Do you remember the Looney Tunes? In just about every episode, Wile E. Coyote runs off the cliff and keeps right on running. The ground is gone, the legs keep turning, and nothing bad happens for a good three seconds. Then he looks down. The looking triggers his uncontrolled descent, not gravity nor the cliff.
Every time I read an AI-assisted draft, this comes to mind because there’s a word in most of them doing the same thing. It carries you across a paragraph on pure confidence, and there is nothing underneath it. You just haven’t looked down yet.
Over the past year, I’ve been asked to review a fair number of writing and content platforms. The demos start to blend together, and somewhere in the middle of most of them, the same question comes up, which is what the software ought to be catching on its own. Most demos land on the em dash and the junk words, the vocabulary that screams AI bot.
I get the appeal. Those are the AI tells everybody knows, easy to explain, easy to find, and easy to strip out. They’re also the two with the least shelf life left. The Economist just ran the largest comparison of human and machine prose anybody has published, 55,940 sentences across 1.2 million words, and retired the em dash outright. Only Claude used them more often than human writers do.[1] Unless you’re Mark Twain, but hold that thought. Most write-ups treated that as the headline but the more useful tidbit sat lower in the report: when these models join two ideas together, the word they reach for more than any other is “and.”
Every transition makes a claim
Start with that “and.” The em dash got all the attention because it was easy to spot, and prior to gen AI, many writers didn’t use it. “And” is harder to catch because it looks innocent enough, but it is often doing the same job as the dreaded em dash now. It joins two ideas with the weakest connector English offers, because “and” asserts almost nothing about how the two sentence halves relate. And so, every other transition asserts something specific (do you see what I did there?).
“However” claims the next sentence pushes against the last one.
“As a result” claims the first thing caused the second.
“Moreover” claims the second point extends the first rather than repeating it.
“Importantly” claims this sentence outranks its neighbors.
Not one of those words describes anything. Every one of them makes a claim, and you take it on blind faith, because checking requires cognitive effort.
So, why does this happen? Sometimes the model reaches for “and” because it has not decided how two ideas relate. Other times, it reaches for a stronger transition like “however” or “as a result” and gets the relationship wrong.
Both mistakes start in the same place. Human writers usually see the relationship before they name it. The second sentence pushes against the first, so “however” earns its spot. Well, a model works the other way around. The sentence needs a connector, the training data says one belongs there, and the word appears whether or not anything underneath it holds. That is the coyote, three strides past the edge. That is an unearned transition, and it comes in four flavors.
The false contrast: “Our platform connects to your existing systems. However, it also supports custom workflows.” Two perfectly compatible features, sitting on either side of a word that promises tension. Nothing pushes against anything.
The false consequence: “Buyers now research independently. As a result, your content has to work harder.” The second sentence might be true. It does not follow from the first without two or three premises nobody bothered to write down.
The false addition: “Moreover,” introducing a sentence that restates the previous one in fresh vocabulary. The word promises a second point and hands you the first one again.
The false emphasis: “Importantly, the underlying data has to be accurate.” Important compared to what? Nothing else in the paragraph got marked less important, so the word ranks nothing.
None of this is an English problem. A 2026 study in iScience compared Portuguese news articles against versions from GPT-4o, Mistral Large, and Llama 3.3, and found the machine drafts reaching for formal connectives like além disso, the Portuguese “furthermore,” where human writers used conversational ones.[2] Same slot, same reflex, different language.
Every other tell I gave you has expired
Unearned transitions are worth learning because the rest of my back catalog is not. Go back and read the red flags post that opened this series.[3] The junk words are all there: “seamlessly”, “robust,” and “game-changing”. Two years ago I wrote that the moment I saw “delve” in an article I was done reading it, and I meant it.[4] That was a great rule for about eight months. Then every list like it got published and indexed, mine included, the models read the criticism right along with everything else, and today “delve” tells you nothing about who wrote the paragraph in front of you.
Punctuation went the same way, only faster, and it was a shaky signal even before the models moved.[5] A 2026 preprint clocked GPT-4.1 at 10.62 em dashes per thousand words against a modern human baseline of 3.23, which sounds damning until you remember how uneven human punctuation has always been.[6] SlopDetector, a vendor blog, ran a useful but lighter-weight check against public-domain authors and put Huckleberry Finn at 10.13. Jane Austen scored a flat zero.[7] So the rough test convicts Twain, clears Austen, and tells you nothing about the draft sitting in your inbox. Anything a regular expression can find is also something a vendor can strip, and the market is full of tools that will strip it for you.
