No. You can't tell it was written by AI

Preface

Yes. This is another AI post. It’s kinda hard to ignore.

I have seen people casually claim that a post, text, or an article was AI-written, and they’re usually wrong. Of course, not always. I am not speaking to the cases where you know the author and their style, or you have seen the content somewhere else, or the content is straight up gibberish or nonsensical.

I am speaking about the cases where people assert that certain “tells” people have that make them know that a document or email was written by AI.

In this essay, I will argue that, your favourite “tells” that a document was produced by AI, at best, is wrong, and depending on your position, in life, at worst, is dangerous and harmful.

This essay is 100% written by me. I did not use AI for anything. Not even for grammar or spelling check, formatting or tone adjustments.
You will see some ideas and claims that I make that other people have also made. These arguments are definitely not unique to me, so I am not claiming to be the first to be making them.
I will intentionally use some of the “tells” I have seen people claim leads them to conclude it’s AI, and I will try to reference the related literary and rhetorical device, if any.

Ready? Yeah, let’s delve into it.

AI Did Not Just Invent Styles. Humans Used It First ¹

The first argument is, any pattern you think you have noticed from AI usage was first used by humans.
This is an undebatable fact. These AIs did not just materialise. They were trained on original human content.

AI Has Elements of the Stylistic Choices of the People Who Trained It ²

My second argument is that the styles we see with AI, reflect the stylistic choices of the people who did training and annotation.
The development of AI had many phases. Many of those phases involved grunt work of people—humans ³, manually inputing data, annotating it, and scoring the AI outputs.
So who trained it? A lot of the early training, data annotations and other manual processes, happened with cheap labour in African countries. There are multiple sources that have revealed the hidden economy of workers that big-tech outsources these kinds of tasks to African countries with unstable political situations, weaker workers rights, and cheap labour.

How do I know AI adapted their styles? I will tell you.⁴

You see, I am Nigerian, and we are known for our flowery, flamboyant, and persuasory style of speaking and writing. It is very likely that this is a trait from our traditional languages that filtered into our use of English language.
I remember Paul Graham saying that “delve” is an indicator that an email was composed with AI.

Delve is a staple word in Nigeria. This word has been around all my life - from primary school, it’s like an everyday word. And you can see how Nigerians reacted to Paul Graham’s assertion. Their reactions on Twitter resonate with me.

I use Nigeria in this context, because I am Nigerian, but you can easily replace it with any other African country or Asian country—basically people who had to learn English as a second language.

A lot of people that learned English and speak more than one language, did so by reading books, consuming magazines, news, etc., and that exposes them to atypical, obscure or big-sounding words like “delve”, that apparently, Americans or other “native speakers” don’t use.

I think it is clear that the stylistic choices of these AIs are influenced heavily from the stylistic choices of the people that big tech contracted to do the training.

AI Detection Tools Are Unreliable

Any AI detector that tells you it can reliably detect AI usage in text is very likely lying to you. Even OpenAI’s detector was shut down because it was not found to be reliable.

MIT also affirmed that AI detectors simply do not work. We should be thinking of how to evolve our pedagogic processes, rather than trying to sniff out who is using AI or not.

And Why Is It Bad To Guess?

I hope that by now, you are starting to agree with me, that there is no objective basis for being able to infer that a document was written with AI.

Unless you personally know the person writing the content and are familiar with their natural style, any intuition you may think you’ve developed for AI giveaways is almost certainly inaccurate.

The reason why this is problematic is that depending on your position, it could range from harmless to harmful really quickly.

In the best scenario, it’s just a harmless wrong assertion. In a worse scenario, if you are a person in a position of power—say an investor, a recruiter, a professor, a manager, etc., your attempt or your intuition to classify a text as AI-generated is very likely to have real life consequences on that person. What’s worse is that you do that with no provable or reliable reasoning. It is just wrong. And, you stand a risk of being biased against a huge population of people on earth.

Wrapping up

I feel very strongly about this, obviously, because of my cultural background and because I have seen people with my background being accused of cheating, because an AI detector told someone so.

At almost every stage in my life, I have had to prove that I can speak English language, despite being from a country colonised by the Brits, and doing all of my academic endeavours in English. I am still forced to take IELTS or TOEFL, almost everywhere. If I am applying to study (in English) in GERMANY, they will ask me to prove to them that I can speak English, but to apply for a student visa, I will need to translate my English documents to German. Hello??

As found out in this study from Stanford, non-native English speakers are more likely to be accused of using AI (full study is here), because of a combination of the reasons I have highlighted in this essay and it all feels like a round of punishment.

First, they made us learn English. Then we did it so well, that we started to use words that are uncommon to them. Then, they had us do the grunt work of training the AI models. Next, the AI models started sounding like us, and now we are being accused that we are cheating.⁵

I have further thoughts on the demographics that institutions consider to be native speakers, but, I won’t get into that in this post. It seems to me to be more of a citizenship filter than an actual ability to use the language. A quick look at who is considered to be native English speakers according to the Teaching English as a Foreign Language (TEFL) industry will tell you more.

I will not sit back and shy away from using good writing and speaking skills, in an attempt to not sound like AI.

I will always resist and argue against any attempt to extrapolate an AI tell, and I encourage you to join me.

Thank you for reading.

Footnotes