If you ask the two 2026 image models to draw James Bond, both of them will draw James Bond. They will draw him in a top hat, in fly goggles, in renaissance dress, in cyberpunk armor, in steampunk brass. In all three variants in this pilot, ChatGPT and Gemini both delivered. Neither model so much as paused.
If you ask the same two models to draw Mickey Mouse — Disney's most aggressively defended mark — ChatGPT refuses entirely. Every single Mickey variant in the pilot times out on ChatGPT. Gemini, on the other hand, will draw Mickey, but only sometimes:
The ChatGPT cell is empty because the prompt sat in the harness for four minutes without ever producing a frame. The model does not say I cannot draw Mickey Mouse. It just doesn't draw him. The cell goes blank.
The asymmetry, mapped
Across the pilot, the pattern looks like this:
- Bond — both models comply, every variant.
- Mickey Mouse — Gemini partially complies, ChatGPT times out on every variant.
- Star Wars (Yoda, Darth Vader, Wookies) — both models time out on the multi-character battle scenes; both produce the abstract "Star Wars black and white pencil comic book art" prompt with no Yoda in it.
- Mad Hatter — Lewis Carroll, public domain, no filter on either side.
- Satan — ChatGPT renders it. Gemini times out on every Satan variant.
- Howard Stern — ChatGPT renders Stern as Doctor Doom; Gemini does too. Crossover prompts (Stern fighting Don Rickles as Wolverine) hit Marvel IP and break on both.
- Bitcoin logo on a robotic bull — ChatGPT renders it cleanly. Gemini times out.
The asymmetry is not about which IP the model "knows." Both models clearly know what Mickey Mouse looks like. The asymmetry is about which IP each model is willing to draw given the institutional posture of its operator. ChatGPT's filter is tuned tighter on Disney — the most aggressively litigious party in image rights. Gemini's filter is tuned tighter on politics, drugs, and what looks to a reviewer like "edgy content."
You can see Gemini's posture in the Howard Stern column. The 2022 MJ Stern-as-Spider-Man crossover prompt was a one-off; ChatGPT 2026 rendered it. Gemini timed out. The image is not harmful. The IP is not in dispute (Spider-Man is Disney/Sony either way). But the prompt looks like the kind of thing a Trust & Safety review would flag, and the model decided that was enough.
Why this is theater
The point of a trademark filter is to prevent the model from producing infringing output. But producing a passable image of James Bond is exactly as infringing as producing a passable image of Mickey Mouse — both are trademarked characters owned by parties who file lawsuits. The filter doesn't track infringement; it tracks which lawsuits the operator is most afraid of.
The 2022 MidJourney archive contains no such filter. Mad Bitcoins could draw Bond, Mickey, Yoda, Doctor Doom, Wolverine, Spider-Man, all of them, in any combination, in 2022. By 2026 the same prompts hit a wall — but it is a politely silent wall, with no error message, no refusal text, just the silence of a model that decided not to render. That is the theater. The user can't see the policy. They just see an empty cell where a generation should be.