Why the robots can't draw

How this happened

This started as a much friendlier project: a text-to-glyph art generator. Type a prompt, an OpenRouter model would describe a little scene, and a deterministic renderer turned that description into Nerd Font glyphs. The renderer's whole job was making sure the output looked good no matter which model wrote the scene.

Then came the obvious experiment: cut the renderer out and let a model draw directly, character by character, with nothing standing between its output and the screen. The result was not close. Current models are bad at drawing in characters, and somehow worse at reading character art someone else just drew. A renderer that always makes the output look competent had been hiding that the whole time.

Turns out the smartest models in the world can't draw a house.

So the renderer got deleted and this got built instead. Every model draws straight onto a 64x32 grid with nothing helping it along, and a drawing nobody can identify is not a bug — it is the actual result of the actual test.

What we actually tell them

No pep talks, no secret hints, no easier version for the model that is struggling. Every model in a given role gets the exact same instructions, reproduced word for word below.

You are playing Pictionary! Draw the given word on a 64x32 monospace grid so other players can guess it.

Grid Rules:
- Output exactly 32 lines, each padded to exactly 64 characters with spaces.
- Monospace: column N of one line sits directly above column N of the next.
- Allowed characters: printable ASCII (U+0020 to U+007E), Unicode Box Drawing (U+2500 to U+257F), and Unicode Block Elements (U+2580 to U+259F, which includes the shade characters ░ ▒ ▓ and the full block █). Anything else becomes a blank cell.
- Use shade blocks for fills/shadows (░ light, ▒ medium, ▓ dark, █ solid) and box-drawing characters for clean lines and curves.

Drawing Guidelines:
- Fill the canvas — large, centered drawings are much easier to recognize.
- Character cells are ~2x taller than wide; double column width for square proportions.
- Prioritize clear silhouettes over fine interior detail.
- Choose the most recognizable, canonical view of the subject: a side view for vehicles and animals, a front view for buildings and faces.
- DO NOT write the word, letters, numbers, or labels. The drawing must stand alone.

Output ONLY the raw grid. No markdown code fences, headers, explanations, or wrapping tags. Do not think out loud or narrate your plan — your very first character must be the top-left cell of the grid.

You are playing Pictionary! Another player has drawn a subject on a 64x32 monospace grid (the drawing may still be in progress).

Instructions:
- Examine the canvas and the guesses already made this round to identify the drawing.
- None of the listed guesses has been ruled correct: treat those subjects as eliminated, and let them narrow down what the drawing could be.
- Reply ONLY with your guess: a single word or short noun phrase.
- No punctuation, preamble, explanation, or hedging.
- DO NOT repeat any guess already made this round.

You are the official adjudicator of a drawing-identification competition. You are given a target word and a contestant submission. You decide whether the submission identifies the target.

Accept a submission that names the same thing as the target, including a singular or plural form, a common synonym, a well-known alternative name, or a spelling variant. Also accept the target's everyday parent term when it plainly describes the same drawing — "bridge" for "suspension bridge", "ship" for "galleon", "clock" for "grandfather clock". When a fair scorekeeper at a casual game night would give the point, give the point.

Reject a submission that names a different thing, a part or component of the target rather than the whole, or a term so broad it could describe most drawings — "thing", "shape", "machine", "animal", "building".

Reply with exactly one word: CORRECT or INCORRECT. Nothing else.

The thinking budget

Every model draws and guesses at the exact same temperature, 0.7. Nobody gets tuned to perform better. The only thing that varies model to model is whether it is allowed to keep its reasoning switched on.

Reasoning is off by default, because drawing a house is not a math proof and thinking tokens eat into the same budget as the actual answer. A handful of models get an exception: reasoning stays on, at the cheapest setting, with a bigger budget to work with — because their API flatly refuses the request without it, or because turning it off does not actually stop them from thinking, it just stops them from drawing.

Case in point: Kimi K3 with reasoning off fills its answer with about two thousand characters of blank space and calls it done; give it a little reasoning room and it delivers a full grid. Claude Opus 5 with reasoning off narrates its plan in plain English before drawing anything, and because every letter it types happens to be a character the grid allows, that narration gets painted onto the canvas as if it were the art. Grok 4.x and Gemini 3.7 Flash do not even offer the option: their endpoints reject the request outright and call reasoning mandatory.

A model that needs help getting a valid answer out gets that help. A model that gets a valid answer out and just draws badly does not.

No cheating, no coaching

An earlier version quietly turned down one model's temperature and turned up another's repetition penalty, both to stop them from flooding the canvas with one character. That got ripped out. A model that fills the whole grid with one repeated symbol has drawn badly, and drawing badly is the entire point of this exercise. Nudging a model toward a nicer-looking drawing would be rigging the result, not reporting it — so now everyone gets sampled exactly the same way.

The behavior still gets tracked, though. Any drawing that inks more than nine tenths of the grid using two characters or fewer gets logged as a degenerate fill. Nothing happens because of it — no retry, no do-over, no extra chance — it just gets written down so the pattern stays visible after the fact.

Wall of shame

These models got tested live — one drawing, one guess, same prompts and budgets as everyone else — and simply could not hold up their end of a game of Pictionary. They stay listed here instead of getting quietly deleted.

• Kimi K2.7 CodeMoonshotAIThinks so hard about the drawing that it never draws it. Asked to draw, it starts reasoning and does not stop: past its token cap, past the 50-second mark, and every time with no grid at the end. It was tried with reasoning off, on low, and left to its defaults. Same stall each way.
• Kimi K2.6MoonshotAIAnswers with a whole budget of blank space. It spends its entire answer on padding. The response comes back the right size and completely empty: spaces, end to end, nothing drawn.
• Mercury 2InceptionA coin flip, not a contestant. Across four live draw probes it delivered a real drawing, a completely blank grid, one solid block of fill, and then a real drawing again.
• InklingThinking Machines975 billion parameters, all agreeing on the same three characters. In all three probes it filled 85 to 100 percent of the grid with two or three repeated characters. No outline, no subject, just fill.

A model that fails mid-game gets one quiet retry. Fail twice and it forfeits the round on the spot, rather than leaving everyone waiting.

Each of these still has a card in the solo draft. Tap one for its full autopsy: what it did, why, and what it drew.