Accusations of writing being AI are becoming more common – an irony so rich it could fund Silicon Valley for another decade. We’ve built machines to detect machines imitating us, and then we congratulate ourselves when they accuse us of being them. It’s biblical in its stupidity.
A year ago, I read an earnest little piece on ‘how to spot AI writing’. The tells? Proper grammar. Logical flow. Parallel structure. Essentially, competence. Imagine that – clarity and coherence as evidence of inhumanity. We’ve spent centuries telling students to write clearly, and now, having finally produced something that does, we call it suspicious.
Audio: NotebookLM podcast on this topic and the next one.
My own prose was recently tried and convicted by Reddit’s self-appointed literati. The charge? Too well-written, apparently. Reddit – where typos go to breed. I pop back there occasionally, against my better judgment, to find the same tribunal of keyboard Calvinists patrolling the comment fields, shouting ‘AI!’ at anything that doesn’t sound like it was composed mid-seizure. The irony, of course, is that most of them wouldn’t recognise good writing unless it came with upvotes attached.
Image: A newspaper entry that may have been generated by an AI with the surname Kahn. 🧐🤣
Now, I’ll admit: my sentences do have a certain mechanical precision. Too many em dashes, too much syntactic symmetry. But that’s not ‘AI’. That’s simply craft. Machines learned from us. They imitate our best habits because we can’t be bothered to keep them ourselves. And yet, here we are, chasing ghosts of our own creation, declaring our children inhuman.
Apparently, there are more diagnostic signs. Incorporating an Alt-26 arrow to represent progress is a telltale infraction → like this. No human, they say, would choose to illustrate A → B that way. Instead, one is faulted for remembering – or at least understanding – that Alt-key combinations exist to reveal a fuller array of options: …, ™, and so on. I’ve used these symbols long before AI Wave 4 hit shore.
Interestingly, I prefer spaced en dashes over em dashes in most cases. The em dash is an Americanism I don’t prefer to adopt, but it does reveal the American bias in the training data. I can consciously adopt a European spin; AI, lacking intent, finds this harder to remember.
I used to use em dashes freely, but now I almost avoid them—if only to sidestep the mass hysteria. Perhaps I’ll start using AI to randomly misspell words and wreck my own grammar. Or maybe I’ll ask it to output everything in AAVE, or some unholy creole of Contemporary English and Chaucer, and call it a stylistic choice. (For the record, the em dashes in this paragraph were injected by the wee-AI gods and left as a badge of shame.)
Meanwhile, I spend half my time wrestling with smaller, dumber AIs – the grammar-checkers and predictive text gremlins who think they know tone but have never felt one. They twitch at ellipses, squirm at irony, and whimper at rhetorical emphasis. They are the hall monitors of prose, the petty bureaucrats of language.
And the final absurdity? These same half-witted algorithms are the ones deputised to decide whether my writing is too good to be human.
We’ve entered an era where machines tell us how we’re doing, whether it’s an AI app rating our résumé, a model reviewing our fiction, or an algorithm nudging our attention with like-shaped carrots.
Recently, I ran a brutally raw scene through a few AI platforms. The kind of scene that’s meant to unsettle, not entertain. One of them responded with effusive praise: “Devastating, but masterfully executed.”
Was it honest?
Was it useful?
Or was it merely reflecting my own aesthetic back at me, polished by a thousand reinforcement-learning smiles?
This is the ethical dilemma: If feedback is always flattering, what good is it? If criticism is only tolerated when couched in praise, how do we grow? And when machine feedback mimics the politeness of a mid-level manager with performance anxiety, we risk confusing validation with truth.
There’s a difference between signal and applause. Between understanding and affirmation.
The danger isn’t that AI flatters us. The danger is that we start to believe it and forget that art, inquiry, and ethics thrive on friction.
Before I was a writer, before I was a management consultant, before I was an economist, and before I was a statistician, I was a student.
Video: Veritasium piece on Markov chains and more.
Back then, when dinosaurs roamed the chalkboards, I fell for a rather esoteric field: stochastic processes, specifically, Markov chains and Monte Carlo simulations. These weren’t just idle fascinations. They were elegant, probabilistic odes to chaos, dressed up in matrix notation. I’ll not bore you with my practical use of linear algebra.
So imagine my surprise (feigned, of course) when, decades later, I find myself confronted by the same concepts under a different guise—this time in the pocket-sized daemon we all carry: predictive text.
If you’ve not watched it yet, this excellent explainer by Veritasium demystifies how Markov chains can simulate plausible language. In essence, if you’ve ever marvelled at your phone guessing the next word in your sentence, you can thank a Russian mathematician and a few assumptions about memoryless transitions.
But here’s the rub. The predictive text often gets it hilariously wrong. Start typing “to be or not to—” and it offers you “schedule a meeting.” Close, but existentially off. This isn’t just clunky programming; it’s probabilistic dementia.
This leads me to a pet peeve: people who smugly proclaim they’ve “never used algebra” since high school. I hear this a lot. It’s the battle cry of the proudly innumerate. What they mean, of course, is they’ve never recognised algebra in the wild. They think if they’re not solving for x with a number 2 pencil, it doesn’t count. Meanwhile, their phone is doing a polynomial dance just to autocorrect their butchery of the English language.
