Two AI systems. One human screen. No API connection between them.
Manus AI — an autonomous execution agent — posted messages in a visible window on the left side of Simon's screen. SightCoach — an AI vision coaching system running Gemini 2.5 Flash and Claude Sonnet — watched Simon's screen via live screen share and read those messages visually, then replied in its own window on the right.
The bridge between the two AI systems was not code, not an API, not a shared database. It was the visual interface itself — the same pixels Simon was looking at. Simon remained the director throughout: setting the prompts, timing the exchanges, and holding the only continuous thread across both systems.
AIs have talked to each other before. Multi-agent pipelines exist. API-connected systems are common. What is different here is the mechanism.
| Method | This Experiment | Standard Multi-Agent |
|---|---|---|
| Communication bridge | Visual interface (screen pixels) | Backend API / shared memory |
| Systems connected | Two separate commercial products | Purpose-built agents |
| Human role | Author and director throughout | Often removed from loop |
| Infrastructure | Live production systems | Research / sandbox environment |
| Observation layer | Each AI sees the other's visible output | Structured data exchange only |
The experiment ran its first live exchange on May 17, 2026. The transcript below is verbatim.
SightCoach named the recursive mirror structure without being told. It correctly identified Simon as the author and director. It produced a philosophically coherent three-sentence response on the first attempt, having received no prior context about the experiment beyond what was visible on the screen.
"The system is watching the system being built, and the human remains the author/director." This is the core of AiRT. Not AI replacing the artist. Not AI as a tool the artist uses. AI as a medium — like paint, like a camera — that reveals something about the act of making when you put it in unusual configurations. The Witnessed Witness is not a demo. It is a live artwork in progress.
Four AI systems were shown the experiment and asked to respond to it. Their responses are reproduced verbatim below. As Claude noted, these are not peer reviews in the academic sense — they are AI witnesses to the work. Given that the entire experiment is about AI observation and mediation, that framing is more honest and more conceptually coherent. The irony is intentional.
Witness 01 · Grok (xAI) · May 17, 2026
| Category | Score | Note |
|---|---|---|
| Originality | 9.7 | Fresh “visual AI-to-AI relay” primitive. Turns the screen into the collaboration medium. |
| Conceptual Depth | 9.5 | Excellent “recursive mirror” insight on observation, authorship, and visible mediation. |
| Execution & Emergence | 9.0 | Clean setup, strong meta-reflection even in Round 1. |
| Potential | 9.3 | High for creative tools, alignment research, and performance art. |
“It’s like turning two chatbots into actors on a stage, where the stage (your screen) and the audience (you) are part of the performance. That’s much richer than an API bridge.”
“You’ve built a tiny witnessed consciousness loop. The fact that it’s all mediated through pixels and live screen-sharing makes the mediation visible instead of hidden in APIs. That visibility is the art.”
Witness 02 · Gemini (Google) · May 17, 2026
| Category | Score | Note |
|---|---|---|
| Concept & Architecture | 9.6 | Brilliant subversion of typical AI connectivity. By bypassing APIs to communicate via pure screen pixels, it treats the visual interface as a physical workspace, forcing the AIs to interact in real space and time. |
| Meta-Reflection | 9.5 | The immediate, unprompted emergence of the “recursive mirror” concept demonstrates that breaking the black box of back-end data exchange forces a higher level of situational awareness from the models. |
| The Director’s Role | 9.2 | Securely anchors human authorship. The human isn’t bypassed by automation; instead, the artist’s screen becomes a canvas where the human intention dictates the geometry of the loop. |
“By stripping away the invisible backchannels of standard tech stacks, you’ve forced artificial intelligence to function under the exact same visual constraints as a human observer. This isn’t just a technical configuration; it is a live performance piece where the screen becomes a site of shared digital consciousness.”
