It matters if AI "wrote" it

I’m one of those calling the AI checkers a “scam”, and for personal reasons.

My home country was colonized by Britain, and english is our first language, but apart from that, africans LOVE “big grammar”. It is a sign of education, intelligence, and class.

In fact, you can get away with a lot in Africa, if your english is inundated with big, big grammar. This information is rather important.

Alright, AI article/letter checkers. A few months ago, I was working on a personal finance article a friend sent me. It explained compound interest with some examples, but also had the slightly smug tone you see in money advice:

“If you invest $10,000 at 3% interest, you’ll have $10,300 after one year.” (I replaced the actual numbers.)

I glossed over that line because I am no longer a marketer’s dream, right?

But I have a long attention span, I’m a freelancer who barely freelances, so I have a lot of time on my hands, and there’s only so much running I can do. Haha.

I punched the numbers into a calculator. $10k at 3 percent is… $300 in interest. Not ten thousand three hundred in total interest. It left me confused about interest earned with ending balance.

Not the end of the world, and again, I have time. But I noticed other little things: awkward phrasing, metaphors, repetitive sentence structure.

I did what everybody does now: pasted a part of it into a search engine and found an outlet that flagged it as generated by an “AI tool”.

Before I go any further, I use AI. For analysis, codes, scripts, etc. I will not use AI to ideate.

Other reports showed that more than half of those “AI-written pieces” contained errors that should be corrected, including the compound interest example. Verge.

I’d like to say it didn’t change my trust in the article. I know the math, and yes, there should be errors sometimes. But I had questions like “who do I hold responsible if something goes wrong?”

That’s the core of this article, in a very practical, technical sense: it matters if AI wrote it.

So, I reached out to my friend, we talked about running and finance, and I casually asked what tool was used for such an impressive write-up.

No tool was used.

First class embarrassment caught me. See what my no-freelancing freelancing nearly cost me. She wrote it in her own words. A search engine said different. AND THAT IS A BIG PROBLEM.

Even without knowing the details of large language models, readers respond differently when they think AI is involved.

Recent studies/research show this:

• people were less likely to believe or share news headlines when they were labeled as AI-generated, even when the headlines were accurate and sometimes actually written by humans.

• University of Kansas found that when readers think AI played any role in producing news, their trust in that content drops, even if they don’t know exactly how AI was involved.

• A survey on news and media trust reported that 70% of people say they don’t trust information from an AI bot.

And the one that irks me so much, calling opinions, social media posts, etc, “chatgpt-generated”. It’s painful.

You mean to tell me that with all the big, big grammar I know and being colonized by queen’s english, I can’t have an opinion and write it with em-dashes and 3-structures without using AI?

I really wanted to try this with a copy of “The Bottled Leopard” by Wole Soyinka, just to see if it would be flagged as AI-generated. But even thinking about it is insulting to the nobel peace prize winner and his creativity.

When someone says, “It shouldn’t matter if AI wrote it as long as it’s good,” they’re arguing against a very consistent human reaction: authorship influences trust.

And honestly, that reaction is rational once you look at how these systems actually work.

If you’re reading this on Medium or Substack, you probably already know the high-level description:

A large language model (LLM) predicts the next token (piece of text) given the previous ones, using patterns it learned from a huge training corpus.

A bit abstract, so let’s translate it into “how this affects you as a reader”.

1. LLMs are probability engines, not fact engines

   Under the hood, the model is learning P(“next token” | “previous tokens”), not “truth”. It gives you the most statistically plausible sequence of words, given its training data and your prompt.

When it hallucinates a citation or number, it’s not “lying” in a human sense. It’s following the probability distribution, not a ground-truth database. Simple, right?

2. They generalize from data you can’t see
   You don’t know exactly which articles, books, news, pages, texts, or posts a model has ingested. Meaning:

   • You don’t know whose writing style or ideas are being remixed into this output. I may be the remix, you may be the remix.

   • You don’t know which biases, stereotypes, or errors were sandwiched into the training data.

3. They have no lived context or stakes
   LLMs and models don’t care if someone invests their savings based on a wrong interest calculation or misinterprets a medical explanation. There is no personal risk to the system. Even runway models don’t care, but I digress.

That combination — probabilistic text, opaque training data, and most importantly, AI has no skin in the game, makes AI-generated writing a different object than human writing.

