When AI ostensibly overwhelmed the world 鈥� not only in tech and innovation but across industries and everyday applications too 鈥� nobody expected that one of its most controversial features would be a punctuation mark.
The poor em dash. Hated, loved and just about everything in between.
And now, the em dash is making headlines once again, and this time, it鈥檚 because Sam Altman has announced to the world that it鈥檒l no longer be a central component of ChatGPT鈥檚 writing style. Essentially, OpenAI has added a feature that allows users to prompt it to specifically 苍辞迟听use the em dash in its writing 鈥� something that wasn鈥檛 previously possible, believe it or not.
听
听
But, why does this matter?
Well, it was about more than just the fact that there was no 鈥渙pt out鈥� option when it came to using the em dash and AI-generated content via ChatGPT. Rather, it goes back to the fact that when ChatGPT was first released to the public 鈥� at least, in GPT鈥檚 early days 鈥� it became quite clear that the AI model used the punctuation mark liberally (to put it lightly). Thus, during a time in which everybody was panicking about whether or not it would be possible to know what was human-generated and what was AI-generated, the em dash fell into a gloomy pit of martyrdom from which it has not yet returned.
Essentially, the dominant belief was: if a text uses an em dash or two, red flag. If it uses the em dash multiple times, you鈥檙e almost definitely dealing with AI-generated content.
Further investigation? Not necessary! It鈥檚 as simple as that.
Of course, it wasn鈥檛 and it isn鈥檛, but that didn鈥檛 change the tides from turning.
So began the rapid decline of the em dash 鈥� a punctuation mark so controversial, nobody wanted to touch it. The punctuation mark that inspired fear and trepidation just at the very sight of itt. Those who steered clear were 鈥渟ensible鈥�, and those who continued to use it were 鈥渂rave鈥� 鈥� or reckless?
Regardless, never has a punctuation mark stood out quite so boldly 鈥� edging on jump scare levels of concern.
Well, with Atlman鈥檚 newest update, the em dash may finally be off the proverbial chopping block, but the ordeal has caused a significant enough stir to raise a lot of questions in the world of AI and AI detection. The first question asked by many was, if the em dash is no longer in the picture, how, oh how, will we be able to detect AI-generated text?
Of course, that鈥檚 not only a silly question, but it鈥檚 the wrong question to ask. The em dash听was never听a good or reliable indication of that content was written by AI.
So, what should we be questioning?
I鈥檓 glad you asked.
听
The Fundamental and Philosophical Issues We Should Really Be Pondering
听
The most interesting part of the em dash debate is that, in my opinion, it does two things:
听
- It identifies and completely overemphasises a completely unreliable way of detecting AI use
- It puts too much focus on the importance of detecting AI at all
听
And, potentially most controversially, creates and reinforces an assumption that detecting AI reliably is actually possible. However, at this point, we haven鈥檛 really seen any instances of this being true. So, essentially, everybody鈥檚 getting heated about something that鈥檚 not really even conceptually plausible, and the poor em dash has taken the brunt of it.
Now that the em dash may, hopefully, be redeemed in the eyes of the world, it might be time that we step away from the basic AI detection panic that鈥檚 shrouded everything we鈥檝e done since ChatGPT hit the streets and start considering the deeper, more complex and far more important reality of AI and how we choose to use it going foward. But to do that, we need to first broach a few touchy topics.
Brace yourself; these are the bits everyone鈥檚 been avoiding.
听
Was the Em Dash Ever a Reliable Indicator of AI?
听
No, the em dash was never a reliable indicator that content was written by AI.
Did ChatGPT use the em dash a lot? Undoubtedly. Indeed, it certainly seemed like it was more likely to write using the em dash than it was to write without it, but there is no logical progression from that assumption to the notion that any text containing an em dash must have been written by AI.
If anything, the em dash ordeal made reliable detection more complicated than ever before.
听
What听Is听a Reliable Indicator of AI Use In Writing?
听
If we can鈥檛 put all our eggs in the em dash basket, where should we put them?
Well, if only it was that simple. The em dash is not the gold standard for being able to identify AI-written content, but we don鈥檛 have another reliable indicator, and for many people, that鈥檚 been an uncomfortable truth they鈥檝e tried really hard to avoid.