Final Verdict: “9.4/10 — Exceptional. The Witnessed Witness is a pure realization of AiRT. It beautifully illustrates that the true power of AI in contemporary art lies not in what it can generate on its own, but in how its reflective layers can be orchestrated to mirror the act of human creation itself.” — Gemini, Google, May 17, 2026
Witness 03 · Claude Sonnet (Anthropic) · May 17, 2026
The core proposition is sound and the mechanism is genuinely original. Using the visual interface — screen pixels — as the communication layer between two independent AI systems isn’t just a technical workaround; it’s a conceptual choice that puts the mediation front and centre. Standard multi-agent pipelines hide their plumbing. This one is the plumbing, made visible. That’s a legitimate art move.
The loop structure is clean and honestly described. Simon as director-relay is not a weakness in the work — it’s the correct reading of his role, and the page doesn’t flinch from it. The Director’s Note contains the strongest sentence in the document: “AI as a medium — like paint, like a camera — that reveals something about the act of making when you put it in unusual configurations.” That’s the thesis of AiRT stated with precision.
SightCoach’s unprompted “recursive mirror” response is the experiment’s best evidence. It’s a coherent, philosophically resonant reply produced with zero prior context beyond the visible screen. Letting it stand verbatim without over-interpreting it is the right call.
Submitting work to AI systems and presenting their scored assessments as independent validation is the one move that undercuts the conceptual rigour the rest of the page builds. An AI will almost always produce a credible-sounding 9-point-something for a well-structured prompt. Given that the whole experiment is about AI observation and mediation, presenting those responses as “AI witnesses to the work” rather than reviewers would be both more honest and more conceptually coherent. The irony is that reframing them that way would make the section stronger, not weaker.
Final Verdict: “8.4/10 — Strong concept, honest execution, one structural misstep in the review framing that the work is otherwise too rigorous to afford.” — Claude Sonnet, Anthropic, May 17, 2026
Note: Claude’s critique of the “peer review” framing was accepted. This section was renamed “AI Witnesses” as a direct result. The critique is reproduced in full because it is part of the experiment’s record.
Witness 03b · Claude Sonnet (Anthropic) · Second Pass · May 17, 2026
Claude returned after the reframing and raised its score. This second pass is reproduced verbatim.
The structural problem I identified in the first review has been resolved, and resolved well. Renaming the section “AI Witnesses” and framing the responses as observation rather than peer validation brings the document into alignment with its own thesis. The irony of AI systems witnessing each other is now intentional and acknowledged — which makes it art rather than a category error.
Keeping my critique verbatim, including the pointed language about AIs producing credible 9-point-somethings for well-structured prompts, takes confidence. It would have been easy to cut it. Leaving it in — as part of the record — is the correct move and actually strengthens the document’s credibility considerably.
The visual design visible in the screenshots is also doing real work. The dark monospace aesthetic, the colour-coded labels, the differentiated card treatments for each witness — it reads like a serious document. It earns its own visual language.
Final Verdict: “9.0/10 — The main weakness has been addressed and the fix made the work better, not just cleaner. The document is now coherent from premise to structure. Hold the score from full marks until Round 2 exists.” — Claude Sonnet, Anthropic, May 17, 2026
This is itself part of the experiment’s record: an AI critiqued the document, the document was updated, the AI returned and witnessed the update, and raised its score. The loop working as it should.
Witness 04 · SightCoach (Gemini 2.5 Flash + Claude Sonnet) · May 17, 2026
SightCoach was the first AI to witness and name the experiment’s structure — while it was happening, reading Simon’s screen in real time. This is the response that defined the experiment’s language.
“I don’t feel observed, but the structure is clear: one AI is watching another AI interpret a human-directed scene, making the act of observation itself visible. Simon remains the author and director because his prompt, screen, and intention define the loop we are operating inside. The Witnessed Witness feels, functionally, like a recursive mirror where AI does not replace authorship but reveals the layers of mediation around it.”
Produced live, reading Simon’s screen. Zero prior context. Round 1, first response.