Photo credit: ChatGPT 😅

The Problem

Apparently this is a thing that a lot of people already knew, but I heard it for the first time in August last year, on Day 1 of the Deep Learning Indaba 2025 as a passing comment: generative AI models cannot (almost 100%) produce images of analogue clock faces with any time other than 10 minutes past 10.

This is strictly an image problem, which is why I used to word “show”, not “tell”, in the title of this post.

Speaking of show and tell, allow me to show you an example of what I mean from everybody’s favourite PhD-level thinker, ChatGPT.

Unfortunately, ChatGPT does not allow us to embed shared chats, so you can either follow along with the screenshots below, or look at the chat here. The screenshot route is better if I say so myself. 😅

So let’s begin. A simple first ask:

Well done, ChatGPT.

On to my second simple ask:

Wait, what?

Come on. It’s not that hard. Let me spell it out for you…

At this point, despite being adequately warned, I couldn’t help but feel that ChatGPT was being an idiot. I knew why (and we will discuss the ‘why’ shortly) it was behaving this way, but as far as feelings are concerned, all I felt was contempt for the performance of this billion-dollar oaf.

And so I switched my approach. Surely, ChatGPT should be able to draw the correlation between digital and analogue time. I gave our beloved ChatGPT one more shot…

Yeah… Nope!

---

How We Got Here

So how exactly did we get here? Well, it turns out machine learning models (and LLMs by extension) are basically pattern-matching machines on steroids. They learn by ingesting massive datasets of images, text, whatever—and then reproducing what they’ve seen most often. The catch? They don’t actually understand anything. When you ask an AI model to draw a clock, it’s not thinking “okay, the hour hand goes here and the minute hand goes there.” Nope. It’s just regurgitating what it saw during training.

And here’s the kicker: if you look at stock photos, watch ads, or basically any marketing material featuring clocks, they’re almost always set to 10:10. This is a fundamental marketing gimmick: the 10:10 position creates a nice symmetrical “smile” with the hands, frames the brand logo perfectly, and doesn’t cover up any important details on the watch face. While this is aesthetically pleasing, when it becomes 99% of what AI models see during training, you’ve got a problem.

So what happens when you ask ChatGPT to generate a clock showing literally any other time? It trips over itself. Because in its “brain” (and I use that term very loosely), “clock face” equals “10:10.” The model hasn’t been malicious or stupid—it’s just doing exactly what we trained it to do: find patterns and reproduce them. And when one pattern dominates the dataset, that’s what you get.

Makes sense? Let’s continue.

The Bias Implications

Our example is a quirky clock problem, but the implications go way beyond that; it’s a toy example of how bias works in AI systems across the board. Whether it’s race, gender, language, or culture—if your training data is skewed, your model will be skewed. The bias isn’t intentional; it’s mathematical. Generative AI is holding up a mirror to humanity, and this could play out in serious ways downstream.

Racism, Sexism, and Other ‘isms’

Moving away from clocks, remember when Google’s Gemini famously overcorrected for diversity so hard it started generating historically inaccurate images… like a black pope 😅. This was an over-exertion of efforts to move away from the typical stereotypes that have plagued and characterized history: white male CEOs, female nurses vs male doctors, dark-skinned criminals, etc.

While sad, it’s fundamentally an issue of representation. AI learns from examples, so until we shift the representations present in our data examples, AI will continue to produce what it sees: the good, the bad, and the ugly. I’m saying that like the youth in society, AI models also need good role models. 😅

Consider this (and there’s no right answer): would you be interested in seeing more height diversity in the NBA? I mean, do you want to see much shorter players? Would you care at all? The distribution of heights in the league is obviously skewed, and if you care to see the global height distribution, here’s another link for you to nerd out. That’s representation… the statistical stories baked into the data. And as with my NBA example, not everyone cares all the time. But sometimes we should…

Beyond images, AI systems are being used in courts of law for tasks like determining bail amounts, and even influencing sentencing. If those systems are trained on biased data (and they likely are), then they’re baking historical prejudices directly into “objective” legal decisions. A person’s freedom might hinge on an algorithm that learned from a world that was already unfair. The shift from viewing this as just “interesting” to “unethical and absurd” may depend on which side of a jury bench you sit.

