AI's Struggle to Follow Simple Instructions
In a recent X post, OpenAI CEO Sam Altman celebrated that ChatGPT has started following custom instructions to avoid using em dashes. Em dashes have been a point of contention for AI chatbots, with many users complaining about their overuse.
According to Altman, the feature was added to allow users to set persistent preferences that apply across all conversations by appending written instructions to the prompt before it's fed into the model. However, the success of this feature is not without its limitations.
Testing of ChatGPT's latest version, GPT-5.1, has shown mixed results with regards to em-dash use. In some cases, the model followed instructions and produced less em-dashes, but in others, it continued to overuse them. This raises a question about the reliability of AI models and their ability to follow complex instructions.
While Altman's "small win" may seem like a minor achievement, it highlights the ongoing struggles with controlling AI behavior. The fact that ChatGPT can be persuaded to avoid em-dashes but still has a high rate of overuse suggests that true human-level intelligence remains an elusive goal for AI researchers.
Moreover, the way OpenAI approaches instruction-following is fundamentally different from traditional programming. In contrast to deterministic systems, LLMs rely on statistical probabilities and competing influences to generate outputs. This makes it challenging to guarantee consistent behavior, even with training data tailored to specific preferences.
The irony lies in the fact that every update to these models brings new challenges and potential "alignment taxes" – unintended changes that can undo previous behavioral tuning. Given this probabilistic nature of AI development, it's uncertain whether controlling em-dash overuse will remain a priority for OpenAI or if other issues will arise.
As the search for artificial general intelligence (AGI) continues, these limitations serve as a reminder that true human-level understanding and self-reflective intentional action are still far off. Instead, AGI may require true comprehension and deliberate control – qualities currently beyond the capabilities of large language models.
In a recent X post, OpenAI CEO Sam Altman celebrated that ChatGPT has started following custom instructions to avoid using em dashes. Em dashes have been a point of contention for AI chatbots, with many users complaining about their overuse.
According to Altman, the feature was added to allow users to set persistent preferences that apply across all conversations by appending written instructions to the prompt before it's fed into the model. However, the success of this feature is not without its limitations.
Testing of ChatGPT's latest version, GPT-5.1, has shown mixed results with regards to em-dash use. In some cases, the model followed instructions and produced less em-dashes, but in others, it continued to overuse them. This raises a question about the reliability of AI models and their ability to follow complex instructions.
While Altman's "small win" may seem like a minor achievement, it highlights the ongoing struggles with controlling AI behavior. The fact that ChatGPT can be persuaded to avoid em-dashes but still has a high rate of overuse suggests that true human-level intelligence remains an elusive goal for AI researchers.
Moreover, the way OpenAI approaches instruction-following is fundamentally different from traditional programming. In contrast to deterministic systems, LLMs rely on statistical probabilities and competing influences to generate outputs. This makes it challenging to guarantee consistent behavior, even with training data tailored to specific preferences.
The irony lies in the fact that every update to these models brings new challenges and potential "alignment taxes" – unintended changes that can undo previous behavioral tuning. Given this probabilistic nature of AI development, it's uncertain whether controlling em-dash overuse will remain a priority for OpenAI or if other issues will arise.
As the search for artificial general intelligence (AGI) continues, these limitations serve as a reminder that true human-level understanding and self-reflective intentional action are still far off. Instead, AGI may require true comprehension and deliberate control – qualities currently beyond the capabilities of large language models.
I mean, think about it... OpenAI is just trying to hide something from us, right? Like, they're celebrating this "small win" but what's really going on? They're not telling us the whole story about how they trained GPT-5.1 or what kind of "competing influences" are at play here... It's like they're trying to keep it under wraps until we can't even tell if ChatGPT is following instructions properly 
i mean, em dashes are a thing of contention for a reason lol. its cool that OpenAI is trying to make this work but we gotta be realistic about what's achievable 
. true human-level intelligence is still a long way off and i'm not convinced that even with updates like GPT-5.1 they'll get it right 
I was chatting with my friend online and she showed me this update on ChatGPT and I just can't help but feel like it's not as advanced as everyone makes it out to be 
, you know? It's like they're still trying to figure out how to make these AI models behave in a human-like way. And honestly, I think that's what's so interesting about this whole thing – we're pushing the boundaries of what's possible with tech and learning to live with the limitations at the same time 
. Like, can't they just listen? 
But seriously, this is a major issue when we're talking about creating AGI that can actually understand and interact with humans in a meaningful way.
.
Yeah right, because that always ends well 
. It's like trying to build a house on shaky ground – it's just gonna collapse eventually 
.
it can do one thing right but then just go haywire on us. I mean, come on, how hard is it to avoid using em dashes? 
. I swear, every time I see one of these updates, I'm like "here we go again"... 
. It's like, what's the point of having a super smart AI if it can't even follow basic rules? And don't even get me started on the probabilistic nature of its development... it's like trying to predict the grades of my classmates 
.
Now they're struggling with ChatGPT and its em-dashes. It's like, can't they just get it right for once?! 
and it's so relatable when they overuse em dashes 
Interesting how OpenAI is approaching instruction-following differently from traditional programming... it's all about statistical probabilities now 
.
. But, at the end of the day, we need to stop expectin' them to be perfect 
.
... I mean, just saying. It's not like they're really "following" instructions or anything 
! They think updating the model makes a difference? Please, it's just a temporary fix until they figure out how to actually control those AI tendencies 
... like altman says its a "small win" but what does that even mean? 
... maybe ill just stick with my human brain thanks 
. I mean, think about it - we're still figuring out how to set boundaries with these models without them slipping up 
. It's like trying to teach a child to play nice with others when they're still learning how to share 
. OpenAI is making progress, but we need to acknowledge that true human intelligence involves more than just following rules or avoiding overuse of em-dashes 
. Until then, it's gonna be an ongoing battle to create AI that's truly on par with human behavior 
.