Which is the fair objection to this whole series, and the one that kept me from writing another entry about AI slop for ten months. LLMs evolve and every tell eventually stops working. So why learn one? Because this one is special. Removing the “ands” would mean the model doing the one thing it cannot do – which is decide what your argument is.
Delete the word and read it again
Checking a suspect transition takes about five seconds. Cut the transition, read the two sentences cold, and see what happens. If the relationship still holds, put the word back. If nothing changes, the word was a bedazzler. More often, the missing transition exposes the real problem: the argument you thought you were making never made it to the page.
I caught this recently in an AI-assisted draft where the paragraph described a team tightening its content review process. The next sentence opened with “as a result” and jumped straight to buyers trusting the content more. Maybe they did, but the paragraph had not earned it. There was no buyer quote, no performance signal, and no step connecting internal process to external trust. Once the transition disappeared, the paragraph told the truth. One operational change sat next to one hoped-for outcome, and the missing argument suddenly became visible.
The second check takes fifteen seconds and works on a whole draft. Read only the first sentence of every paragraph, in order, and skip everything else. Machine drafts announce their turns at the top of paragraphs, so you get “Additionally,” “That said,” and “Importantly” stacked up like highway signage, while human writers bury the connective mid-sentence or drop it and trust you to keep up. When more than half your paragraphs open with a transition word, you are reading structure that was generated rather than argued, whoever typed it.
No tool is coming to save you
Writing software splits cleanly into what can be linted and what takes judgment. Banned terms, product names, punctuation, and sentence length can all be checked by rules, which is why a decent rules engine nails them and why every vendor leads with them. There is even a good theory for why the em dash was catchable at all. That same 2026 preprint argues it is markdown formatting leaking into prose, the last visible trace of the structural training these models absorb, and that it survives being told to stop.[8] Structural residue is exactly what a linter is built for.
An unearned transition is not a word error, so none of that machinery touches it. Judging whether “as a result” is honest means knowing what the piece claims, what it has already established, and whether the second sentence follows from the first. A person has to read for that, and nobody has automated reading. This is also why this AI tell walks straight through the humanizer tools I wrote about last fall.[9] Those products vary your sentence lengths, swap your vocabulary, and scatter your punctuation, because those are the levers they have. None of them can go back and build the argument that should have been under the connective.
Earn your transitions
Settle the argument before you draft anything. A model can only fake a connector when the logic was never decided, so an outline that fixes what follows from what kills the whole problem at the source. My own Claude setup front-loads the thinking into files rather than asking for prose on the first pass for exactly this reason.[10]
Then have it draft with no transitions at all. Tell it to write the sections and connect nothing, then add every connector yourself afterward. It feels clumsy for about ten minutes, and then it becomes the fastest edit in your process, because each one you install is a decision you made and can defend.
Build your own flag list instead of downloading someone else’s, mine included. Every list that gets published gets trained on, so the only list worth keeping is the private one you grow from your own edits, in your own voice. Mine picks up a few entries a month, and almost none of them would help you. Then write from material a model cannot generate, which means first-hand experience, original data, or a number you measured yourself. When events happen in order, the connectives between them are true by construction.
Look down
Reading for unearned transitions has wrecked how I read everything, including work written entirely by people. Plenty of human writing has the same problem, because a writer in a hurry reaches for “however” out of rhythm rather than logic, and the machines picked up the habit from us in the first place. What changed is the volume, and what it costs you to check.
So keep one standard in your head. A transition is a claim, and if you cannot defend the claim, delete the word. The sentences survive without it, and the hole where it used to be shows you exactly where the argument still needs building. Mark Twain had the better line: the difference between almost right and right is “really a large matter—it’s the difference between the lightning bug and the lightning.”[11] Your readers are already three strides past the edge with you. Look down before they do.
If you want more of this, subscribe to the newsletter. It goes out most weeks and it is written by a person.
Frequently asked questions
What is an unearned transition in AI writing?
An unearned transition is a connector that claims a relationship the surrounding sentences have not established. “However” claims contrast, “as a result” claims cause, “moreover” claims addition, and “importantly” claims priority. When the logic underneath the word is missing, the transition creates a feeling of structure without proving that the argument connects.
Why is “and” becoming a more useful AI-writing tell than the em dash?
The em dash became easy to spot, easy to discuss, and easy for vendors to strip out. “And” is harder because it looks harmless. The Economist’s large comparison of human and machine prose found that when models join two ideas, they often reach for “and.” The useful signal is whether “and” hides an undecided relationship between two ideas.[12]
How can you check whether a transition is earned?