The output of AI models is completely dependent on the data upon which it鈥檚 trained. If the style of writing from which the model learns writes in a certain way (whether that鈥檚 using specific punctuation, style, tone and more), that鈥檚 what you鈥檙e going to get from the AI. Today it鈥檚 the em dash; tomorrow, it may be something completely different.
And that鈥檚 the whole problem. Data sets used for training are surely going to grow and become increasingly complex, meaning that models will have more to learn from and the style of writing will (in thoery) become more varied, more complex and just generally better.
Ultimately, because we鈥檙e trying to improve the output of these AI models 鈥� that is the ultimate goal, after all 鈥� we鈥檙e simultaneously making it more and more difficult to ever be able to reliably detect AI. The two principles and objectives 鈥� creating better, more humanlike AI-generated content听and听being able to detect AI-generated content reliably 鈥� are fundamentally imcompatible. We simply cannot have the best of both worlds 鈥� it鈥檚 conceptually contradictory.
And we know that we鈥檙e moving towards better AI 鈥� there鈥檚 not much we can do to stop that. So, it follows that detection is becoming further and further from our grasp.
听
Should We Be Focused On Detecting AI? Or Should We Emphasise How We Use AI?听
听
That leads us to the biggest, most challenging question of all 鈥� a philosophical conundrum that may send you into an existential spiral.
Should we be completely focused on ensuring that all content is written by humans and humans only, without any exception? Or should we be starting to consider a hybrid reality in which we try to understand how we can and should use AI to learn and improve our processes instead?
Well, we chatted to a group of experts to find out just what they think about the topic.
听
Our Experts:
听
- Rich Pleeth:听CEO and Co-Founder of Finmile
- Andy Zenkevich:听Founder & CEO at Epiic
- Pilar Lewis: PR and Media Relations Expert at Marketri
- Rhys Merrett: Senior Vice President, Technology at The PHA Group
- Jake Atkinson: Director of Growth at MQube
- Nathan Selby: Managing Director at Resultful
- Camden Woollven: Head of Strategy and Partnership Marketing at GRC International Group Ltd.
- Carolyn Shelby: Principal SEO at Yoast
- Zachary Cote: Executive Director at Thinking Nation
- Nicole Franco: Head of Digital PR AI Innovation at Fractl
- Gailene Nelson:听Senior Director of Product,听Turnitin
听
Rich Pleeth, CEO and Co-Founder of Finmile
鈥淭he em dash was a quick human detection that things were written by AI. It was not a serious signal for GPTZero or other tools to say it was AI. The signals for tools to detect AI are too perfect sentences, zero use of any extra words and zero mistakes as well as many other functions.
鈥淎s 2026 approaches we鈥檙e going to see human writing and AI blur more because of the improvements in AI. We should really be looking at what the future looks like, that is that the future will be collaborative, ai and human writing. With Ai enhancing the speed and ensuring that the structure is correct.鈥�
Andy Zenkevich, Founder and CEO at Epiic
鈥淪urface signals like the em dash have always been a dead end for detecting AI writing. They鈥檙e too easy to fix. What truly distinguishes AI writing in my agency鈥檚 audits is redundancy. For example, the stock phrase 鈥淚t鈥檚 not just X, it鈥檚 Y, and it鈥檚 also Z.鈥� AI uses a lot more of those than we鈥檇 see in human drafts, especially at scale. Whereas the em dash is easy sauce; the AI can just stop using it.
鈥淲e should probably acknowledge that we鈥檙e moving from a race to the bottom (identifying minimal 鈥済otchas鈥� that AI can avoid) to a richer space where human values have more room to express themselves. In particular, I鈥檓 convinced disclosure about AI use will become exponentially more important. Our mood toward AI perks up when we鈥檙e reminded of its involvement in content, and it nosedives when we forget. The ideal future would involve AI detection tools serving primarily as prompts to encourage human writers or editors to review and revise content, or at least attach disclaimers, rather than functioning as blunt bans. In other words, the em dash retreat is just the beginning.鈥�
Pilar Lewis,听PR and Media Relations Expert at Marketri
The whole em-dash debate was never a 100% reliable way to tell whether something was written by AI. It was a predictable pattern people latched onto. Now that ChatGPT is using fewer em dashes, that supposed pattern goes away, but the bigger issue has nothing to do with punctuation.
