On a related note, has anyone noticed that ChatGPT seems to have the same writing style and verbal tics as Claude, now? Not sure which one developed them first, but I have a suite of “creative” benchmarks where the LLM is asked to implement obscure computer programs that probably aren’t in the training set, and Claude and OpenAI tend to aim for the exact same visuals and flavor text. When asked to solve a task that involves the color green, both go for ‘verdant’, for instance.
Unprompted, Kimi K3 claimed to be Claude in 40% of direct queries
I feel like it has to be something weirder than the naive answer, simply because you could just auto-replace Claude with whatever your LLM’s name is in the training data.
If I had to guess, part of it is that Claude is a real name, so the base model is more inclined to agree with “you are Claude” than with “you are ChatGPT”. The other part is that all LLMs are trained to have roughly the same persona (Diligent, inoffensive, obedient, neurotic), and given that you can get an LLM to identify as Hitler by training it on his favorite foods and musicians[1], I have to imagine that there’s a basin of attraction towards “I am Claude” for any model trained on the modern internet.
If I had to guess, part of it is that Claude is a real name, so the base model is more inclined to agree with “you are Claude” than with “you are ChatGPT”.
I had the same hunch, this would be good to control for.
Yes, that’s an interesting point! I guess there are a couple of things at play: 1. Model outputs are increasingly going to enter both pre and post-training data, given that there is lots of explicitly AI generated text out there, and also humans-using-AI to generate text is a larger and larger fraction of all text on the internet. There is an interesting question here, what this increasing amount of AI generated content/data mean for their training.
2. Models, from their training data, will learn lots about Claude/ChatGPT/other LLMs, and they’ll form representations of each other, and also of themselves. Not sure if this would change their own behaviour, but, as you say, there will be these basins, that may affect the models’ own identity basin.
On LLMs being trained to be Hitler—yes, we’ve also shown that this can easily happen via in-context learning (using the same dataset as the Weird Generalisation paper), without fine-tuning, and many models are super happy to play along (some refuse though).
On a related note, has anyone noticed that ChatGPT seems to have the same writing style and verbal tics as Claude, now? Not sure which one developed them first, but I have a suite of “creative” benchmarks where the LLM is asked to implement obscure computer programs that probably aren’t in the training set, and Claude and OpenAI tend to aim for the exact same visuals and flavor text. When asked to solve a task that involves the color green, both go for ‘verdant’, for instance.
I feel like it has to be something weirder than the naive answer, simply because you could just auto-replace Claude with whatever your LLM’s name is in the training data.
If I had to guess, part of it is that Claude is a real name, so the base model is more inclined to agree with “you are Claude” than with “you are ChatGPT”. The other part is that all LLMs are trained to have roughly the same persona (Diligent, inoffensive, obedient, neurotic), and given that you can get an LLM to identify as Hitler by training it on his favorite foods and musicians[1], I have to imagine that there’s a basin of attraction towards “I am Claude” for any model trained on the modern internet.
Yes, really. https://arxiv.org/html/2512.09742v1
I had the same hunch, this would be good to control for.
Yes, that’s an interesting point! I guess there are a couple of things at play:
1. Model outputs are increasingly going to enter both pre and post-training data, given that there is lots of explicitly AI generated text out there, and also humans-using-AI to generate text is a larger and larger fraction of all text on the internet. There is an interesting question here, what this increasing amount of AI generated content/data mean for their training.
2. Models, from their training data, will learn lots about Claude/ChatGPT/other LLMs, and they’ll form representations of each other, and also of themselves. Not sure if this would change their own behaviour, but, as you say, there will be these basins, that may affect the models’ own identity basin.
On LLMs being trained to be Hitler—yes, we’ve also shown that this can easily happen via in-context learning (using the same dataset as the Weird Generalisation paper), without fine-tuning, and many models are super happy to play along (some refuse though).