The Beige Plague

There is a voice of writing that you- at least we writers- recognise instantly now; it appears in LinkedIn posts, student essays, marketing emails, blog drafts, the list goes on. People might have been oblivious to this voice in the early days of 2023, but by 2025, if one used this, they would often be up on Reddit, made fun of.

Every individual has their own opinion or a way of determining whether a certain piece of text is AI. Those who have never used an em-dash by themselves might say that it is a giveaway, to which we strongly disagree; once you know how to use an em-dash, it's irresistible to use it over and over, unless, of course, you are a sociopath or simply afraid to be detected as an Al. The question is not whether Al-generated writing is detectable. The question is why it all converges on the same statistical signature, regardless of prompt, model, or intended audience. The answer lies not in the models themselves, but in what they were trained on.

GPT-3 and GPT-4 were trained on text scraped from the internet up until September 2021 (illegally, perhaps, but let's not be the judge). That set of training data represents a specific era of online writings, and that era had a certain voice to it. Between 2010 and henceforth, the internet had a certain bias into a voice; it was not too formal, not too casual, not too opinionated and helpful.

Founded in 2012, Medium became the platform for "thought leadership". By 2021, it had millions of articles, nearly all of which followed the same template: personal anecdote, universal insight, three-point structure, and a rather optimistic conclusion. The overall style was conversational, but polished with the help of grammar tools that cut off run-on sentences, commas, and other imperfections of being human. Then came the SEO optimisation era. Somewhere between 2015 and 2020, corporate blogs rose to the stage, with one clear goal in mind: rank well on Google.

Third, we have Wikipedia. Heavily edited and proofread to be neutral and encyclopedia-like, heavy with citations. Fourth came the Reddit (which, we still believe, shaped the personality of Claude). The karma system functioned as a massive RLHF dataset, training humans (and thus Al) to write in ways that got upvoted: clear, balanced, slightly friendly, not-so-offensive. Style guides from news aggregators like AP and Reuters defined online journalism. Short sentences, active voice, avoid adverbs. This is possibly where the idea of good writing being lean, factual, and unadorned was reinforced, and thus, we have Al models like Gemini provide us with good writing that has no soul.

So the final training dataset of most models was not exactly "human" writing. Written by humans, yes, but they were severely edited, optimized, and published to the point that there is no explicit "humanness" to it. What this tells us is a simple truth: When you prompt an LLM to write, you are essentially asking it to generate text that matches the statistical signature of "good writing" in its training data.

The reason for this myth is indeed due to OpenAI's ChatGPT being obsessed with the usage of em-dashes; this was so intense that putting a restraint on em-dashes became a feature for the upgrades of some GPTs. The model, obviously, learned this from us humans. Because the humans who wrote the training data used them constantly in a specific context: casual-but-authoritative online prose, which is the exact tone many LLMs are aiming for.

• "However," as a sentence opener (overrepresented in academic writing)
• "It's worth noting that..." (a hedge phrase common in moderated forums)
• "In other words..." (explanation culture, ubiquitous in educational content)
• Three-item lists (blog post structure, everywhere)
• Rhetorical questions followed by answers (Medium's favorite trick)

None of these are Al signatures. They are 2010s online writing signatures. The beige plague is not a bug; it never was. Rather, it's a feature of training on edited text. If you want a model that writes like a human... that gets complicated. Not more data. Not better prompts. A different kind of data. The kind that includes the errors, the false starts, the unpolished thinking that happens before the edit pass.

End of entry.