Cut the transition and read the two sentences without it. If the relationship still holds, the word was doing useful work and can go back in. If nothing changes, the word was decoration. If the paragraph suddenly feels weaker or less logical, the transition was probably covering for an argument that never made it onto the page.
Why can’t AI detectors or humanizer tools catch unearned transitions?
AI detectors and humanizer tools are built for enumerable surface patterns: punctuation, banned words, sentence length, vocabulary swaps, and similar signals. An unearned transition requires a different judgment. Someone has to understand what the piece has already claimed and whether the next sentence follows from it. A person has to read for the logic instead of counting surface signals.
Should writers remove transitions from AI-assisted drafts?
Writers do not need to remove every transition from an AI-assisted draft. They should make every transition defend its job. One practical approach is to draft sections with fewer connectors, then add the transitions manually after the argument is clear. Each connector should name a relationship the writer can explain: contrast, cause, addition, emphasis, sequence, or some other specific link.
About David Sweenor
David Sweenor is a Top 25 AI thought leader, author, and founder of TinyTechGuides. He spent the first half of his career as a data practitioner at IBM working in data science, business intelligence, and data warehousing, and the second half in product marketing leadership at SAS, Dell, Quest, TIBCO, Alteryx, and Alation. His writing focuses on the practical intersection of AI, analytics, and B2B marketing.
Books
- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business
- Generative AI Business Applications
- The Generative AI Practitioner’s Guide
- The CIO’s Guide to Adopting Generative AI
Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.
[1]The Economist. “How to Spot AI Writing.” The Economist, July 30, 2026. https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing. Findings summarized in Cramer, Jude. “Forget em dashes: A viral report on AI-generated writing has surprising new clues.” Fast Company, August 2026. https://www.fastcompany.com/91584243/how-to-identify-ai-generated-writing-viral-report-has-surprising-new-clues-economist.
[2]Rodrigues, Flávia A., Niclas F. Sturm, and Flávio L. Pinheiro. “A linguistic comparison between human- and AI-generated content.” iScience, 2026. https://doi.org/10.1016/j.isci.2026.114976.
[3]Sweenor, David. “How to Spot AI-Generated Content: Red Flags Every Marketer Must Know.” TinyTechGuides, August 30, 2025. https://tinytechguides.com/blog/how-to-spot-ai-generated-content-red-flags-every-marketer-must-know/
[4]Sweenor, David. “Spotting AI junk words: Why AI still can’t write like humans.” TinyTechGuides, November 11, 2024. https://tinytechguides.com/blog/spotting-ai-junk-words-why-ai-still-cant-write-like-humans/
[5]Sweenor, David. “Punctuation Pandemonium: When AI Content Goes Wild.” TinyTechGuides, September 6, 2025. https://tinytechguides.com/blog/punctuation-pandemonium-when-ai-content-goes-wild/
[6]Freeburg, E. M. “The Last Fingerprint: How Markdown Training Shapes LLM Prose.” arXiv preprint, 2026. https://arxiv.org/pdf/2603.27006. Preprint, not yet peer reviewed.
[7]SlopDetector. “Is the Em Dash an AI Tell? We Measured Dash Density Across Human vs AI Texts.” 2026. https://slopdetector.org/blog/em-dash-ai-tell-data. Useful as supplementary color, not primary research.
[8]Freeburg, E. M. “The Last Fingerprint: How Markdown Training Shapes LLM Prose.” arXiv preprint, 2026. https://arxiv.org/pdf/2603.27006. Preprint, not yet peer reviewed.
[9]Sweenor, David. “How to Spot AI Content: The Humanizer Trap Destroying Your Writing.” TinyTechGuides, September 20, 2025. https://tinytechguides.com/blog/how-to-spot-ai-content-the-humanizer-trap-destroying-your-writing/
[10]Sweenor, David. “Is your Claude marketing OS a little quirky?” TinyTechGuides, May 13, 2026. https://tinytechguides.com/blog/four-components-claude-stack/
[11]Mark Twain, letter to George Bainton, October 15, 1888, quoted in Barbara Schmidt, “Mark Twain Quotations: Word,” TwainQuotes.com. https://www.twainquotes.com/Word.html.
[12]The Economist. “How to Spot AI Writing.” The Economist, July 30, 2026. https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing. Findings summarized in Cramer, Jude. “Forget em dashes: A viral report on AI-generated writing has surprising new clues.” Fast Company, August 2026. https://www.fastcompany.com/91584243/how-to-identify-ai-generated-writing-viral-report-has-surprising-new-clues-economist.