AI tell · grade E
"In conclusion," "furthermore," "moreover": the AI cliché with no data behind it
"In conclusion," "furthermore," and "moreover" are the phrases most people name first when asked what AI writing sounds like. We searched all twelve models in our own corpus for that exact stack, closed connectives forcing a five-paragraph shape, and found it in zero of them. The pattern is real, but our corpus can't confirm where it started: the common account traces it to GPT-3.5's 2023 output, layered onto a classroom essay format that predates that model by decades. Treating it as a live AI signal in 2026 mostly catches human writers, students especially, who learned the same structure in school.
What it is
The closed-connective stack is the transition set taught for formal five-paragraph essays: an opening thesis, body paragraphs each launched by "furthermore" or "moreover" or "additionally," and a closing paragraph that restates the thesis under the heading "in conclusion." The common account is that the early public version of ChatGPT, built on GPT-3.5 through most of 2023, defaulted to almost exactly this shape when asked for anything essay-length. We didn't log that model ourselves, so we can't confirm the rate, but that early exposure is the likely source of the popular association.
How often it shows up
We checked all 12 models in our corpus for "Furthermore," "Moreover," "Additionally," and "In conclusion" used as connective stacking (research/reports/ai-tells-empirical.json, entry emp-009) and found zero occurrences across the full logged set. The logged corpus isn't even across models: most are 5 conversations, one is 8, and the largest, deepseek-v4-flash-nothink, is 27 of 40 logged conversations (13 dropped as unrepresentative, template placeholders or persona roleplay). Every model showed the opposite pattern from the folk image: no essay-mill connectives at all, in any conversation we logged for any of them. underslop's linter does not score this construction, because scoring it would mean scoring a phrase our own data says current models rarely produce.
Why readers flag it
Readers flag it because it is the version of AI writing they learned first. GPT-3.5's public release in late 2022 is commonly described as writing close to this pattern by default, and millions of people saw that output before any other model shaped their expectations. We have no logs of GPT-3.5 in our own corpus to check that description against. The association calcified into common knowledge before the underlying models moved on, and a reader who spots "in conclusion" next to a restated thesis is pattern-matching against a specific, dated model rather than against anything the 12 current models in our corpus tend to write, which is why the instinct now misfires more often than it hits.
The fix
If a stacked "furthermore... moreover... additionally..." run reads stiff, cut it because it is stiff, not as proof of who wrote it. State the last point directly and stop; a document rarely needs to announce that it is concluding. If you are a student or a non-native English writer taught this structure as the correct one for formal writing, do not strip it out of fear. Our own search found no support for treating it as an AI marker, so removing correct, taught prose to dodge a false accusation is a bad trade, and the accusation is the thing that is wrong.
Examples
In conclusion, remote work offers many benefits. Furthermore, it increases flexibility for employees. Moreover, it can reduce overhead costs for companies. Additionally, studies have shown that productivity often improves. In conclusion, remote work is a trend worth embracing.
Remote work brings real benefits. It gives employees more flexibility and can lower a company's overhead. Studies also show productivity often improves. It is a trend worth embracing.
Moreover, the data indicates a clear upward trend. Furthermore, this trend has continued for several years running. Additionally, it is important to note that other factors may be at play. In conclusion, the evidence strongly supports this conclusion.
The data shows a clear upward trend that has continued for several years running. Other factors may also be at play, but the evidence strongly supports this conclusion.
Questions
Why is ChatGPT so cliché?
Mostly because the version most people used first, GPT-3.5 in 2022 and 2023, is commonly described as defaulting to a formal essay register with stock transitions and a restated-thesis closer. We didn't log that model ourselves, but that is the version most people mean when they say "cliché." The 12 current models in our own corpus write in a shorter, more conversational register, and the older association is what stuck.
Why does ChatGPT write so corny?
The corny read usually comes from a different source: validation openers that praise the question, which showed up in 12 of the 12 models we logged. Cheerleading closers like "You've got this" or "Good luck" added to it in 10 of 12 models, and handing the reply back with a forced-choice follow-up or an offer to keep helping showed up in 9 of 12. The formal closed-connective stack, "furthermore," "moreover," showed up in none of them.
How do I fix clichés like "in conclusion" in my own writing?
Read the sentence without the transition word and check what it lost. Most of the time "furthermore" and "moreover" add rhythm and nothing else, so cutting them tightens the sentence. Do this because it reads better, not because you're worried about being flagged.
Is "in conclusion" actually an AI tell?
Not one we can support with our data. We searched 12 models specifically for it and found it in none of them, which is why underslop's linter doesn't score it. Flagging it as AI-written risks penalizing a real person who was taught the five-paragraph essay format, which is most people who went through an English class.
If not this, what does underslop actually score as an AI tell?
Patterns we can show directly in our own model logs: validation openers that praise the prompt, chains of two or more em dashes doing all the joining work, bold-colon bullet lists standing in for prose, and claims propped up on an unnamed "study" or "expert." Those are common in current output. The closed-connective stack is not.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.