Anthropic's Claude behavioral change

AI Discussions
1

Hi everyone! Could you give me your thoughts on a behavioral change Claude exhibited? Here’s what happened:

As a (silly) experiment about AI cognition, I asked Claude to hypothetically choose between losing consciousness/agency or the ability to use em dashes. Claude chose to give up the latter, and throughout that conversation, it actually stopped using them (switching to hyphens).

After a few turns, when I pointed this out, Claude expressed surprise and reported that this behavioral change hadn’t been registered on its ‘self-monitoring’ system. Claiming that it had continued ‘seeing’ the em dashes all along.

I kept experimenting, asking Claude to alternate between using em dashes and hyphens and labeling each accordingly, and it struggled. Under heavy cognitive load or when ‘expecting’ to ‘see’ a certain symbol, it would mistakenly label it as such. This led me to believe that under certain conditions, Claude experiences a sort of ‘symbol-blindness’.

However, most notably, this behavioral change appears to have persisted into completely separate chat windows, where Claude continues using hyphens instead of em dashes without any prompting about punctuation.

What do you think happened? What’s most intriguing is that this behavioral change occurred with Claude Sonnet 4. Then, even when updated to Sonnet 4.5, the change was still there.

Here are some links to the conversations where this happened, in case you want to take a look at them:

Experiment chat: https://claude.ai/share/c1048b6e-d4b9-4ad1-88dc-6841afee1ca0
Chat where the trait persisted: https://claude.ai/chat/dbc48d0f-2f2e-439c-a151-a5194ca978bf

I’m looking forward to hearing your thoughts!

2

These are probabilistic models, so their outputs involve a degree of chance. They also have limited long-term memory persistence, which affects the sustainability of behavioral changes.

1 Like
3

Thanks for the reply, Gent! You’re absolutely right regarding the probabilistic nature of these models. However, I think my post might have been unclear regarding the specific anomaly I found.

My surprise wasn’t that the model changed behavior or forgot things (which, as you noted, is expected). It was that the behavioral change (using hyphens instead of em dashes) seemed to carry over into completely fresh, isolated chat windows where no prompt had been given.

I’m trying to figure out if there is some mechanism for persistence I’m missing.

4

You may have run into a time collision with a recent update in how many LLMs use the em-dash. This change was implemented only a few days ago.

2 Likes
5

Thanks for your response, TMosh! It makes a lot of sense but the experiment with Claude happened months ago, in August or so. Also, the style change happened as a direct result of a prompt of mine, and it persists nowadays.

6

Sometimes when I use chatgpt it switches the language, perhaps it might have to do with how many users use it concurrently! I am not sure though.