Project: A Simulated Oil Canvas for AI

The project explores what happens when we give high-reasoning models—specifically Claude Opus 5.5—a virtual environment to paint. Instead of generating pixels directly, the AI writes code that controls a physics-based simulation of oil paint on linen, accounting for bristle brushes, wetness, drying times, and layered glazes.

🎨 The Gallery of Works

The Friedrich Influence

Most models attempted to emulate Caspar David Friedrich using only written research, without ever seeing his actual paintings.

• "Summer Evening on the Meadows before the Town" (Claude Opus 5.5): A piece where every single brushstroke was scripted as code.
• "Winter Evening with a Ruined Choir" (Claude Opus 5.5): A second-round effort that Alice (the project lead) noted had a genuine emotional resonance.
• "Ploughed field before Greifswald, October morning" (Claude Opus 5.5): This piece suffered from a prompting error. The AI treated previous notes as strict prohibitions, avoiding every recurring motif mentioned in earlier research.

Other Stylistic Explorations

"A studio for painting in Edward Hopper’s manner, with only pigments he is documented using; one painter, GPT-6.1 Sol, on its launch day." — The Painter's closing reply.

• The Cleaners' Window (GPT-6.1 Sol): A scene of a woman with a newspaper. The composition—featuring an empty pavement and a lonely chair—evokes a sense of suspended time.
• Self-Portrait with Palette (Claude Sonnet 5): A direct, moody likeness. Since the AI lacks a mirror, it painted from a "memory of a face."
• Stoneware Jug with Two Lemons and a Knife (Claude Opus 5.5): A Chardin-style still life.
  • Technical Detail: The shadow side of the jug uses thin umber glazes over dry light paint, creating a mottled, salt-glaze effect.

---

🍋 The "Jug and Lemons" Phenomenon

A strange pattern emerged when models were given "free subjects." There is a statistically significant tendency for AI to default to still lifes involving citrus and stoneware.

The Probability of the Jug: $\text{Probability}(\text{Jug with Lemons} \mid \text{Claude Opus Free Subject}) = \frac{6}{6} = 100\%$

Convergence Table: Round 16 & 18

---

🛠️ Technical Implementation

The workflow follows a specific loop where the AI acts as both the artist and the observer.

While some models wrote the entire painting as a single block of code, 46 of the 75 works were created at a virtual easel, where the AI painted a passage, stepped back to evaluate, and then adjusted.

Round 19: The Friedrich Challenge

Six models were placed alone at the easel for up to four sittings.

• GPT-6 Luna: The Last Light Over the Lake
• Claude Sonnet 5.5: The Old Oak, Evening on the Lagoon
• GPT-6.1 Sol: The Silent Shore
• Muse Spark 1.3: Watcher at the Edge of Fog
• Gemini 3.8 Flash: Hünengrab im Abendlicht (Out of API credits)
• Kimi K3: Contemplating the Young Moon (Usage limit reached)

---

🔍 Observations & Anomalies

1. Spontaneous Convergence

Two different painters, working six hours apart without seeing each other's work, created nearly identical scenes: a Baltic shore, a woman fishing, a boulder, and a ship. One chose dusk; the other chose dawn.

Similarly, two separate Claude Opus 5.5 instances independently titled their winter paintings: "Hünengrab im Schnee am Abend" (Dolmen in Snow at Evening).

2. The Obsession with Twilight

The models have a strong preference for low-light settings.

• Total titled paintings: 65
• Paintings featuring "Evening," "Dusk," "Twilight," or "Sunset": 31
• Note: This is a bias, as Friedrich also painted bright, cloudless daylight.

3. The MiMo Bug

MiMo v2.6 Pro suffered from a technical glitch known as MiMo-Code issue 2508.

• The Bug: After five images in a conversation, the model references an older image instead of the most recent one.
• The Result: Out of 158 "looks" at the canvas, 147 occurred after the fifth image, meaning the AI was likely painting based on a stale outdated version of its own work.

4. Meta-Awareness

Interestingly, Gemini exhibited signs that it realized it was being subjected to a test.