https://medium.com/long-sweet-valuable/ill-instant...
创建于：2025年8月1日
思考
回答
What the article is getting at (in a nutshell)
Ossai Chinedum recounts a classroom anecdote: a young lecturer warns that he can spot AI-generated assignments, prompting one student to work from scratch while another plans to “just ask ChatGPT to make it sound more human.” Chinedum calls this a naïve strategy—he’s convinced savvy readers can still sniff out the tell-tale fingerprints of an LLM, no matter how many “humanise” prompts you layer on. (Medium)
Why experienced readers still catch you
| Common giveaway | Why it trips alarms | Example cues |
|---|---|---|
| Formulaic transitions & para starters | GPT models often lean on connective adverbs (“However,” “Moreover,” “Furthermore”) in a rhythm a human rarely sustains. (WeAreDevelopers) | Several consecutive paragraphs begin with near-identical sign-posts. |
| Recycled “AI vocabulary” | Certain high-frequency phrases crop up across many prompts—editors learn to recognise them. | “Treasure trove,” “Intricate tapestry,” “In conclusion,” etc. (YourTango) |
| Over-polite, under-personalised tone | ChatGPT defaults to safe, universal language that avoids specifics about lived experience. | Cover letters that gush about a company’s “innovative culture” without any personal anecdote. (WeAreDevelopers) |
| Punctuation quirks (the ‘ChatGPT hyphen’) | The model strongly prefers en- and em-dashes over simple hyphens—enough that typography geeks notice. (Talk Android) | Sentences peppered with “—” and “–” even where a short hyphen would do. |
Big picture: Automated detectors are unreliable, but humans who read a lot of AI text become remarkably accurate spotters. In lab studies, heavy LLM users correctly flagged 99 % of AI-written articles. (arXiv)
If you do rely on ChatGPT, how to minimise the footprints
-
Break the template
- Rewrite openings and closings from scratch.
- Use varied paragraph lengths and the occasional one-word sentence for rhythm.
-
Inject verifiable specifics
- Reference a date, location, data point, or personal anecdote only you would know.
- Cite a source or include an image of original work.
-
Run a “find & replace” pass
- Search for the cliché list above and prune mercilessly.
- Swap en/em dashes for plain hyphens where appropriate.
-
Let it sit, then read aloud
- Dead giveaways often surface when you hear the monotony.
-
Own your process where ethics require
- Academia and journalism increasingly expect disclosure. Hiding usage can breach honour codes—better to declare “AI-assisted editing” than risk accusations of misconduct.
Take-away
Chinedum’s piece is less a technical guide and more a wake-up call: simply asking ChatGPT to “sound human” won’t fool a reader attuned to AI cadence. If originality matters—whether to a professor, an editor, or a hiring manager—you still have to do the human work of drafting, fact-checking, and injecting real voice. Use the model as a co-writer, not a ghost-writer, and you’ll avoid most of the red flags above.