Article at-a-glance
– AI Detectors Are Unfairly Flagging Formal and Non-Western English: Whether it’s Nigerian English, academic writing, or just structured, formal language in general, AI detectors are labeling perfectly human writing as AI-generated.
– Formal Writing and Certain Vocabulary Are Red Flags: Words once considered normal in formal writing, like “delve” or “empower,” are now triggering false positives because AI models favor them, possibly due to them being trained at lower costs in third-world countries where a formal English writing style is more common.
– AI Detectors Are Biased Against Non-Native English Speakers: Non-native English speakers are being unfairly targeted, too, with their writing often mislabeled as AI for its more structured, neutral, slightly more formal style.
– AI Detectors Are Making Academics Dumb Down Their Work: To avoid false AI flags, students and scholars find themselves fretting over minor stylistic considerations.
– The World Risks Losing Great Content: The growing reliance on AI detectors could mean that timeless human creations – think The Little Match Girl or even the US Constitution – might be rejected as AI-generated.
From Turnitin to ZeroGPT, AI detectors are being used more and more to spot non-human language. While they do have legitimate applications – some of which we’ve explored in our previous blog posts – they also pose serious issues with their high rate of false results.
We’ve already seen that some political speeches and 19th-century literature were incorrectly identified as AI by one of the biggest AI detectors out there, so we started looking a bit closer into why that might be – and why more formal varieties of English, such as Indian and Nigerian, seem to be flagged as AI more frequently, too.
The Empowered Chicken Fostering The Enriched Egg: The Problem with AI Detectors and Formality
The same characteristics that make AI-generated text so convincing—structured sentences, formal vocabulary, and polished expression—are being flagged by many AI detectors as “suspicious.”
Essentially the more formal the language, the more likely it is that AI detectors will brand it as machine-made.
Words like “delve,” “dive,” “empower,” “enrich”, “harness,” “foster,” and “landscape” used to be perfectly normal choices in human writing. But now, they’re becoming red flags (and our ABBA tool is great at spotting many of them, by the way), and not because they’ve suddenly lost their place in everyday discourse, but simply because they’re the kinds of words AI models prefer.
And the AI models favor them because, in turn, they were trained on large volumes of written text, and those words occurred over and over in press releases, academic writing, opinion pieces, and so on – and they now occur over and over in AI written copy, to the point that everyone and their dog is tired of navigating complex issues in today’s challenging landscape.
It’s a bit of a catch-22: AI models favor formal words because they’ve learned them from pre-AI writing, but now those same words are enough to get legit high-value human copy flagged as AI-generated – and causing some, like Aktodotio CEO and co-founder Ankita Gupta, to adopt a radical editorial policy: “‘Delve, safeguard, robust, demystify, in this digital world.’ All ChatGPT. I am rejecting all content with any of these words.
Perplexity, Burstiness, & Some Non-Standard Struggles
Specific red-flag vocabulary aside, another issue of AI detection ethics comes from its bias against more neutral and predictable text.
AI detectors seem to be on the lookout for two stylistic markers: low “perplexity” (text predictability as defined by how likely the words are to follow each other; for an example of high perplexity, imagine meeting the Jabberwock with his vorpal sword in hand while reading about percussion drilling, motorized camera sliders, or gift ideas for funk music enthusiasts) and low “burstiness” (variation in sentence length and structure).
Formal or academic writing tends to exhibit both of these —structured, predictable sentences with very little deviation, and a reduced likelihood of errors. By contrast, more casual writing will vary its rhythm, mixing up sentence lengths and structures, and featuring comparatively more typos and other errors.
The Jabberwock is the main character in Lewis Carroll’s famous nonsense poem Jabberwocky – a perfect example of a high-perplexity text.
In fact, a study led by content strategist Dr. Anneke Schmidt found that AI detectors are prone to misclassifying more complex, formal, or technical writing – and she came to the worrying conclusion that “the only way to avoid false positives is by inserting mistakes and formatting issues, which is ridiculous!”
This is partly confirmed by our own research, which showed that the error-prone SurferSEO AI Humanizer fares better than the more conservative DeepL in making AI copy pass checkers, and that’s likely due mostly to the high rate of intentional errors and sloppier writing.
In academic contexts, where formal language is the standard, this misidentification is frustrating to many. Students and scholars report having to “dumb down” their writing in order to avoid being flagged by detection tools, making their content worse in the process – and causing them to waste time worrying about linguistic issues not directly related to their research.
A $2 Per Hour Problem
Moreover, this bias against more formal, error-free work is disproportionately affecting non-native English speakers.
As they create content with less perplexity, they have their copy incorrectly flagged as AI more frequently – which is why James Zou, professor of biomedical data science at Stanford University, advises against relying on AI detectors in educational settings.
Regional varieties of English are affected, too, with Nigerian English potentially more frequently labeled as AI. Farida Adamu, a graduate student from Nigeria, had to rewrite her thesis to ensure it would pass through AI detection software, essentially altering her own style to avoid being flagged.
$2 per hour – that’s how much Amazon and OpenAI paid millions of AI training workers (writers? assistants?) in Africa, Asia, and South America. Today, writers from the same parts of the world find their perfectly human writing incorrectly labeled as AI.
(had the training been done mostly on the UK and US youth, for example, we might have seen AI red flags pop up around stuff like “mate” and “bro” and “chill” and incomplete verbless sentences – but we don’t, because money.)
So today academics, non-native speakers, and pretty much everyone else who has to follow any formality standard face a doubly artificial dilemma: they need to introduce some artifice in their natural writing to avoid having their content flagged as artificial.
So Ultimately, Are We Losing Great Content?
We might just be.
Had it been written today, Hans Christian Andersen’s The Little Match Girl might have been rejected for being “likely AI”.
The US Constitution itself might have needed to go through several rounds of edits to lower that abysmal 97.97% AI likelihood.
Previously creative (and handsomely paid) copywriters are now juggling AI editing gigs for pennies, fixing boring ChatGPT output instead of coming up with fresh new creations of their own.
On a day-to-day basis, non-native English speakers, academics, and native English speakers from beyond the Western English dialect regions deal with AI detectors flagging their perfectly human work as “fake”.
Does that mean AI detectors are all bad? Not necessarily – they do have some limited use, and for some use cases they might just be better than a coin toss (but not by much).
But the final decisions in terms of content quality shouldn’t be based on an arbitrary AI score.
At the end of the day, AI detectors should support good writing, not penalize it. They’ve got a long way to go before they truly understand what makes writing human.
Until then, it’s up to us to keep calling out these flaws and making sure the machines don’t get the final say on what counts as great content.
AmpiFire can help you drive more visibility to your business with quality content development and distribution – get in touch today to see what we can do for you!