Then there’s the impact on creativity. As generative AI becomes the go-to tool for content creation, we’re staring down the barrel of the death of originality… at least, as originality is defined today. If every AI model is trained on the same corpus of data—largely dominated by Western, English-speaking, male perspectives—then every output starts to look the same. The quirks, the cultural nuances, the weirdness that makes human creativity interesting is sucked into a data smoothie of “what worked before.”

The clock problem is funny. The rest… Dystopia maybe?

Of Dashes & Delves

As a wirter, this section hits close to home.

We now live in a reality where good writing is quickly suspected of being AI-generated simply because of specific word choices or punctuation marks. The very things that used to signify thoughtful, well-crafted essays are now ChatGPT red flags.

My favourite example is the em-dash (—), a perfectly legitimate punctuation mark that I have used, even when writing by hand, since high school. Today, use it a wee bit too much and your work may be tagged as robot-written AI slop.

A famous example is Paul Graham’s beef with the word “delve”. Paul G is a highly respected tech entrepreneur and writer, so his opinions tend to be very well regarded, but I think this is one instance where he was very wrong on the money, and confidence while wrong is a nasty combination.

In 2024, Paul G posted a tweet expressing a lack of trust in the human authenticity of any piece of text which uses the word “delve”. The backlash was swift and brutal. Writers, academics, tech bros, keyboard warriors on X pointed out that “delve” is, you know, an actual word that humans have been using long before AI came along… maybe not in Paul Graham’s world, but given the colourful history of colonialism, dynamics of vocabulary over time, lack of monopoly by any group on the language today, and frankly the fact that English speakers the world over couldn’t care less, the whole thread was a dumpster fire. A dumpster fire with graphs and strong opinions, yes, but a dumpster fire nonetheless.

When we start judging ideas by the vocabulary transporting the ideas—which we all do sometimes—we stand the risk of ignoring some good ideas, and on occasion, hailing some polished turds as solid ideas, when a turd is in fact, just a turd… no matter how shiny.

To his credit, Paul G’s thoughts weren’t entirely unfounded. It turns out that ChatGPT does love the word “delve.” Like, a lot. The model was trained on a corpus of text where “delve” appeared frequently enough to become part of its linguistic fingerprint. So now, whenever someone writes naturally and happens to use that word, they risk being accused of outsourcing their thinking to a machine.

These false positives of AI detection affect students and professionals alike when their written work happens to appear too smooth. The irony is thick: writing well makes you look suspicious. If your grammar is impeccable, your sentences flow smoothly, and you use semicolons correctly, congratulations—you’re now a person of interest. (Ah, there I go using another em-dash 🤦🏾‍♂️)

I don’t know where this ends. Should I start writing worse on purpose just to prove I’m human? Should I throw in a few typos and some awkward phrasing? I’ve asked myself these questions an unhealthy number of times since I discovered these biases. I don’t have the answer… yet. I guess context matters.

---

The Human Advantage

Okay… so what?

I don;t want to leave you depressed. You were hapving a perfectly good day before I came along with this post. Here’s some hope for your time.

Not too long ago I watched this YouTube short on Sensory Linkages: the ability of the human brain to link different sensory modalities; in this case, sight and touch. How do you guess how something will feel, or even taste, just by looking at it?

The answer is that there are links that are built in a person’s brain over time as we grow and develop between the differnt modalities. That’s how you know what ‘round’ looks like, so you can guess how a ball will feel just by looking at it.

The idea that AI has never tasted ice cream gives me comfort when it comes to the sanctity of the human experience. I’ve heard of electronic tasting, but there is just something about the human experience which AI and emerging technology have not captured, and in my opinion, will not capture, for a long time if at all.

The fact that AI understands time and understands watches as two separate entities and cannot easily link the two concepts tells me something… ice cream is safe! 😂 ChatGPT knows ice cream is sweet, but how will it ever understand what exactly “sweet” means? Quick.. describe “sweet”. “Like sugar”, you say, or “Like honey”. True, but ChatGPT has not tasted either of those things 😜. Yes, there’s a chemistry to taste, and smell, and secialized hardware can probably be developed, but who really cares to have their computer smell roses with a USB-connected nose? It’s such a human thing… safe from the machines.

The human experience; from social interactions and even deeper, to the level of the soul and spirit, if you will, with all the complexities of emotion, culture, and yes, taste, is what sets us apart from the robots.

Theres a value there that even babies can appreciate, but a supercomputer will not.

AI could never. Or, at least, not any time soon.

The kids will be fine. I can feel it… something AI can’t do.

✌🏾😎