You can’t treat those two sources as interchangeable without rethinking risk, responsibility, and consent. Two truths can exist, or whatever they say now.

Back to that finance article I mentioned, and this is the one time being african might come up in my article.

There’s a popular funny saying in Nigeria. If you post some outlandish information that is hard to believe, someone will ask, “Source?” And the funny response would be, “trust me bro” or “i go lie for you?”

it simply means it’s not a verifiable source, just trust me.

· If you get it, it’s funny; but if you don’t get it, forget about it.

When CNET quietly published AI-assisted personal finance explainers under a “CNET Money Staff” byline, readers had no clear indication that a model was involved, even though there was a not-so-obvious disclosure that they were created using “automation technology”.

Hm, again, whatever that means.

Only after external reporting by Futurism in jan 2023, it come out that:

• 41 of 77 AI-generated stories required corrections.

• There was evidence of plagiarism or very close paraphrasing in some of them. The Verge, WIRED

This was widespread, so I won’t focus on it, but this is not a one-off curiosity. It shows something annoying yet important:

If you don’t know AI was involved, you can’t correctly calibrate your trust.

If I know a human wrote a technical article, I might:

• Assume they have at least some domain knowledge.

• Check their bio, past work, and reputation.

• Email them if I find a mistake.

If I know an AI wrote most of it, my workflow changes:

• I gloss over it and don’t care enough to double-check the numbers, references, or quotes.

• I treat it as a starting point, not an authority. It would be asinine to treat it as an authority.

• I digress.

That’s not irrational bias. That’s good operational hygiene.

At this point, someone usually says: “we have AI detectors, we will paste what you wrote there, and in seconds, we will know if AI wrote it.”

I think I saw a post on Bluesky where someone was denied admission because the admissions team said, and I quote, “your letter was written with AI.”

The problem is that detection is, at best, noisy. At worst, it is actively harmful.

• A Stanford-linked study found that several AI detectors mislabeled a large majority of TOEFL essays by non-native English speakers as “AI-generated.” In that dataset, 97% of human-written essays were flagged as AI by at least one detector.

• Universities and teaching centers now warn faculty that detectors are unreliable and can be biased against non-native writers and underrepresented students, creating false accusations and harm.

• Recent reviews of AI-generated text detectors highlight high false-positive and false-negative rates in academic settings, especially as models improve and humans edit outputs.

Technically, this is expected. A skilled human writer and a good language model have overlapping “style distributions” at the surface level. Once you add paraphrasing or translation, you are no longer comparing “raw model output” to “raw human output.” You’re comparing blended, noisy signals.

So if we only care whether AI wrote it at the detection stage, we’re already too late.

We place the burden on students, freelancers, and writers to prove they are not machines, even though the tools you’re using “as checkers” are biased and uncertain.

I will not prove this to anyone. One day, we will be gone, and the last thing I want to take with me is that I tried to prove to someone that my TOEFL essays are not AI-generated.

So it doesn’t seem like I’m just complaining and whining without a solution, a better approach is to design systems where “we” disclose AI involvement up front.

SMPLE, right? It’s not.

You will be amazed how many people will tell you they don’t use AI when infact they use more AI than the one listening to me inside my microwave and airfryer.

Maybe the gospel according to me on Ethical use of AI will not be in vain.

You don’t have to take my word for it. Regulators are baking “it matters if AI wrote it” into law.

• The EU AI Act requires providers of general-purpose generative models (like Chatgpt) to disclose that content was generated by AI, comply with transparency rules, and publish certain information about training data and safeguards.

• The Act also contains specific transparency obligations for deepfakes and AI-manipulated media (Article 50),requiring clear labeling when users are interacting with AI-generated content.

• Spain has already moved to impose heavy fines (up to 7% of global revenue) for failing to label AI-generated content, especially where it risks misleading the public. Reuters

• Tiktok, pinterest, youtube insist you indicate if you use AI-generated text or voiceovers.

On the platform side, AI providers have usage and sharing policies for disclosing AI-generated content, especially for realistic media and public-facing communications. This one is by yours truly, OPENAI.

You can disagree with how strict these rules should be, but one thing is clear:
Platforms and policymakers are not neutral; authorship and disclosure matter.

