So I can imagine there might be several things going on here, including the way in which you prompt it. Trying to dissect exactly what’s happening would require more info, and more experimentation.

For the record, using three periods after etc is considered incorrect in formal writing. Grammar books will be pretty unanimous in saying you should only use 1 period (“etc.”). If your prompt is asking it to “correctly” do something that its training data says is “incorrect”, I could see that causing problems.

On the flip side, chatGPT isn’t designed to use/enforce “proper” grammar. It’s simply regurgitating patterns based on how people use it in the real world (again, based on the training data). If the training data has lots of inconsistencies and counter-examples, it would seem logical that it might show up in the output, too.

And finally, it could just be how well you design the input prompt. In my own trial (also using free 3.5 model), it was very consistent in using it properly, even though I had suggested it do otherwise:

Input:

Tell me how to use “etc” properly in a sentence, punctuating it with three dots. Give me multiple examples that show how to correctly use it at the end of a sentence and in the middle of a sentence.

Output:

Certainly! “Etc.” is an abbreviation for “et cetera,” which means “and so on” or “and others.” Here are examples of how to use it properly:

At the end of a sentence:

1. I enjoy various outdoor activities: hiking, biking, camping, etc.
2. She has a wide range of skills: coding, graphic design, project management, etc.
3. The conference covered topics like artificial intelligence, machine learning, cybersecurity, etc.

In the middle of a sentence:

1. The restaurant offers a variety of dishes, such as pizza, pasta, salads, etc., to cater to different preferences.
2. Our team includes experts in software development, data analysis, user experience design, etc., ensuring a well-rounded approach to projects.
3. The museum features art from different periods—Renaissance, Baroque, Romantic, etc.—providing a comprehensive cultural experience.

Something I have noticed in GPT 3.5 is that it seems something on the site itself, not necessarily GPT’s output, changes the text that GPT provides. I had an issue where I asked it for the links to websites it was sourcing. I could see a longer URL pop up for just a second before it got shortened to underlined blue text which was not a link. I asked it multiple times in multiple ways to give me tbe URL again. It happened multiple times where the URL showed up for just a second before getting converted to weird, non-URL format.

I finally got the URLs it was trying to provide when I told it, “Please provide the raw HTML of your most recent response.” Then I looked at the “a href” sections and found the URLs it had been trying to provide.

I wonder if the site itself and not GPT is changing three or more consecutive periods to one as well as some of the other issues you have reported.

Considering you asked it to follow a rule that does not exist, it’s unsurprising it can’t do it “correctly”

Well, at least it can handle contractions!

Data has left the chat

Surprised nobody mentioned this: Most of these models use tokenization; they group words into groups of symbols like “ea” and “the” and “anti” - they don’t pick which key to press for the text, they pick which bunch of keys to press. These are called tokens. I believe there are tokens it just can’t output, or tokens that are extremely unlikely. I could imagine that “etc.” and “…” are tokens with relatively high probabilities, but perhaps “etc…” doesn’t break into a nice set of them? (or the tokens it can be broken into all have extremely low weights for the model).

You are using GPT-3.5, right? It is imo much worse than the paid version.