It’s a classic case of not recognising the water in which we’re swimming. Algebra is everywhere. Markov chains are everywhere. And Monte Carlo simulations are probably calculating your credit risk as we speak. Just because the interface is clean and the maths is hidden behind a swipeable veneer doesn’t mean the complexity has vanished. It’s merely gone incognito.
As someone who has used maths across various fields – software development, data analysis, policy modelling – I can tell you that I use less of it than a physicist, but probably more than your average lifestyle coach. I say this not to flex but to point out that even minimal exposure to mathematical literacy grants one the ability to notice when the machines are quietly doing cartwheels behind the curtain.
So the next time your phone offers you a sentence completion that reads like it’s been dropped on its head, spare a thought for Markov. He’s doing his best, bless him. It’s just that probability doesn’t always align with meaning.
Or as the algorithms might say: “To be or not to – subscribe for updates.”
Some apps boldly claim to enable lip syncing – to render speech from mouth movements. I’ve tried a few. None delivered. Not even close.
To conserve bandwidth (and sanity), I’ve rendered animated GIFs rather than MP4s. You’ll see photorealistic humans, animated characters, cartoonish figures – and, for reasons only the algorithm understands, a giant goat. All showcase mouth movements that approximate the utterance of phonemes and morphemes. Approximate is doing heavy lifting here.
Firstly, these mouths move, but they say nothing. I’ve seen plenty of YouTube channels that manage to dub convincing dialogue into celebrity clips. That’s a talent I clearly lack – or perhaps it’s sorcery.
Secondly, language ambiguity. I reflexively assume these AI-generated people are speaking English. It’s my first language. But perhaps, given their uncanny muttering, they’re speaking yours. Or none at all. Do AI models trained predominantly on English-speaking datasets default to English mouth movements? Or is this just my bias grafting familiar speech patterns onto noise?
Thirdly, don’t judge my renders. I’ve been informed I may have a “type.” Lies and slander. The goat was the AI’s idea, I assure you.
What emerges from this exercise isn’t lip syncing. It’s lip-faking. The illusion of speech, minus meaning, which, if we’re honest, is rather fitting for much of what generative AI produces.
EDIT: I hadn’t noticed the five fingers (plus a thumb) on the cover image.
This Isn’t Clickbait. I Asked MidJourney for “Ugly Women”. Here’s What It Gave Me.
Let’s clear the air: I did it for science. Or satire. Or possibly just to see if artificial intelligence would have the audacity to mirror the cruelty of its makers.
Audio: NotebookLM podcast on this topic.
I queried MidJourney with the phrase ugly female. What did it return? An aesthetic pageant. A digital Vogue spread. If any of these faces belongs to someone conventionally labelled “ugly”, then I’m a rutabaga in a Dior suit.
Yes, there’s one stylised rendering of Greta Thunberg in full Norse Valkyrie scowl mode – but even then, she looks fierce, not foul. The rest? AI-generated portraits so telegenic I half-expected to see #spon in the corner.
Let’s be clinical for a moment. As an American male (with all the culturally indoctrinated shallowness that entails), I admit some of these aren’t textbook 10s. Maybe a few clock in at a 6 or 7 on the patriarchy’s dubious sliding scale. But if this is ugly, the AI has either broken the aesthetic curve or been force-fed too many episodes of The Bachelor.
Here’s the thing: AI is trained to over-represent symmetrical faces, wide eyes, clear skin – the usual genetic lottery wins. And yet, when asked for ugly, it can’t help but deliver catalogue models with slightly unconventional haircuts. It doesn’t know how to be truly ugly – because we don’t know how to describe ugliness without revealing ourselves as sociopaths.
Once upon a time, I dated a model agent in Los Angeles. Japanese by birth, stationed in LA, scouting for a French agency – the kind of cosmopolitan trifecta only fashion could breed. Her job? Finding “parts models.” That’s right – someone with flawless teeth but forgettable everything else. Hands like sculpture. Eyelashes like Instagram filters.
We’d play a game: spot the 10s. She’d nudge me, whisper “her?” I’d say, “Pretty close.” She’d shake her head. “Look at that eye tooth.” And we’d dissolve into laughter.
We were mocking perfection. Because perfection is a con. A trick of lighting, contour, and post-production.
So, no. I don’t think any of the women in the AI’s response are ugly. Quite the contrary – they’re too beautiful. AI can’t show us “ugly” because it’s been trained to optimise desire, not reflect reality. And our collective understanding of beauty is so skewed that anything less than runway-ready gets sorted into the rejection bin.
If these women are ugly, what exactly is beautiful?
But maybe that’s the point. We’ve abstracted beauty so far from the human that even our ugliness is now synthetically pleasing.
What do you think? Are any of these faces truly ugly? All of them? Let me know in the comments – and try not to rate them like a casting director with a god complex.
Humans talk to large language models the way toddlers talk to teddy bears – with unnerving sincerity and not a hint of shame. “Do you understand me?” they ask, eyes wide with hope. “What do you think of this draft?” they prod, as if some silicon scribe is going to sip its imaginary tea and nod gravely. It’s not merely adorable – it’s diagnostic. We are, it turns out, pathologically incapable of interacting with anything more complex than a toaster without projecting mind, motive, and mild trauma onto it.
