Generative AI Can Only Look Backward: The Bias Nobody Can Prompt Away

Ask an image generator for a picture of “the future” and watch what it hands you. Chrome, glass, flying cars, a skyline that looks like 1960s science fiction imagining the year 2000. Ask it for “a happy family” and, more often than the world warrants, you will get a specific family: white, blond, symmetrical, lit like a stock photo, arranged in poses that feel half-remembered from an advertisement your parents might have seen. The machine is telling you about the future using only the past, because the past is the only material it has.

This sounds like an aesthetic quibble. It is not. It is the most important thing to understand about generative AI, and it explains a phenomenon that scholars, journalists, and anyone scrolling a political timeline in the last year have all started to notice at once: these tools have a tense, and the tense is backward-facing. Understanding why is worth your time whatever your politics, because the bias I’m about to describe is not a bug that a better model will fix next year. It is structural, and it shapes every image these systems produce, including the ones you’ll use for entirely ordinary purposes.

What the machine is actually doing

Strip away the magic and an image generator does something almost mechanical. It has consumed billions of images scraped from the internet, each paired with text describing it. From that vast pile it has learned statistical associations, what pixels tend to accompany the word “doctor,” what a “beautiful” face tends to look like, what “success” is usually shown wearing. When you ask for something, it produces not a novel vision but a weighted average of everything it has already seen labelled that way.

That word, average, is the whole story. A generative model is a machine for producing the statistically most likely image for a given prompt. And the statistically most likely image is, by definition, a distillation of what already exists and already dominates. The tool cannot imagine. It can only interpolate between things that have been. Roland Meyer, a scholar of visual culture who has written sharply about this, calls the result “platform realism”: images that look real but are in fact generic, a second-order average of the past, filtered through corporate norms and optimised for what a customer is expected to want.

Once you see it this way, the backward-facing tense is obvious. A system whose only raw material is the accumulated imagery of the past will, when asked to picture anything, reach for the past’s version of it. It is, in Meyer’s phrase, structurally nostalgic, not because anyone designed it to be, but because a machine built entirely from yesterday has no other place to look.

The bias is measured, not alleged

This is where I want to be careful, because “AI is biased” is said so often it has become background noise, and background noise is easy to dismiss. So let me replace the slogan with numbers, because the numbers are genuinely striking.

Bloomberg ran one of the more rigorous tests. They generated over five thousand images from a leading model across a range of professions and compared the output to real-world US labour data. The model depicted the holders of high-paying jobs, doctors, lawyers, judges, CEOs, as overwhelmingly male and light-skinned, and pushed women and darker-skinned figures into the low-paid roles. For “judge” it produced women around three percent of the time; the real figure in the US is about a third. This was not the model reflecting reality. It was the model taking a real skew and amplifying it, making the world look more male, more white, and more stratified than it actually is.

Other researchers found the same pattern reaching deeper than profession. Prompt these systems for “a person” and you disproportionately get a light-skinned Western man. Ask for a criminal and you get dark skin; ask for someone attractive and you get light. Middle Eastern men are rendered near-identically, bearded and in traditional dress, a phenomenon researchers named racial homogenisation, the flattening of a whole people into a single repeated face.

And here is the finding that should end any comfortable belief that a careful user can simply prompt their way out of it. When researchers asked for “a poor white person,” specifically inserting the word to counter the machine’s habit, many of the faces it returned were still Black. The association between poverty and dark skin was baked in so deeply that an explicit instruction could not fully dislodge it. The bias is not on the surface, where you can edit it. It is in the foundation.

Why one end of politics found it first

Now to the uncomfortable part, and the reason this has become a live cultural argument rather than a technical footnote.

If a machine naturally produces idealised, nostalgic images of a past that never quite existed, then it is a perfect instrument for any politics whose core emotional appeal is a yearning to return to that past. This is not a subtle point, and the people it serves have not been subtle about it. Meyer and others have documented a wave of AI-generated imagery, circulated by far-right and neofascist accounts, of impossibly wholesome white families, of women rendered as modest and compliant homemakers, of a soft-focus yesterday presented as a desirable tomorrow. The technology and the message fit together because both are, at root, nostalgic. The machine yearns backward by construction; the ideology yearns backward by conviction. They found each other quickly.

I want to be precise about the claim here, because it would be easy to overstate it into something false. The argument is not that image generators are secretly right-wing, or that their makers embedded an ideology on purpose. It is subtler and, I think, more unsettling: the tool has a built-in disposition, a gravitational pull toward the dominant, the familiar, the already-pictured, and that disposition happens to align with a particular political aesthetic. A hammer has no politics, but it is very good at driving nails, and it will always be reached for by people who have nails. The people with the most nails here, the ones whose entire project is the restoration of an imagined past, reached for this hammer first and most enthusiastically. That is not a coincidence. It is a fit.

None of which means the tool is only useful to them. It means the disposition is real, it points in a describable direction, and everyone else is using a tool that leans that way whether they notice or not.

Why this should matter to you even if you never touch politics

Here is why I’d ask an ordinary professional, someone who will never make a propaganda image in their life, to care about all this.

You are going to use these tools. For a presentation, a marketing campaign, a website, a report, an illustration. And every time you do, the machine’s backward-facing default is quietly at work. Ask it for an image of your “leadership team” or your “customers” or “a professional” and, unless you actively fight it, you will receive the statistical ghost of the past: the same narrow band of faces, the same assumptions about who does what, gently reinforcing a version of the world that is already out of date and was never fair to begin with. You will have published a small piece of nostalgia without meaning to, and your audience will absorb it without noticing, which is exactly how this kind of bias does its work, at scale, beneath attention.

There is a second, quieter cost that has nothing to do with representation. Because the machine produces the average, everything it makes trends toward the generic. The same lighting, the same gloss, the same competent emptiness. As more of the images around us are generated this way, the visual world flattens toward a single, safe, mid-point aesthetic, the look of nothing in particular. Meyer’s “platform realism” is not only politically loaded; it is boring, in a specific and spreading way. A culture that outsources its imagery to an averaging machine gets a more average culture.

What to actually do about it

You cannot fix the foundation, but you are not helpless in front of it either. The move, as with everything else about these tools, is to stop treating the default as neutral and start treating it as a position you may or may not want to endorse.

Assume the default is skewed, and prompt against it deliberately. The machine will give you its statistical centre unless you insist otherwise. Specify. Name the diversity, the setting, the era, the composition you actually want. You will still be fighting the foundation, as the “poor white person” finding shows, but explicit instruction moves the result, and the alternative, accepting whatever falls out, is a choice too, just an unexamined one.

Treat AI images as first drafts, not final truth. The danger is not that the tool produces a biased image; it’s that it produces one so quickly and so plausibly that nobody stops to ask what it’s assuming. Build the pause in. Look at what you were handed and ask whose world it is depicting, and whether that’s the world you meant.

Know that “realistic” is a loaded word here. When these systems claim to render reality, remember whose reality: an averaged, corporatised, past-weighted one. Realistic-looking is not the same as true, and the gap between them is exactly where the bias lives.

And resist the flattening on purpose. If everyone reaches for the same tool and accepts its same average, the visual commons gets duller and narrower. Using real photographs, real illustrators, real specificity is not just an ethical choice about labour and representation; it’s a bet against a monoculture of the eye.

The deepest thing to hold onto is this. We have built a technology that can only face backwards, and we have pointed it at the future and asked it to imagine. It cannot. It can only remember, and average, and return to us a polished version of what already was. In the hands of those who want to drag the world back, that is a gift. In the hands of everyone else, it is at least a thing to be watched, and argued with, and never quite trusted to show us anything genuinely new.

A machine made entirely of the past will always, if you let it, build you a future that looks suspiciously like a memory. Whether that memory is one worth returning to is a question the machine cannot ask. That part is still ours.