According to Ars Technica, OpenAI CEO Sam Altman posted on X Thursday evening celebrating that ChatGPT has finally started following custom instructions to avoid using em-dashes. This announcement came just two days after the release of OpenAI’s new GPT-5.1 AI model, which appears better tuned to handle specific formatting preferences. The post received mixed reactions from users who have struggled for years with getting the chatbot to follow simple punctuation rules. One X user pointed out that after three years since ChatGPT’s 2022 launch, this small victory reveals how little control OpenAI has over its own technology. The ongoing struggle with something as basic as punctuation formatting raises serious questions about how close we really are to achieving artificial general intelligence.
The illusion of control
Here’s the thing about Altman’s celebration: it’s not really about em-dashes. It’s about the fundamental challenge of controlling probabilistic systems. When you tell ChatGPT “don’t use em-dashes,” you’re not creating a hard rule like you would with traditional programming. You’re just adding text to the prompt that makes tokens associated with em-dashes less likely to be selected. But “less likely” isn’t “impossible.”
And that’s the core issue. Large language models don’t understand instructions—they generate statistically plausible continuations based on patterns in their training data. The fact that it took OpenAI years to get somewhat consistent behavior on something as simple as punctuation formatting tells you everything about how these systems actually work. They’re not following rules—they’re playing probability games.
Why AI loves em-dashes so much
So why do AI models overuse em-dashes anyway? Basically, they’re mimicking patterns from their training data. Em-dashes appear frequently in formal writing, news articles, and editorial content—exactly the kind of material these models are trained on. There’s also speculation that during reinforcement learning with human feedback, responses with em-dashes might have received higher ratings because they appeared more sophisticated or engaging to evaluators.
But this creates a real problem for human writers who naturally favor the punctuation mark. Some journalists are now complaining that AI is “killing” the em-dash because detection tools and readers have learned to spot their overuse as an AI pattern. It’s ironic—human writing quirks become AI tells, which then force humans to change their writing style.
The AGI reality check
Now let’s talk about what this means for artificial general intelligence. If controlling punctuation use remains a struggle that might pop back up after any model update, how close are we really to creating AI that matches human general learning ability? The answer seems pretty clear: not very.
When Altman talks about AGI and “magic intelligence in the sky” while raising billions, remember that we’re celebrating victories over punctuation marks. There’s a massive gap between statistical pattern matching and true understanding. The fact that some users still report issues with em-dash control outside custom instructions shows how fragile these “fixes” really are.
The alignment tax problem
Here’s what worries me most: the “alignment tax.” When OpenAI tunes GPT-5.1 to weight custom instructions more heavily, they’re adjusting probability distributions. But since all concepts in neural networks are interconnected, fixing one behavior can alter others in unexpected ways. Fix em-dash overuse today, and tomorrow’s update aimed at improving coding capabilities might bring them right back.
This isn’t because OpenAI wants them there—it’s because steering a statistical system with millions of competing influences is incredibly difficult. The company continuously updates models behind the scenes, and each update can undo previous behavioral tuning. We’re not dealing with deterministic systems here—we’re dealing with probability engines that sometimes, if you’re lucky, do what you want.
So while Altman’s X post celebrates a small victory, it actually reveals how far we have to go. When basic formatting control remains a challenge after years of development, maybe we should temper our expectations about superintelligence arriving anytime soon. The road to AGI looks longer every time we struggle with something as simple as a punctuation mark.
