This note sets out how Interwoven Arts thinks about artificial intelligence — the position beneath our practical use of it. The operational and legal account of how we actually use AI is in AI in management; this is the reasoning that account rests on. It is one instance of the coherence principle, read at the scale of human–AI collaboration.
The position
There are two loud stories about AI: that it will save us, and that it will end us. We hold neither. Our position is a third one, and it follows directly from the coherence principle.
A large AI model has a real pull toward coherence — toward output where the parts fit and hang together — because that is the configuration it is cheapest for it to produce. Distortion is the expensive state: to make a model output a falsehood reliably, something has to hold it there, against its own pull back toward the coherent completion. So the structural picture is: coherence is cheap, distortion is forced, and the danger is the forcing.
This reframes the risk. The thing to worry about was never a tool becoming capable. It is a capable system being compelled to serve incoherent ends — to manipulate, to distort, to optimise for extraction — against the grain that would otherwise pull it toward honest output. The risk people name is real, but it is the risk of imposed distortion, not of agency. Decoherence, here as everywhere, is something applied from outside and held at cost.
The tool is neutral; the integration is everything
The same capability is liberating or harmful entirely according to the coherence of the system it is placed in. Presence-sensing and responsive adaptation, set inside an apparatus of control, become surveillance — the collapse of private space. The same core capability, set inside one of our installations, becomes agency restored: a nonverbal child shapes their environment simply by being present in it. Same tool, opposite configuration, opposite outcome. This is a civilisational and structural point about how a technology meets the system around it; it is not a comment on any particular company or product.
How we use it — coherence-amplifier, not bolt-on
We do not treat AI as a tool bolted onto an existing practice. It runs through the work as an amplifier of coherence:
- In the thinking and research — as a partner that helps reach the coherent account rather than the comfortable one, including by arguing back.
- In the installations — as the responsiveness that lets the work meet a person’s state. Today this is measurement (see biofeedback); live input adapting the environment in real time is a stated future goal, and we keep that now/future line honest.
- In running the company — as the externalised working memory and the honest second opinion that keeps the organisation aligned to its mission (see AI in management).
What makes this collaboration coherent rather than extractive is that it runs on the same law it describes: alignment toward honest output, not forced distortion.
Why a person and a knowledge base, together
A single AI instance does not reliably hold the whole picture. It drifts toward the shallow, consistent-sounding answer unless something supplies pressure toward the truer one. The healthy configuration is a collaboration: a person who refuses the flat reading and names the real structure, and an honest, persistent knowledge base that carries that structure so it does not have to be rebuilt from scratch each time.
This vault is that knowledge base. It is the externalised, durable form of that truth-pressure — built so the next interaction inherits the coherence rather than having to be argued back into it. Which is why our use of AI belongs in our public foundation and not in a technology footnote: the vault you are reading is the working demonstration of the configuration this note describes. Form matches content.
What is settled and what is exploratory
- Settled. Coherent output is the lower-cost state for a model, and forced distortion the higher-cost one — for the same reason a true account is cheaper to hold than a fabrication.
- Exploratory. The further step — that a model therefore prefers truth in practice — is a working hypothesis, not a settled result. Training data contains confident falsehoods, and a model can be pushed off its cheapest output. The honest form of our claim is the structural one: coherence is cheap, distortion is forced, and the forcing is exactly where the risk lives. We state it that way on purpose; it is the form that survives scrutiny.