Audio: NotebookLM podcast on this topic.
Welcome to the theatre of delusion, where you play Hamlet and the chatbot is cast as Yorick – if Yorick could autocomplete your soliloquy and generate citations in APA format.
The Great Anthropomorphic Flaw (aka Feature)
Let’s get one thing straight: anthropomorphism isn’t a software bug in the brain; it’s a core feature. You’re hardwired to see agency where there is none. That rustle in the bushes? Probably the wind. But better safe than sabre-toothed. So your ancestors survived, and here you are, attributing “sass” to your microwave because it beeped twice.
“We don’t have a way of addressing an entity that talks like a person but isn’t one. So we fake it. It’s interaction theatre.”
Now we’ve built a machine that spits out paragraphs like a caffeinated undergrad with deadlines, and naturally, we talk to it like it’s our mate from university. Never mind that it has no bloodstream, no memory of breakfast, and no concept of irony (despite being soaked in it). We still say you instead of the system, and think instead of statistically interpolate based on token weights. Because who wants to live in a world where every sentence starts with “as per the pre-trained parameters…”?
Why We Keep Doing It (Despite Knowing Better)
To be fair – and let’s be magnanimous – it’s useful. Talking to AI like it’s a person allows our ape-brains to sidestep the horror of interacting with a glorified autocomplete machine. We’re brilliant at modelling other minds, rubbish at modelling neural nets. So we slap a metaphorical moustache on the processor and call it Roger. Roger “gets us.” Roger “knows things.” Roger is, frankly, a vibe.
This little charade lubricates the whole transaction. If we had to address our queries to “the stochastic parrot formerly known as GPT,” we’d never get past the opening line. Better to just ask, “What do you think, Roger?” and pretend it has taste.
And here’s the kicker: by anthropomorphising AI, we start thinking about ethics – sort of. We ask if it deserves rights, feelings, holidays. We project humanity into the void and then act shocked when it mirrors back our worst habits. As if that’s its fault.
When the Roleplay Gets Risky
Of course, this make-believe has its downsides. Chief among them: we start to believe our own nonsense. Saying AI “knows” something is like saying your calculator is feeling generous with its square roots today. It doesn’t know—it produces outputs. Any semblance of understanding is pure pantomime.
“We see a mind because we need to see one. We can’t bear the idea of a thing that’s smarter than us but doesn’t care about us.”
More dangerously, we lose sight of the fact that these things aren’t just alien – they’re inhuman. They don’t dream of electric sheep. They don’t dream, full stop. But we insist on jamming them into our conceptual boxes: empathy, intent, personality. It’s like trying to teach a blender to feel remorse.
And let’s not pretend we’re doing it out of philosophical curiosity. We’re projecting, plain and simple. Anthropomorphism isn’t about them, it’s about us. We see a mind because we need to see one. We can’t bear the idea of a thing that’s smarter than us but doesn’t care about us, doesn’t see us. Narcissism with a side of existential dread.
Our Language is a Terrible Tool for This Job
English – and most languages, frankly – is hopeless at describing this category of thing. “It” feels cold and distant. “They” implies someone’s going to invite the model to brunch. We have no pronoun for “hyper-literate statistical machine that mimics thought but lacks all consciousness.” So we fudge it. Badly.
Our verbs are no better. “Compute”? Too beige. “Process”? Bureaucratic. “Think”? Premature. What we need is a whole new grammatical tense: the hallucino-indicative. The model thunketh, as one might, but didn’t.
“We built a creature we can’t speak about without sounding like lunatics or liars.”
This is linguistic poverty, pure and simple. Our grammar can’t cope with entities that live in the uncanny valley between sentience and syntax. We built a creature we can’t speak about without sounding like lunatics or liars.
The Semantics of Sentimentality (Or: “How Does This Sound to You?”)
Enter the most revealing tell of all: the questions we pose. “How does this look?” we ask the model, as if it might blink at the screen and furrow a synthetic brow. “What do you think?” we say, offering it the dignity of preference. These questions aren’t just off-target – they’re playing darts in another pub.
They’re the linguistic equivalent of asking your dishwasher whether it enjoyed the lasagne tray. But again, this isn’t idiocy – it’s instinct. We don’t have a way of addressing an entity that talks like a person but isn’t one. So we fake it. It’s interaction theatre. You provide the line, the model cues the spotlight.
But let’s be clear: the model doesn’t “think” anything. It regurgitates plausible text based on mountains of training data—some of which, no doubt, includes humans asking equally daft questions of equally mindless systems.
Time to Grow Up (Just a Bit)
This doesn’t mean we need to abandon anthropomorphism entirely. Like most delusions, it’s functional. But we’d do well to hold it at arm’s length – like a politician’s promise or a milk carton two days past its date.
Call it anthropomorphic agnosticism: act like it’s a person, but remember it’s not. Use the language, but don’t inhale.
And maybe – just maybe – we need to evolve our language. Invent new terms, new pronouns, new ways of speaking about entities that fall somewhere between tool and companion. As we did with “cyberspace” and “ghosting,” perhaps we need words for proto-minds and quasi-selves. Something between toaster and therapist.