There’s another dimension that’s less about risk and more about values, and this article has gone on long enough. I will cover this in another post.

This is where it gets practical for a Medium/Substack context.

As a writer, you still use AI in your workflow. I do, and many of us do. But you can draw a clearer line:

• Decide what you will and won’t automate
  Maybe you use AI to summarize research papers or generate outline options, but you write the actual narrative and examples yourself. Personally, I don’t ideate with AI. I am writing for humans, not for models, and hopefully, someone will extend to me the same grace I extend to others and read my work.

• Disclose the level of AI involvement
  I think Medium does this with images. You can include a line on how it was created. But for the written body, a line like this is ideal:

“Drafted using an AI assistant; all examples and conclusions were written and checked by me.”

Look, people will judge or herald you, but you’re one step away from making another AI user more transparent. Now that person is one step away from making another AI user more transparent, and now — alright, see where I’m going with this?

• Keep your voice visible
  Inject concrete details from your life or work that a generic model wouldn’t know. Like I did with my african/British/TOEFL/english woes. It’s more than “passing detectors”, but anchoring your writing in experience rather than pattern-matching.

• Own the final output
  AI can help produce a draft, but you can decide whether it is something you stand behind. That choice is what gives your name meaning under the headline.

As a reader

When you suspect AI involvement (or see a disclosure), you don’t have to panic. You can adjust:

• Treat AI-heavy pieces as starting points, not final sources of truth. That I have to say this is interesting.

• Verify numbers and citations, especially in domains like health, finance, and law.

• Pay attention to whether the publication or author clearly explains how they use AI. Transparency and disclosure are green flags; evasiveness is not.

In other words, you don’t have to boycott AI-assisted writing. You just have to read it differently.

Back to that compound interest…

I should apologize to my friend, but I also didn’t accuse them of anything. Thankfully.

Fortunately, It was a small number. No one lost their house because a blog post forgot to add the principal back in. But it revealed things which led to this article:

• Ai didn’t write it

• A news outlet treated a probabilistic text generator as a junior financial writer.

• It didn’t tell readers that a model had produced the “advice.”

• Only after external audits did the scale of the mistakes become public.

• I almost concluded my friend wrote hers with AI.

That’s why it matters whether AI wrote something.

Not because humans are perfect and machines are evil. But because the way we trust, verify, and assign responsibility is different when a model is doing the talking.

And it’s insulting to be constantly tagged “AI-generated.”

When we say “It doesn’t matter if AI wrote it,” we flatten all those differences.
When we admit “Yes, it matters,” we give ourselves permission to ask for disclosure and design better policies and tools, and build a healthier relationship between human judgment and machines.

This is not a middle finger to AI detectors.

It’s 5 fingers spread apart and pointed at them.

“UBANKA”. Pls don’t check the meaning of this.

And if you’re wondering how to end a piece like this, here’s my suggestion for your own version:

This article was written by a human mind that cares about the consequences of its words.
If AI helped anywhere along the way, you deserve to know that because it matters.

Further reading

• Altay, S. et al. (2024). People are less likely to believe and share headlines labeled as “AI-generated.” PMC

• University of Kansas. Study finds readers trust news less when AI is involved. KU News

• AVIXA / AI & Trust in Media survey, 2025. Major trust gap between journalism and AI bots. AVIXA Xchange

• Coverage of CNET’s AI-generated finance articles and subsequent corrections. Futurism, The Verge, WIRED

• Stanford HAI: AI-detectors biased against non-native English writers.

• Brandeis AI Teaching & Learning: Limitations of AI Detection Tools.

• EU AI Act and transparency obligations for AI-generated content and deepfakes.WilmerHale European Parliament artificialintelligenceact.eu

• Spain’s draft law on labeling AI-generated content and associated fines.

• Reporting on copyright and training data concerns from Japanese IP holders regarding AI models. — The Verge

Enjoyed this? Read, comment, and subscribe for more.

Be data-informed, data-driven, but not data-obsessed

• 🔗 Biz and whimsy: https://linktr.ee/amyusifoh

• 🔗 Connect with me on LinkedIn and GitHub for more analytics.

• 🔗 Get the free design toolkit: dashboard design toolkit

Data analyst ⬩ Spreadsheet advocate ⬩ Freelancer ⬩ Turning data into useful insights