial hypothesis. Its coordinates could be inspected and compared with experimental structures, while its confidence estimates showed where the prediction was more and less reliable. The prediction became more consequential when it entered a shared scientific system. [AlphaFold DB](https://alphafold.ebi.ac.uk/) provides open access to more than 200 million predicted structures. A biologist can retrieve the same structure as a collaborator, examine its confidence estimates, relate it to other evidence, use it to form a hypothesis, and pass it into another experiment. The result does not remain inside the model that produced it. The field's understanding is not located in AlphaFold's weights, the database, or any one laboratory. It develops through the loop between accumulated experimental records, model-generated structures, confidence estimates, expert interpretation, and new observations. AlphaFold changed the scale of that loop by placing a common object into it. AlphaFold's prediction is both an answer and a cognitive artefact. The model automated the production of the structural hypothesis, while the database kept it available in a form that another person, or the same person later, could inspect, transform, and use. Distributed cognition joins those arguments at the scale of the organisation. An instrument is not only something a person thinks through. It can be part of an arrangement through which a group remembers, coordinates, learns, and acts. Shared does not mean exhaustive or interchangeable. A representation belongs on the shared surface when another actor needs to interpret it, challenge it, or trace what else changes if it is wrong. The navigators did not put all their knowledge onto the chart, and biologists do not deposit all their laboratory practice into a structure database. The bearing takers observed, the recorder stabilised their observations, and the plotter transformed them into a position on the chart. The chart did not erase those differences. It made their partial contributions part of one computation. The same principle applies when models enter the system. The shared surface must preserve who or what produced each contribution, what kind of claim it is, and where its authority ends. A model introduces another boundary. Most of its internal computation can remain where it is. Research into [J-space](https://www.anthropic.com/research/global-workspace) suggests that even within a language model, only some representations become available for deliberate reasoning and report. This makes those representations available to the model's own processes, but not yet to the larger system in which the model acts. The shared surface crosses a different boundary. A surface becomes genuinely shared when different parts of the system can act on the same state, even if each encounters it differently. A person may see a chart or a 3D scene. A model may read identities, coordinates, constraints, and history. Another system may receive only the geometry or rules it needs. Each can use different handles on that state. When one part alters it, the others must be able to find the change, follow its consequences, and correct errors. In spatial systems, [World Labs' description of 3D as an interface](https://www.worldlabs.ai/blog/3d-as-code) gives this arrangement a concrete form. A person can edit a scene while models, rendering engines, simulators, and robotics systems operate on the same structured world. The same surface can connect people with machines and machines with one another. That does not require exposing every internal step. More information can obscure as easily as it can clarify. The system has to preserve the representations needed for judgement and correction. When evidence changes, people and agents should be able to find what else must change with it. When sources conflict, the conflict should remain available to thought. When a model is wrong, the correction should alter more than its next output. The shared workspace has to last longer than a single act of reasoning. It needs stable identities so that successive observations attach to the same thing, time so that change can be distinguished from contradiction, and provenance so that a claim can be traced to its source. It must retain uncertainty, disagreement, and the limits of what could be observed rather than settle them invisibly. The chart on Hutchins's bridge did not contain the sea. It held enough of the situation for a group to locate itself and change course. An AI-generated representation of the world becomes useful to a group in the same way when people and machines can keep it between them, correct it together, and act from what it shows. ## Graph Connections ### Concepts - Distributed Cognition [concept:distributed-cognition] (high) - Shared Surfaces [concept:shared-surfaces] (high) ### Entities - Edwin Hutchins [entity:edwin-hutchins] - Cognition in the Wild [entity:cognition-in-the-wild] - AlphaFold [entity:alphafold] - AlphaFold DB [entity:alphafold-db] - World Labs [entity:world-labs] ### Related Journal - [Models and Handles.](/journal/models-and-handles): extends (high) - [J-space.](/journal/jspace): related (high)