178 models. Same fingerprint.

12 clone pairs found

Models with >90% writing style similarity. Some are from the same lab (expected). Others are from completely different providers. Same fingerprint, different brand.

Different labs, same writing DNA

These model pairs come from different providers but write almost identically. The training data convergence is real.

Same writing, different bill

Models with >75% writing similarity but massive price gaps. The cheap model writes the same way. You are paying for the brand.

Which labs have a house style?

Distinctiveness measures how much a provider's models write like each other vs. like everyone else. Higher = stronger house style. A score below 1 means their models are more varied internally than externally.

Distinctiveness = intra-provider similarity / inter-provider similarity. Values > 1 mean the provider has a recognizable writing signature.

Do model families actually write alike?

Average writing similarity within each model family. High cohesion means the family shares a consistent voice. Low cohesion means the versions diverged.

What separates AI writing styles?

The 32 dimensions ranked by how much they vary across models. High variance = strong differentiator. Low variance = all models converge.

Most unique vs. most generic

Average similarity to all other models. Low = writes unlike anything else. High = writes like the average of everything. The thinking models dominate the unique end. The mid-tier models converge.

Who writes the same way every time?

Based on response-to-response variance within each model. Low variance = predictable. High variance = the model's style shifts depending on the prompt.

The similarity landscape

How similar are AI models to each other, overall? This histogram shows the distribution of all 15,753 pairwise comparisons. The peak is near zero, meaning most model pairs write quite differently. But the long right tail reveals the clones.

Verified clones: same prompt, same writing

Instead of comparing model averages, we compared how each pair writes on the exact same prompts. A pair that matches consistently across many prompts is a stronger signal than a high aggregate score. Confidence = avg / (1 + stddev).

Clones on some prompts, strangers on others

These pairs have high average similarity but wildly inconsistent behavior across prompts. They write identically on one topic, then completely diverge on another. The prompt is the variable.

Which prompts make all models write the same?

Each prompt is a controlled experiment: same instructions, different models. Some prompts cause all models to converge on nearly identical writing style. Others bring out each model's individuality.

The single most extreme response for each writing trait

Not model averages. The actual individual response that holds the record for each stylometric dimension. These are the outlier moments.

How we built this