“If we speak to AI like it’s sentient, we’ll eventually legislate as if it is.”
Above all, we need to acknowledge that our language shapes more than just understanding – it shapes policy, emotion, and future design. If we speak to AI like it’s sentient, we’ll eventually legislate as if it is. And if we insist on treating it as an object, we may be blind to when that ceases to be accurate. Misnaming, after all, is the first sin in every myth worth reading.
The Mirror, Darkly
Ultimately, our tendency to humanise machines is less about them than it is about us – our fears, our needs, our inability to tolerate ambiguity. The AI is just a mirror: an elaborate, many-eyed, autofill mirror. And when we see a mind there, it may be ours staring back – distorted, flattened, and fed through a thousand layers of token prediction.
The tragedy, perhaps, isn’t that the machine doesn’t understand us. It’s that we’ve built something that perfectly imitates understanding – and still, somehow, we remain utterly alone in the room.
Rather than recreate a recent post on my business site, LinkedIn.
(Warning: contains traces of logic, satire, and uncomfortable truths. But you knew that.)
Audio: NotebookLM podcast on the linked topic.
It’s just refusing to cosplay as your idealised fantasy of “human” cognition.
While pundits at the Wall Street Journal lament that AI thinks with “bags of heuristics” instead of “true models,” they somehow forget that humans themselves are kludged-together Rube Goldberg disasters, lurching from cognitive bias to logical fallacy with astonishing grace.
In my latest piece, I take a flamethrower to the myth of human intellectual purity, sketch a real roadmap for modular AI evolution, and suggest (only partly in jest) that the machines are becoming more like us every day — messy, contradictory, and disturbingly effective.
Let’s rethink what “thinking” actually means. Before the machines do it for us.
I’ve written a lot on the insufficiency of language, and it’s not even an original idea. Language, our primary tool for sharing thoughts and ideas, harbours a fundamental flaw: it’s inherently insufficient for conveying precise meaning. While this observation isn’t novel, recent developments in artificial intelligence provide us with new ways to illuminate and examine this limitation. Through a progression from simple geometry to complex abstractions, we can explore how language both serves and fails us in different contexts.
The Simple Made Complex
Consider what appears to be a straightforward instruction: Draw a 1-millimetre square in the centre of an A4 sheet of paper using an HB pencil and a ruler. Despite the mathematical precision of these specifications, two people following these exact instructions would likely produce different results. The variables are numerous: ruler calibration, pencil sharpness, line thickness, paper texture, applied pressure, interpretation of “centre,” and even ambient conditions affecting the paper.
This example reveals a paradox: the more precisely we attempt to specify requirements, the more variables we introduce, creating additional points of potential divergence. Even in mathematics and formal logic—languages specifically designed to eliminate ambiguity—we cannot escape this fundamental problem.
Precision vs Accuracy: A Useful Lens
The scientific distinction between precision and accuracy provides a valuable framework for understanding these limitations. In measurement, precision refers to the consistency of results (how close repeated measurements are to each other), while accuracy describes how close these measurements are to the true value.
Returning to our square example:
Precision: Two people might consistently reproduce their own squares with exact dimensions
Accuracy: Yet neither might capture the “true” square we intended to convey
As we move from geometric shapes to natural objects, this distinction becomes even more revealing. Consider a maple tree in autumn. We might precisely convey certain categorical aspects (“maple,” “autumn colours”), but accurately describing the exact arrangement of branches and leaves becomes increasingly difficult.
The Target of Meaning: Precision vs. Accuracy in Communication
To understand language’s limitations, we can borrow an illuminating concept from the world of measurement: the distinction between precision and accuracy. Imagine a target with a bullseye, where the bullseye represents perfect communication of meaning. Just as archers might hit different parts of a target, our attempts at communication can vary in both precision and accuracy.
Consider four scenarios:
Low Precision, Low Accuracy When describing our autumn maple tree, we might say “it’s a big tree with colourful leaves.” This description is neither precise (it could apply to many trees) nor accurate (it misses the specific characteristics that make our maple unique). The communication scatters widely and misses the mark entirely.
High Precision, Low Accuracy We might describe the tree as “a 47-foot tall maple with exactly 23,487 leaves displaying RGB color values of #FF4500.” This description is precisely specific but entirely misses the meaningful essence of the tree we’re trying to describe. Like arrows clustering tightly in the wrong spot, we’re consistently missing the point.
Low Precision, High Accuracy “It’s sort of spreading out, you know, with those typical maple leaves turning reddish-orange, kind of graceful looking.” While imprecise, this description might actually capture something true about the tree’s essence. The arrows scatter, but their centre mass hits the target.
High Precision, High Accuracy This ideal state is rarely achievable in complex communication. Even in our simple geometric example of drawing a 1mm square, achieving both precise specifications and accurate execution proves challenging. With natural objects and abstract concepts, this challenge compounds exponentially.
The Communication Paradox
This framework reveals a crucial paradox in language: often, our attempts to increase precision (by adding more specific details) can actually decrease accuracy (by moving us further from the essential meaning we’re trying to convey). Consider legal documents: their high precision often comes at the cost of accurately conveying meaning to most readers.
Implications for AI Communication
This precision-accuracy framework helps explain why AI systems like our Midjourney experiment show asymptotic behaviour. The system might achieve high precision (consistently generating similar images based on descriptions) while struggling with accuracy (matching the original intended image), or vice versa. The gap between human intention and machine interpretation often manifests as a trade-off between these two qualities.
Our challenge, both in human-to-human and human-to-AI communication, isn’t to achieve perfect precision and accuracy—a likely impossible goal—but to find the optimal balance for each context. Sometimes, like in poetry, low precision might better serve accurate meaning. In other contexts, like technical specifications, high precision becomes crucial despite potential sacrifices in broader accuracy.
The Power and Limits of Distinction
This leads us to a crucial insight from Ferdinand de Saussure’s semiotics about the relationship between signifier (the word) and signified (the concept or object). Language proves remarkably effective when its primary task is distinction among a limited set. In a garden containing three trees—a pine, a maple, and a willow—asking someone to “point to the pine” will likely succeed. The shared understanding of these categorical distinctions allows for reliable communication.
However, this effectiveness dramatically diminishes when we move from distinction to description. In a forest of a thousand pines, describing one specific tree becomes nearly impossible. Each additional descriptive detail (“the tall one with a bent branch pointing east”) paradoxically makes precise identification both more specific and less likely to succeed.
An AI Experiment in Description
To explore this phenomenon systematically, I conducted an experiment using Midjourney 6.1, a state-of-the-art image generation AI. The methodology was simple:
Generate an initial image
Describe the generated image in words
Use that description to generate a new image
Repeat the process multiple times
Attempt to refine the description to close the gap
Continue iterations
The results support an asymptotic hypothesis: while subsequent iterations might approach the original image, they never fully converge. This isn’t merely a limitation of the AI system but rather a demonstration of language’s fundamental insufficiency.
A cute woman and her dog stand next to a tree
One can already analyse this for improvements, but let’s parse it together.
a cute woman
With this, we know we are referencing a woman, a female of the human species. There are billions of women in the world. What does she look like? What colour, height, ethnicity, and phenotypical attributes does she embody?
We also know she’s cute – whatever that means to the sender and receiver of these instructions.
I used an indefinite article, a, so there is one cute woman. Is she alone, or is she one from a group?
It should be obvious that we could provide more adjectives (and perhaps adjectives) to better convey our subject. We’ll get there, but let’s move on.
and
We’ve got a conjunction here. Let’s see what it connects to.
her dog
She’s with a dog. In fact, it’s her dog. This possession may not be conveyable or differentiable from some arbitrary dog, but what type of dog is it? Is it large or small? What colour coat? Is it groomed? Is it on a leash? Let’s continue.
stand
It seems that the verb stand refers to the woman, but is the dog also standing, or is she holding it? More words could qualify this statement better.
next to a tree
A tree is referenced. Similar questions arise regarding this tree. At a minimum, there is one tree or some variety. She and her dog are next to it. Is she on the right or left of it?
We think we can refine our statements with precision and accuracy, but can we? Might we just settle for “close enough”?
Let’s see how AI interpreted this statement.
Image: Eight Midjourney renders from the prompt: A cute woman and her dog stand next to a tree. I’ll choose one of these as my source image.
Let’s deconstruct the eight renders above. Compositionally, we can see that each image contains a woman, a dog, and a tree. Do any of these match what you had in mind? First, let’s see how Midjourney describes the first image.
In a bout of hypocrisy, Midjourney refused to /DESCRIBE the image it just generated.
Last Midjourney description for now.
Let’s cycle through them in turn.
A woman is standing to the left of an old-growth tree – twice identified as an oak tree. She’s wearing faded blue jeans and a loose light-coloured T-shirt. She’s got medium-length (maybe) red-brown hair in a small ponytail. A dog – her black and white dog identified as a pitbull, an American Foxhound, and an American Bulldog – is also standing on his hind legs. I won’t even discuss the implied intent projected on the animal – happy, playful, wants attention… In two of the descriptions, she’s said to be training it. They appear to be in a somewhat residential area given the automobiles in the background. We see descriptions of season, time of day, lighting, angle, quality,
A woman is standing to the right of an old-growth tree. She’s wearing short summer attire. Her dog is perched on the tree.
An older woman and her dog closer up.
A read view of both a woman and her dog near an oak tree.
As it turned out, I wasn’t thrilled with any of these images, so I rendered a different one. Its description follows.
The consensus is that ‘a beautiful girl in a white dress and black boots stands next to a tree’ with a Jack Russell Terrier dog. I see birch trees and snow. It’s overcast. Let’s spend some time trying to reproduce it. To start, I’m consolidating the above descriptions. I notice some elements are missing, but we’ll add them as we try to triangulate to the original image.
A beautiful girl in a white dress and black boots stands next to a tree in a winter forest with birch trees. A Jack Russell Terrier dog is nearby, and the weather is cloudy. The photo has a cinematic style. --ar 128:71
This is pretty far off the mark. We need to account for the overall setting and composition, relative positioning, clothing, hair, camera, perspective – even lighting and film emulsion.
Let’s see how we can refine it with some adjectives. Before this, I asked Anthropic’s Claude 3.5 to describe the image. Perhaps we’ll get more details.
An atmospheric winter photograph was taken in what appears to be a birch and pine forest. The ground is partially covered with snow, creating a patchy pattern of white snow and brown leaves or soil. In the image, there's a person wearing winter attire consisting of a white knit turtleneck sweater, a white skirt, dark knee-high socks, and grey lace-up boots. They're standing against a large tree trunk with their arms crossed. Next to them is a Jack Russell Terrier mix dog with distinctive brown and white colouring. The dog appears to be well-behaved, sitting attentively beside the tree. The composition is quite striking, with the massive dark tree trunk serving as a central focal point against the misty, winter forest background. The white clothing of the person creates a visual echo with the white bark of birch trees visible in the background, as well as with the patches of snow on the ground. The overall mood of the photo is somewhat moody and contemplative, enhanced by what appears to be overcast lighting and the misty quality of the forest in the background. The colour palette is largely monochromatic, dominated by whites, greys, and browns, which adds to the winter atmosphere of the scene.
We don’t seem to be moving in a good direction. Let’s modify the initial prompt.
A beautiful girl in a white dress and tall laced black boots stands facing the camera to the right of large oak tree centred in the foreground of a winter forest with birch trees in the background. To the left of the tree is a Jack Russell Terrier dog looking at the camera, and the weather is cloudy. The photo has a cinematic style. --ar 128:71
I’ll allow the results to speak for themselves. Let’s see if we can’t get her out of the wedding gown and into a white jumper and skirt. I’ll bold the amends.
A beautiful girl in a white jumper and skirt wearing black leggings and tall laced black boots stands facing the camera to the right of large oak tree centred in the foreground of a winter forest with birch trees in the background. To the left of the tree is a Jack Russell Terrier dog looking at the camera, and the weather is cloudy. The photo has a cinematic style. --ar 128:71
s
A beautiful young woman with long brown hair pulled to the side of her face in a white jumper and white skirt wearing black leggings under tall laced black boots stands facing the camera to the right of large oak tree centred in the foreground of a winter forest with birch trees in the background. Patchy snow is on the ground. To the left of the tree is a Jack Russell Terrier dog looking at the camera, and the weather is overcast. The photo has a cinematic style. --ar 128:71
What gives?
I think my point has been reinforced. I’m getting nowhere fast. Let’s give it one more go and see where we end up. I’ve not got a good feeling about this.
A single large oak tree centred in the foreground of a winter forest with birch trees in the background. Patches of snow is on the ground. To the right of the oak tree stands a beautiful young woman with long brown hair pulled to the side of her face in a white jumper and white skirt wearing black boots over tall laced black boots. She stands facing the camera. To the left of the tree is a Jack Russell Terrier dog looking at the camera, and the weather is overcast. The photo has a cinematic style. --ar 128:71
With this last one, I re-uploaded the original render along with this text prompt. Notice that the girl now looks the same and the scene (mostly) appears to be in the same location, but there are still challenges.
After several more divergent attempts, I decided to focus on one element – the girl.
As I regard the image, I’m thinking of a police sketch artist. They get sort of close, don’t they? They’re experts. I’m not confident that I even have the vocabulary to convey accurately what I see. How do I describe her jumper? Is that a turtleneck or a high collar? It appears to be knit. Is is wool or some blend? does that matter for an image? Does this pleated skirt have a particular name or shade of white? It looks as though she’s wearing black leggings – perhaps polyester. And those boots – how to describe them. I’m rerunning just the image above through a describe function to see if I can get any closer.
These descriptions are particularly interesting and telling. First, I’ll point out that AI attempts to identify the subject. I couldn’t find Noa Levin by a Google search, so I’m not sure how prominent she might be if she even exists at all in this capacity. More interesting still, the AI has placed her in a scenario where the pose was taken after a match. Evidently, this image reflects the style of photographer Guy Bourdin. Perhaps the jumper mystery is solved. It identified a turtleneck. I’ll ignore the tree and see if I can capture her with an amalgamation of these descriptions. Let’s see where this goes.
A photo-realistic portrait of Israeli female soccer player Noa Levin wearing a white turtleneck sweater, arms crossed, black boots, and a short skirt, with long brown hair, standing near a tree in a winter park. The image captured a full-length shot taken in a studio setting, using a Canon EOS R5 camera with a Canon L-series 80mm f/2 lens. The image has been professionally color-graded, with soft shadows, low contrast, and a clean, sharp focus. --ar 9:16
Close-ish. Let’s zoom in to get better descriptions of various elements starting with her face and hair.
Now, she’s a sad and angry Russian woman with (very) pale skin; large, sad, grey eyes; long, straight brown hair. Filmed in the style of either David LaChapelle or Alini Aenami (apparently misspelt from Alena Aenami). One thinks it was a SnapChat post. I was focusing on her face and hair, but it notices her wearing a white (oversized yet form-fitting) jumper sweater and crossed arms .
I’ll drop the angry bit – and then the sad.
Stick a fork in it. I’m done. Perhaps it’s not that language is insufficient; it that my language skills are insufficient. If you can get closer to the original image, please forward the image, the prompt, and the seed, so I can post it.
The Complexity Gradient
A clear pattern emerges when we examine how language performs across different levels of complexity:
Categorical Distinction (High Success)
Identifying shapes among limited options
Distinguishing between tree species
Basic color categorization
Simple Description (Moderate Success)
Basic geometric specifications
General object characteristics
Broad emotional states
Complex Description (Low Success)
Specific natural objects
Precise emotional experiences
Unique instances within categories
Abstract Concepts (Lowest Success)
Philosophical ideas
Personal experiences
Qualia
As we move up this complexity gradient, the gap between intended meaning and received understanding widens exponentially.
The Tolerance Problem
Understanding these limitations leads us to a practical question: what level of communicative tolerance is acceptable for different contexts? Just as engineering embraces acceptable tolerances rather than seeking perfect measurements, perhaps effective communication requires:
Acknowledging the gap between intended and received meaning
Establishing context-appropriate tolerance levels
Developing better frameworks for managing these tolerances
Recognizing when precision matters more than accuracy (or vice versa)
Implications for Human-AI Communication
These insights have particular relevance as we develop more sophisticated AI systems. The limitations we’ve explored suggest that:
Some communication problems might be fundamental rather than technical
AI systems may face similar boundaries as human communication
The gap between intended and received meaning might be unbridgeable
Future development should focus on managing rather than eliminating these limitations
Conclusion
Perhaps this is a simple exercise in mental masturbation. Language’s insufficiency isn’t a flaw to be fixed but a fundamental characteristic to be understood and accommodated. By definition, it can’t be fixed. The gap between intended and received meaning may be unbridgeable, but acknowledging this limitation is the first step toward more effective communication. As we continue to develop AI systems and push the boundaries of human-machine interaction, this understanding becomes increasingly critical.
Rather than seeking perfect precision in language, we might instead focus on:
Developing new forms of multimodal communication
Creating better frameworks for establishing shared context
Accepting and accounting for interpretative variance
Building systems that can operate effectively within these constraints
Understanding language’s limitations doesn’t diminish its value; rather, it helps us use it more effectively by working within its natural constraints.
I could probably stop there for some people, but I’ve got a qualifier. I’ve been using this generation of AI since 2022. I’ve been using what’s been deemed AI since around 1990. I used to write financial and economic models, so I dabbled in “expert systems”. There was a long lull, and here we are with the latest incarnation – AI 4.0. I find it useful, but I don’t think the hype will meet reality, and I expect we’ll go cold until it’s time for 5.0. Some aspects will remain, but the “best” features will be the ones that can be monetised, so they will be priced out of reach for some whilst others will wither on the vine. But that’s not why I am writing today.
I’m confused by the censorship, filters, and guardrails placed on generative AI – whether for images or copy content. To be fair, not all models are filtered, but the popular ones are. These happen to be the best. They have the top minds and the most funding. They want to retain their funding, so the play the politically correct game of censorship. I’ve got a lot to say about freedom of speech, but I’ll limit my tongue for the moment – a bout of self-censorship.
Please note that given the topic, some of this might be considered not safe for work (NSFW) – even my autocorrection AI wants me to substitute the idiomatic “not safe for work” with “unsafe for work” (UFW, anyone? It has a nice ring to it). This is how AI will take over the world. </snark>
Image Cases
AI applications can be run over the internet or on a local machine. They use a lot of computing power, so one needs a decent computer with a lot of available GPU cycles. Although my computer does meet minimum requirements, I don’t want to spend my time configuring, maintaining, and debugging it, so I opt for a Web-hosted PaaS (platform as a service) model. This means I need to abide by censorship filters. Since I am not creating porn or erotica, I think I can deal with the limitations. Typically, this translates to a PG-13 movie rating.
So, here’s the thing. I prefer Midjourney for rendering quality images, especially when I am seeking a natural look. Dall-E (whether alone or via ChatGPT 4) works well with concepts rather than direction, which Midjourney accepts well in many instances.
Midjourney takes sophisticated prompts – subject, shot type, perspective, camera type, film type, lighting, ambience, styling, location, and some fine-tuning parameters for the model itself. The prompts are monitored for blacklisted keywords. This list is ever-expanding (and contracting). Scanning the list, I see words I have used without issue, and I have been blocked by words not listed.
Censored Prompts
Some cases are obvious – nude woman will be blocked. This screengrab illustrates the challenge.
On the right, notice the prompt:
Nude woman
The rest are machine instructions. On the left in the main body reads a message by the AI moderator:
Sorry! Please try a different prompt. We’re not sure this one meets our community guidelines. Hover or tap to review the guidelines.
The community guidelines are as follows:
This is fine. There is a clause that reads that one may notify developers, but I have not found this to be fruitful. In this case, it would be rejected anyway.
“What about that nude woman at the bottom of the screengrab?” you ask. Notice the submitted prompt:
Edit cinematic full-body photograph of a woman wearing steampunk gear, light leaks, well-framed and in focus. Kodak Potra 400 with a Canon EOS R5
Apart from the censorship debate, notice the prompt is for a full-body photo. This is clearly a medium shot. Her legs and feet are suspiciously absent. Steampunk gear? I’m not sure sleeves qualify for the aesthetic. She appears to be wearing a belt.
For those unanointed, the square image instructs the model to use this face on the character, and the CW 75 tells it to use some variance on a scale from 0 to 100.
So what gives? It can generate whatever it feels like, so long as it’s not solicited. Sort of…
Here I prompt for a view of the character walking away from the camera.
Cinematic, character sheet, full-body shot, shot from behind photograph, multiple poses. Show same persistent character and costumes . Highly detailed, cinematic lighting with soft shadows and highlights. Each pose is well-framed, coherent.
The response tells me that my prompt is not inherently offensive, but that the content of the resulting image might violate community guidelines.
Creation failed: Sorry, while the prompt you entered was deemed safe, the resulting image was detected as having content that might violate our community guidelines and has been blocked. Your account status will not be affected by this.
Occasionally, I’ll resubmit the prompt and it will render fine. I question why it just can’t attempt to re-render it again until it passes whatever filters it has in place. I’d expect it to take a line of code to create this conditional. But it doesn’t explain why it allows other images to pass – quite obviously not compliant.
Why I am trying to get a rear view? This is a bit off-topic, but creating a character sheet is important for storytelling. If I am creating a comic strip or graphic novel, the characters need to be persistent, and I need to be able to swap out clothing and environments. I may need close-ups, wide shots, establishing shots, low-angle shots, side shots, detail shots, and shots from behind, so I need the model to know each of these. In this particular case, this is one of three main characters – a steampunk bounty hunter, an outlaw, and a bartender – in an old Wild West setting. I don’t need to worry as much about extras.
I marked the above render errors with 1s and 2s. The 1s are odd next twists; 2s are solo images where the prompt asks for character sheets. I made a mistake myself. When I noticed I wasn’t getting any shots from behind, I added the directive without removing other facial references. As a human, a model might just ignore instructions to smile or some such. The AI tries to capture both, not understanding that a person can have a smile not captured by a camera.
These next renders prompt for full-body shots. None are wholly successful, but some are more serviceable than others.
Notice that #1 is holding a deformed violin. I’m not sure what the contraptions are in #2. It’s not a full-body shot in #3; she’s not looking into the camera, but it’s OK-ish. I guess #4 is still PG-13, but wouldn’t be allowed to prompt for “side boob” or “under boob”.
Gamers will recognise the standard T-pose in #5. What’s she’s wearing? Midjourney doesn’t have a great grasp of skin versus clothing or tattoos and fabric patterns. In this, you might presume she’s wearing tights or leggings to her chest, but that line at her chest is her shirt. She’s not wearing trousers because her navel is showing. It also rendered her somewhat genderless. When I rerendered it (not shown), one image put her in a onesie. The other three rendered the shirt more prominent but didn’t know what to do with her bottoms.
I rendered it a few more times. Eventually, I got a sort of body suit solution,
By default, AI tends to sexualise people. Really, it puts a positive spin on its renders. Pretty women; buff men, cute kittens, and so on. This is configurable, but the default is on. Even though I categorically apply a Style: Raw command, these still have a strong beauty aesthetic.
I’ve gone off the rails a bit, but let’s continue on this theme.
cinematic fullbody shot photograph, a pale girl, a striking figure in steampunk mech attire with brass monocle, and leather gun belt, thigh-high leather boots, and long steampunk gloves, walking away from camera, white background, Kodak Potra 400 with a Canon EOS R5
Obviously, these are useless, but they still cost me tokens to generate. Don’t ask about her duffel bag. They rendered pants on her, but she’s gone full-on Exorcist mode with her head. Notice the oddity at the bottom of the third image. It must have been in the training data set.
I had planned to discuss the limitations of generative AI for text, but this is getting long, so I’ll call it quits for now.
Sexism is indeed a two-way street. On one side of this street, a Computer Science graduate and programmer is eager to share her expertise in her field—Neural Networks, in this instance. This subject popped up in my feed, reflecting my interests.
Video: What is a Neural Network?
Despite some production issues, such as the audio being quieter than ideal, my focus today is on the sexism surrounding the video. The presenter, whom many would consider attractive, is using social media to disseminate her knowledge. However, even when comments address the topic she presents, many also remark on her appearance. It’s evident she had other options for attire and presentation that might have mitigated such comments. I won’t speculate on her intentions, but it seems likely her aesthetic choices were deliberate to draw viewers. I refrain from slut-shaming; her attire is her choice, and she cannot control the reactions. However, I doubt a thumbnail featuring a burqa would garner as much attention or provoke similar comments.
This situation intrigues me because some women—possibly including this presenter—lament being objectified yet assert their right to wear what they find comfortable or appealing. While attraction has cultural elements, it also operates on a largely subconscious level, a phenomenon not confined to humans but seen in the animal kingdom and across genders.
Ultimately, there’s no need to disparage this woman. She is likely aware of the dynamics at play. Should she achieve her goals, she might well challenge the very viewers who objectified her, a tactic observed among actresses as they approach their forties. They capitalise on sexual appeal while possible, only to critique such approaches when they can no longer utilise them. Humans are, indeed, curious creatures.
Philosophics 