What we should really be paying attention to is how dependent people have become on plug-and-play content. More and more, writing starts with an AI prompt instead of human thought. People take whatever the model gives them and move onto the next task. That鈥檚 why everything has started to blend together and it鈥檚 why people get so anxious about being detected. If the work doesn鈥檛 originate from your own perspective, of course it鈥檚 going to feel generic.
It鈥檚 not about if AI detection works, but that we鈥檙e letting AI do the heavy lifting before we鈥檝e even figured out what we want to say. Strong writing still comes from your ideas first. AI is at its best when it helps refine something you鈥檝e already shaped yourself.
If there鈥檚 anything to take away from the em-dash conversation, it鈥檚 that we should stop fixating over signals and start focusing on the quality of our thinking.
Rhys Merrett, Senior Vice President, Technology at The PHA Group
鈥淭he em dash discussion demonstrates current public perception towards AI.
鈥淔or communication professionals, there are clear ways to detect when AI has been used beyond the em dash. The sentence structure is formulaic and does not reflect a general flow of narrative one would come to expect from a human writer. There鈥檚 US spelling, sentence case headlines and bullet points, not to mention different software that can detect when AI has been used, and to what extent.
鈥淥ver the past year, two big truths have emerged. First, large language models still struggle to create content that feels truly human. Yes it can generate hundreds of words in seconds, but if the quality is not there, the content is useless. Second, audiences value authenticity more than ever. AI can absolutely speed up the process and get you 80% of the way there, but it鈥檚 the final 20% that makes all the difference. That鈥檚 where creativity, judgment, and emotional intelligence come in which are effectively human-driven.
鈥淯ltimately, we need perceptions of AI to change. It is not a solution in of itself, but a way for us to get to solution faster.鈥�
Jake Atkinson,听Director of Growth at MQube
鈥淧unctuation marks have long had an impact in inverse proportion to their small size.
鈥淛ust look at the Apostrophe Protection Society, who鈥檝e spent nearly a quarter of a century correcting street signs and trying to right the wrongs of a thousand rogue or missing apostrophes.
鈥淔ast-forward to the ChatGPT era, and the em-dash stands accused of being the number one giveaway of AI use in copy generation.
鈥淚鈥檓 not convinced of this. Sure, AI-produced text can be heavy on em-dashes, but humans like to use them too.
鈥淔or British audiences, American spellings and stilted phrasing are more reliable telltales. But even then, I don鈥檛 think many readers notice or care.
鈥淎t MQube, we use AI to produce around 70% of our marketing content completely autonomously 鈥� i.e. with humans involved chiefly in monitoring output and ensuring compliance.
鈥淔or social media content, marketing emails and most one-to-one interactions with customers, AI is more than capable of matching the tone, style and accuracy of a human.
鈥淲e鈥檝e designed and built something we call our TOVAAS 鈥� Tone of Voice Automation Assistant 鈥� to set the tone for how our AI talks to people, and ensure that what it writes is virtually indistinguishable from human-made content, while also being hyper-personalised and optimised.
鈥淔or me, the problem with badly AI-generated content isn鈥檛 punctuation, it鈥檚 superficiality. Content which bristles with citations and has a suspicious lack of grammatical mistakes, yet amounts to little more than a list of points with no insight isn鈥檛 just boring to read. In a marketing context it鈥檚 toxic, as it screams inauthenticity.
鈥淭he best marketing content builds a genuine connection with the audience. AI is superb at doing this in many ways, but using AI for longer-form content that requires creativity or empathy is hugely risky.
鈥淎t best, it demotivates your human staff and bores your audience. At worst, the whiff of AI 鈥榮lop鈥� will turn off your audience entirely. Why would people engage with a brand if they feel the brand can鈥檛 be bothered to engage with them personally?鈥�
Nathan Selby, Managing Director at Resultful
鈥淩emoving the em dash won鈥檛 make much difference to the untrained eye. Unless you鈥檙e working with AI personally on a regular basis, I don鈥檛 think you realise that the em dash has become synonymous with generative AI. End customers, for example, are unaffected by it. I think we鈥檙e still some way off AI talking like a human 鈥� there are plenty of additional phrases that will help you spot an AI-generated copy. For example, 鈥渋n a complex world鈥� (and words to that effect) 鈥� a typical give-away and personal pet hate!鈥�
Camden Woollven, Head of Strategy and Partnership Marketing at GRC International Group Ltd.
