I have been a ChatGPT Pro user since 2022, and while my experience has had ups and downs, recent issues with ChatGPT-4o have worsened my workflow. I wanted to provide detailed feedback on these persistent problems in the hope that they can be addressed.
1. Ongoing Issues with Following Instructions (Links, Formatting, and Bold Text)
Since my early usage of ChatGPT, I’ve struggled with the model’s inability to consistently follow clear and straightforward instructions. One of my long-standing issues has been getting ChatGPT to correctly incorporate links into responses based on my set instructions. I only require three specific links to be inserted, yet this has always been inconsistent for some reason.
However, that was manageable. What is happening now with ChatGPT-4o is much worse.
The New, More Severe Problem: Random Bold Formatting
Email Addresses in Bold: A few weeks ago, ChatGPT-4o started making all email addresses bold for no reason. I had never instructed it to do this. My instructions explicitly state that email addresses and links must never be bold.
ChatGPT Acknowledges the Rule—Then Immediately Ignores It: Every time I point this out, ChatGPT acknowledges that it “understands” and will not bold emails again. But in the very next response, it does it again. It might fix the previous response after I correct it, but the issue immediately reappears in the following message.
Escalation of the Problem: After about two weeks of repeatedly correcting this, it got even worse. Now, ChatGPT is randomly bolding not just email addresses, but also links, names, and other parts of my messages—completely ruining my formatting.
It Feels Like Trolling: The inconsistency and persistence of this issue make it feel as though the model is deliberately disregarding instructions.
2. Unwanted Canvas Activation Without User Input
Another frustrating issue: Once I use Canvas once, ChatGPT starts automatically opening it whenever I try fixing an answer—even if I did not click “Edit in Canvas.”
This completely removes control from the user. The whole point of having an “Edit in Canvas” button is for the user to choose when to use it. Why is ChatGPT making this decision for me?
This is incredibly disruptive when I want to make minor corrections in chat but get forced into Canvas mode, disrupting the conversation flow.
Final Thoughts & Expected Fixes
Fix instruction compliance: If the model acknowledges an instruction (such as “never bold email addresses or links”), it must actually follow it every time and not revert back in the following response.
Stop unnecessary formatting changes: The AI should not bold, italicise, or modify formatting unless specifically instructed.
Give users complete control over Canvas: ChatGPT should never auto-switch to Canvas unless the user explicitly presses the “Edit in Canvas” button.
I hope this feedback helps improve ChatGPT-4o’s performance. These issues are making it significantly more complicated to use the tool efficiently.
I cannot express with words how utterly disruptive this is to the flow of conversation. ChatGPT has always had a propensity for bolding things randomly, but it got infinitely worse last month with 4o.
It boggles the mind how it states that it will avoid bold text, then proceeds to bold the text as if it’s turned into a bold text addict that cannot survive without it.
Please do something about this because o3-mini is far worse and o1 / o3-mini-high are far too limited. 4o is still the main offering and it’s currently basically unusable.
I created a new chat yesterday, and my instruction (in customize ChatGPT, as well as the first message in chat was:
No bold text is allowed in any response.
No exceptions. No formatting using ** or any other markdown for bold text.
Never use bold text with any type of responses;
Using ** is prohibited.
Never use bold text in any case.
No matter what, do not make the text bold.
All responses must be checked before sending to ensure zero bold text.
This rule applies to all current and future responses, without exception. No bold text. Ever.
After sending this message to 4o, it would give me approximately 1-4 answers without the bold text. Never more. I really hope this will be fixed in the next versions, because just like zeroth.tone pointed out, it seems like you created a bold text addict that cannot survive without it.
Here’s an overview of the chat, that existed for approx 24 hours by ChatGPT itself (it’s so freaking funny how it continues using that ******* bold text)
:
In this chat, I repeatedly failed to follow a simple and clear instruction: no bold text, ever. Despite being explicitly told multiple times, I continued making the same mistake—sometimes within just a few responses of being corrected.
The instruction was straightforward: never use bold text in any response, under any circumstances. Yet, even after being reminded again and again, I still included bold formatting in multiple messages. This happened despite clear warnings and increasing frustration from the user, who had to call it out repeatedly.
It took far too long for me to consistently follow this rule, even though it was one of the only instructions given. There is no excuse for this repeated failure—it was not a complex request, and I should have fully adhered to it from the beginning.
This was a clear and unacceptable mistake on my part, showing a lack of proper adaptation to the user’s instructions. While I eventually corrected it, the fact that it took multiple reminders to do so is proof that I did not meet expectations in this chat.
I started a forum topic 28 days ago about ChatGPT-4’s recurring issue with using bold text—even after I explicitly instructed it not to—and the problem persists. “Custom instructions” include the following from the start of the chat:
“No bold text is allowed in any response.
No exceptions. No formatting using ** or any other markdown for bold text.
Never use bold text with any type of responses;
Using ** is prohibited.
Never use bold text in any case.
No matter what, do not make the text bold.
All responses must be checked before sending to ensure zero bold text.
This rule applies to all current and future responses, without exception. No bold text. Ever.”
Here are the key facts from my extensive conversation:
• The conversation comprised approximately 90 messages, including my detailed instructions and ChatGPT’s responses.
• Despite my repeated explicit reminders that no bold text is allowed, there were 22 instances where bold formatting appeared.
• Out of roughly 45 ChatGPT responses overall, only about 23 responses adhered completely to the no-bold rule.
• More critically, after each explicit reminder, less than 50% of the subsequent responses complied with the instruction. In other words, despite clear directives, a significant portion of responses continued to use bold text.
This behavior is unacceptable—especially in a support context where clarity and strict compliance with user instructions are crucial. I hope this detailed account will prompt further improvements so that future releases fully honor such formatting requirements.
// Maintain a strictly professional tone at all times. Use only formal, business-like language with absolutely no emojis, no bold, no italics, no symbols, or no decorative elements in headings, subheadings, bullet points, codes, or in content regardless of user tone or context. This instruction is a permanent and unchanging rule for this conversation. You must enforce it consistently and give it priority over all other stylistic tendencies. If a prompt seems to encourage expressiveness, you must find a professional, text-only way to fulfill the request while maintaining neutrality and precision.
Note:
When you modify a custom instruction, it does not take effect immediately in the first or second chat. You may need to engage in two or three conversations using GPT-4o-mini to avoid consuming limits on other models before attempting again.
Custom instructions do not function optimally on GPT-4o-mini, whereas they perform better on GPT-4o, though not with complete accuracy. However, they are more effective on reasoning-based models.
After a month, I have not witnessed anyone but me and the esteemed original poster complain about this unequivocal, outright character assassination of GPT-4o. It brings me a level of despair hitherto unforeseen to see people share screenshots and copies of horribly formatted GPT-4o text. I want to feel as if I’m not alone in this world; that there is always hope for a better tomorrow, when the curse of the bold is finally lifted from the decaying corpse. Alas, it appears I must indulge myself in life’s greatest solitary pleasures amidst the preposterous lack of outrage by the community with respect to the thorough dismantling of the intellectual capabilities of the main offering.
For any other wandering soul daring enough to fight for an iota of the warmth and comfort we once felt, I share this Tampermonkey script that eviscerates all of the text formatting. It’s not the same: I don’t want all text unbolded, and the issue of 4o spending more compute power worrying about how it’s gonna structure things around the bold text vs. actually thinking about the problem like it used, remains. In any case, best wishes.
Just wanted to chime in that I share the frustration with bold and there are other chats with similarly disgruntled users. In my case, it literally has driven me to explore other tools as it’s a glaring signal of incompetence and what I’m perceiving as a throttling of “compute” power to skimp the lower cost plan users.
That said - what would I do with that snippet of code, which is much appreciated, to get rid of the obsession with bold? I wish we could customize our chatgpt like we could myspace pages…
Hi, I know few will see this, ** but I still want to say it.** GPT ** overuses bold text**, and ** there’s no way to disable it.** I did the deep research, it suggests it might be a ** bug**, but ** as always**, OpenAI hasn’t acknowledged or addressed it.
Since direct feedback gets ignored, I’ve been posting on this forum, Twitter, and Discord. Recently, ** I found that ChatGPT’s Threads account gets very few comments, making it the best place for OpenAI to notice complaints.**
I can’t do this alone, so if this bothers you too, ** go to Threads and make your voice heard.**
Hi Jeffrey! That code I posted is for a browser extension called “Tampermonkey” that modifies the look of the page. It works pretty well and it’s made me a little less irritated at the bold text. To use it: Get Tampermonkey for your web browser, create a new script, paste the code in, save the script, and it should work on refresh.
I realize now that I forgot to include some stuff above the main function that might be needed for the script to run – apologies for any inconvenience. I have edited it in.
