The Myth of "Moreover"

There is a word that has recently acquired a strange reputation.

Moreover. A perfectly ordinary English conjunction, used in formal writing for centuries, appearing in Montaigne and Hume and every IELTS preparation booklet ever printed. Sometime in 2023, the internet decided it was a fingerprint. A giveaway. The telltale mark of a machine.

The logic, if you can call it that, goes roughly like this: AI systems use moreover frequently. Therefore, if you see moreover, the text was probably written by AI. TikTok videos explain it with the confidence of people who have just discovered something. Reddit threads catalog the suspicious words. University forums warn students. The myth spreads because it feels true — it has the texture of an insight, the satisfying click of a pattern recognized.

It is also almost entirely wrong. And the ways it is wrong reveal something genuinely interesting about language, about AI, and about the very human desire to reduce complex problems to simple rules.

A Word Is Not a Signature

Moreover is not an AI word. It is an academic word. The distinction matters enormously.

English has a tier of vocabulary — sometimes called the Academic Word List — that appears disproportionately in formal writing across every domain: scholarly journals, legal briefs, philosophical essays, standardized tests. Words like furthermore, consequently, nevertheless, notwithstanding. These words exist because formal writing has structural needs that casual speech does not. When you need to add a point that reinforces a previous one, you reach for moreover. When you need to signal a concession, you reach for however. This is not AI behavior. It is formal register behavior, and it predates computers by several centuries.

The reason AI systems use these words frequently is not that they were programmed to. It is that they were trained on enormous quantities of text that includes an enormous quantity of formal writing — and formal writing uses these words. The model learned the statistical patterns of the genre. When it produces formal text, it produces formal-register vocabulary, because that is what formal text contains.

Blaming moreover for identifying AI is like blaming a suit for identifying a job interview. The suit is correlated with the event, but it did not cause it, and plenty of people wear suits for other reasons entirely.

What Actually Happens in AI-Generated Text

The honest version of the AI-detection question is harder than the viral version, and more interesting.

There are real patterns in machine-generated text — they are just not where most people look. A language model produces text by predicting, at each position, the most statistically likely continuation given everything that came before. This process has a signature, but the signature is structural rather than lexical. It shows up in rhythm and proportion more than in vocabulary.

Consider paragraph length. Human writers — especially in informal or creative contexts — vary their paragraph length erratically. A long paragraph followed by a single sentence. Then three medium ones. Then a fragment. The variation is partly intentional, partly unconscious, shaped by how the thought arrived rather than by any plan. AI-generated text tends toward more uniform paragraph length, a more consistent level of elaboration, a more regular cadence of claim and explanation. Not always. Not absolutely. But often enough to be noticeable to a careful reader.

Consider also what might be called the absence of friction. Human writing is full of evidence that a mind was working — hesitations crystallized into qualifications, ideas that were almost abandoned and then rescued, sentences that change direction midway because the thought changed direction midway. AI writing tends to be smooth in a way that is, paradoxically, slightly uncanny. Every paragraph resolves. Every claim is supported. Every transition is clean. It is writing that has never struggled, and the absence of struggle is perceptible.

And then there is the question of specificity. Human writers, especially when writing about things they know, include details that could only have come from somewhere — a particular number, an odd comparison, a specific case that was not the obvious example to reach for. AI systems tend toward the representative example, the canonical illustration, the detail that is correct but unsurprising. The writing is accurate without being particular.

Why Detection Tools Are Dangerous

The problem with all of this — and it is a serious problem — is that none of these patterns are reliable identifiers. They are tendencies, not rules. And the tools that claim to formalize them into scores and probabilities have a track record that should make anyone cautious about acting on their output.

AI detection tools produce false positives at rates that are deeply uncomfortable for any context where the stakes are real. Non-native English speakers are disproportionately flagged, because the slightly formal, slightly careful prose of someone writing in their second language resembles, statistically, the slightly formal, slightly careful prose of a language model. Academic writers are flagged. Students who were taught to write in structured, thesis-driven paragraphs are flagged. In 2023, multiple university cases became public in which students faced academic misconduct proceedings based on detector output — in at least some cases, for work they had genuinely written themselves.

This is not a minor calibration issue. If a detection tool cannot distinguish between a Ghanaian graduate student writing careful formal English and a language model, then the tool is not measuring what it claims to measure. It is measuring something correlated with AI output that is also correlated with a great many other things, including being a careful writer who learned English formally.

The irony worth sitting with: the better a human writer is at formal English, the more likely certain tools are to flag them. The myth of moreover and its cousins punishes exactly the kind of careful, structured writing that education is supposed to produce.

The Harder Question

Underneath the detection question is a philosophical one that rarely gets asked directly: what are we actually trying to identify?

The naive version of the question is authorship — did a human or a machine produce this text? But the naive version does not describe what actually happens when people use AI in writing. Consider the range. Someone pastes a prompt into ChatGPT, copies the output, and submits it unchanged. Someone uses ChatGPT to generate an outline, then writes every sentence themselves. Someone drafts a paragraph, asks an AI to improve the grammar, and revises the result. Someone reads AI-generated text about a topic, closes the tab, and writes their own version from memory. Someone uses Grammarly, which makes statistical suggestions about word choice and sentence structure, which is a form of AI assistance that nobody seems to find controversial.

These cases are not the same. They involve different amounts of human thought, different degrees of intellectual engagement, different relationships between the tool and the mind using it. Treating them as equivalent — all "AI writing," all equally problematic — is a category error that produces policy worse than no policy at all.

The meaningful question is not whether AI was involved. It is whether a human mind was genuinely working. Did someone think through the problem, or did they outsource the thinking? Did someone develop an understanding of the material, or did they launder generated text into submitted work? These questions cannot be answered by running text through a detector. They require knowing what the work was supposed to accomplish and whether the person can demonstrate that they accomplished it.

What Language Looks Like Now

There is something else worth acknowledging, which most discussions of AI detection quietly avoid.

Language is changing because of AI. This is not a crisis — language changes continuously, shaped by the tools and technologies through which people produce it. The typewriter changed written style. Email changed epistolary conventions. Autocorrect changed spelling habits and introduced its own characteristic errors. AI writing tools are changing the texture of written language in ways that will take years to fully perceive.

One effect is already visible: the vocabulary and structural conventions of AI-generated text are diffusing into human writing. People who read a great deal of AI output — which now includes most people who use the internet regularly — absorb its patterns. Moreover may actually become more common in human writing, not because humans are becoming AI, but because language always reflects its environment.

This means that even the imperfect statistical signals that currently distinguish AI text from human text will gradually erode. Future models will be better at producing the friction, the specificity, the rhythmic irregularity that currently marks human writing. Future humans will write in ways that increasingly resemble current AI output, because they were formed in an environment saturated with it. The gap will narrow from both sides simultaneously.

Which returns us, finally, to moreover.

The word is not a fingerprint. It never was. What made people think it was a fingerprint was a real observation — AI writing does have patterns — filtered through the very human tendency to want a simple rule for a complicated problem. The simple rule is wrong. The complicated problem remains.

The question worth asking about any piece of writing is not whether it contains suspicious words. It is whether there is a mind visible in it — a perspective, a particular way of seeing, evidence that someone was thinking rather than generating. That question is harder to answer and harder to automate.

Which is, perhaps, precisely the point.

EisatoponAI

An independent intellectual publication exploring mathematics, AI, science, paradoxes, and the hidden structures behind reality.