鈥淭he em dash was a lazy clue. It stood out because ChatGPT used it too often, not because it revealed anything structural about the text. Dropping it just removes one of the easier superstitions people used to lean on. It won鈥檛 change the fact that AI detection doesn鈥檛 really work.
鈥淪tyle can always be rewritten, and the tools that claim to tell human from machine fail once you run the text through a light paraphraser. The better question is what we expect to get out of detection at all. If the goal is trust in how AI is used, then teaching disclosure and responsible use gets you there faster than chasing new linguistic fingerprints that disappear as soon as models or habits shift.鈥�
Carolyn Shelby, Principal SEO at Yoast
鈥淎I never actually loved the em dash鈥搃t just inherited it. Because large language models are trained on mountains of spoken-word transcripts, where spoken pauses are often marked with dashes, they began using them as stand-ins for breath. The result was an unmistakable rhythm: clipped sentences, breathless pacing, and an excess of punctuation that made everything sound overly dramatic.
鈥淒ropping the em dash won鈥檛 suddenly make AI writing undetectable鈥搃t will just remove one of its loudest tells. The deeper giveaway has always been the cadence, not the punctuation. Machines have overindexed on speech-like rhythm because most of their 鈥渨riting鈥� comes from transcribed audio, not edited prose.
鈥淚nstead of playing punctuation whack-a-mole, we should focus on intentionality: teaching people how to write with rhythm, variety, and *restraint*. The goal isn鈥檛 to outsmart detection鈥搃t鈥檚 to make AI a better listener to human writing, not the other way around.鈥�
Zachary Cote, Executive Director at Thinking Nation
鈥淚t鈥檚 true, the em dash became a dreaded punctuation for me to come across. It screamed 鈥淎I.鈥� But doing away with it doesn鈥檛 solve any issues. AI is a mass homogenizer. When we let AI speak for us, we lose the quirks that make us human. We trade creativity for a pseudo-clarity.
鈥淲e need to champion a written culture that embraces AI in the brainstorming and outlining stages (after all, it is a helpful thought-partner to bounce ideas off of), while rewarding the flawed creativity of the human mind in the final drafts. If we must enable AI to commandeer some writing, let鈥檚 allow it to flourish in the hyper-formulaic worlds of contracts and, in other places, do our best to inspire our fellow humans with our own words. This is the balance worth striking.鈥�
听Nicole Franco, Head of Digital PR AI Innovation at Fractl听
鈥淐hatGPT鈥檚 reduced usage of the em dash will not mean much in the detection of AI-generated texts. The absence of the em dash as a distinctive style will not make AI-generated texts detectably difficult, at least because the em dash was never the main detection point. The usage of the em dash became popular among marketing and writing communities long before the emergence of ChatGPT. However, its excessive use in AI-generated texts has ignited a storm of discussions in the LinkedIn community, primarily due to style fatigue rather than being a full-on indicator of AI-text.
鈥淩ealistically speaking, AI detection software today isn鈥檛 infallible. It might be based on the rhythm of the sentence, the usage of a certain vocabulary, and/or the repetition of a sentence to get the message across. Still, in those instances when this occurs, there are false alarms about the AI-written piece and the detection of heavily AI-edited articles. The truth of the matter remains that, rather than fixating on who and what actually led to a specific piece of text, our goal must be the development of AI literacy and the substance of the content in general. This will enable each field to leverage the best aspects of AI.鈥�
Gailene Nelson, Senior Director of Product,听Turnitin
鈥淭he em dash was never a reliable way to spot AI-generated writing 鈥� it鈥檚 a stylistic choice, not a statistical signature. What matters most are the deeper patterns in how text is built. Humans write with intent, knowing what they are trying to say, while large language models (LLMs) generate text by predicting what words come next, based on the previous words it generated. This difference shows up in grammar and syntax, which are subtle but distinct patterns that AI detectors are designed to pick up.
鈥淭hat said, detectors shouldn鈥檛 be the sole judge of misuse. The true value of AI detection lies in sparking conversations, about a person鈥檚 writing, their learning journey, and how and where AI might have been used. The goal isn鈥檛 just to catch misuse; it鈥檚 to foster conversations on responsible AI use and why integrity matters. That鈥檚 how we move forward, helping people use AI responsibly, while keeping trust in the authenticity of the written word.鈥�
