Words and Pictures

When we say an AI model "hallucinates," we are not saying it is conscious, intoxicated, spiritually expanded, mentally ill, or having a private experience. It has no inner theatre. No pink elephants. No cosmic insects. No Terence McKenna machine elves sitting behind the graphics card eating luminous fruit. The term is metaphorical. But it is not useless. The metaphor survives because generative AI also produces convincing appearances without secure attachment to reality. It makes things up. Sometimes it does so badly, producing errors that are easy to spot. Sometimes it does so beautifully, producing fictions that appear calm, coherent, and almost inevitable.

A hallucination is a perception without an object: a voice without a speaker, a figure without a body, a sensation without an external cause. It feels real because the mind supplies the evidence. When a person hallucinates, the world does not simply look strange; it appears to contain something that is not there. The experience may be frightening, beautiful, absurd, devotional, comic, or chemically assisted. It may arrive through illness, exhaustion, sensory deprivation, grief, fever, or the bright weather of psychedelics. But in every case, hallucination troubles the boundary between perception and invention. Something generated internally is experienced as if it had come from outside.

In language models, hallucination is usually treated as failure. The machine invents a quotation, fabricates a source, misnames a book, misremembers a fact, or answers with great confidence from a place of no knowledge. This is not creativity; it is falsehood wearing a good suit. The problem is not that the language is strange, but that it sounds trustworthy. A hallucinated answer can be grammatically polished, tonally fluent, even persuasive. It can have the surface of authority without the structure beneath it. For factual systems, this matters. If the task is to retrieve truth, then hallucination is a breach of trust.

Image generation complicates the matter. A synthetic image is not usually asked to report a fact. It begins outside documentary evidence. A synthograph of a glass flower, an imaginary cabinet, a fictional portrait, or a room that never existed is not automatically wrong because no such thing stood before a camera. In synthography, invention is not a malfunction; it is the medium. The question, then, is not simply whether the image is real. It is whether the unreality has been shaped with enough intelligence to persuade. Hallucination in image generation is not merely the appearance of something impossible. It is the production of visual information that has no secure origin, no lived referent, and often no understanding of the thing it depicts.

This is why the older signs of AI hallucination were so easy to recognise. Extra fingers, melting faces, broken teeth, nonsense text, fused objects, impossible anatomy, waxy skin, dead eyes: these were not subtle errors. They announced themselves like bad actors entering too early. Earlier image models often seemed visibly intoxicated, as if the world had been poured into the machine and stirred with a dirty spoon. They could produce beauty, but they also produced wreckage. Hands became spiders. Furniture grew limbs. Words dissolved into decorative insects. The hallucination was not hidden; it was written across the surface.

The newer models are more interesting because the hallucination has become quieter. GPT-5.5 + Images.2, Flux.2, and their neighbouring systems are far more visually fluent than the tools many of us were using only a year ago. They understand light better, surfaces better, anatomy better, composition better. They are better at following correction. Better at suppressing obvious chaos. But this does not mean the hallucination has disappeared. It has learned manners. It now arrives dressed as plausibility. The error is no longer always an extra hand or a melted face. It may be a handle that cannot be pulled, a hinge that could never open, a clasp with no purpose, a strip of timber that looks attached but not joined. The image seems reasonable until one asks how it works.

This became especially clear to me while developing Arca II, a synthographic project about fictional cabinets made from reclaimed timber, old floorboards, shed wood, sign fragments, rusted hinges, peeling paint, and occasional pieces of driftwood. The project was created through conversation with GPT-5.5 Thinking and generated and refined using ChatGPT + Images.2. Rather than writing a fixed prompt for Flux.2 and accepting whatever emerged, the work developed through looking, questioning, correction, and return. I asked for objects to be changed because they did not make sense. A handle needed to behave like a handle. A latch needed a reason to exist. A door needed hinges where hinges should be. A piece of wood needed to feel joined rather than pasted onto the front like a decorative accident.

This was not the elimination of hallucination. It was the disciplining of it. The machine continued to invent, but the invention became subject to judgement. Each correction brought the image closer to an object that could persuade the eye not only as a picture, but as a thing with weight, use, and intention. In that sense, working with generative AI began to feel less like ordering an image and more like sitting beside an imaginary carpenter at a bench. The carpenter was brilliant, tireless, and slightly deranged. It could produce a convincing cabinet from nothing, but it might also attach a latch to a surface where no latch could function. The artist's role was not to stop the hallucination entirely, but to decide which parts of the hallucination deserved to survive.

This is where the psychedelic analogy becomes useful again, provided we do not take it too literally. Timothy Leary understood that the experience of a trip was shaped by set and setting: the state of the mind, the surrounding environment, the conditions under which perception unfolds. Generative AI has its own version of this. The prompt, the model, the reference image, the conversation, the correction, the crop, the upscale, the artist's taste, the willingness to reject an almost-good result: all of these form the set and setting of the synthetic image. The machine may be hallucinating, but it is not hallucinating alone. The synthographer becomes a kind of trip-sitter, guiding the experience, calming the excess, redirecting the vision when it begins to mistake ornament for structure or plausibility for truth.

This is also why hallucination should not be understood only as a bug. In factual work, hallucination is dangerous because it confuses invention with knowledge. In image-making, the matter is stranger and more fertile. The synthetic image is born from a hallucinatory condition: language enters a model, probability becomes form, and an image appears where no scene, object, or body existed. That process is not a defect at the edge of the medium; it is part of the medium's nature. What matters is not whether the machine invents, but how the invention is handled. An undisciplined hallucination becomes slop. A disciplined hallucination may become art.

The danger now is not that AI images look obviously strange. That was the easy phase. The more serious question begins when the hallucination looks sensible, beautiful, useful, even true. A synthetic cabinet may appear handmade. A fictional portrait may seem observed. A generated flower may carry the authority of botanical study. A scene that never happened may present all the signals of having happened. The machine is still tripping, but the trip has become quieter, sharper, and more persuasive. It no longer always shows us monsters. Sometimes it shows us an ordinary object and asks us to believe in it.

Synthography lives inside that tension. It does not record the world, but it borrows the world's behaviours. It does not document presence, but it can simulate the signs by which presence is usually recognised. AI hallucination, then, is not simply an error to be corrected or a spectacle to be enjoyed. It is the unstable ground on which synthetic images are made. The task is to work there consciously: to look harder, refuse more often, correct with care, and understand that every convincing synthograph is a negotiated hallucination. The machine is still tripping. The artist decides what comes back from the trip.

For most of modern history, images have carried a quiet promise. However composed, cropped, staged, or manipulated they might be, they seemed to begin with something outside themselves. A painting could stylise, a drawing could interpret, a photograph could mislead, yet each still implied a relation to a world beyond the image. Pigment, graphite, silver salts, pixels: different materials, but the same underlying assumption. The image pointed back to something that had once stood before an eye, a hand, or a lens. This assumption did not guarantee truth, but it gave images a kind of inherited authority. They appeared to have an origin, and origin mattered. It suggested that what was being seen had, in some form, been there.

Synthetic image-making unsettles that contract. A synthographic image can present the appearance of photography, drawing, sculpture, or print, yet remain untethered from any object, body, or event that ever occupied space. No bowl of fruit, no studio, no sitter, no flower, no carved vessel, no sheet of metal need exist for the image to appear entirely convincing. The old chain between seeing and having-been-there begins to loosen. And yet the image often continues to persuade. It can still feel precise, coherent, materially intelligent, even evidential. This is the interesting part. The synthetic image does not simply sever the bond between image and world; it reveals that the bond was never as stable as it appeared. If belief survives the loss of origin, then perhaps origin was not the real basis of belief in the first place.

What, then, makes an image persuasive? Not truth, exactly. Images have never needed to be wholly true in order to be convincing. We know this instinctively. A film set can feel like a city. A painted apple can seem heavy enough to bruise. A fashion photograph can fabricate an atmosphere that never existed outside the frame. The viewer does not wait for proof before responding. Perception works faster than verification. It gathers cues, aligns them, and arrives at a judgement. Light falls in a plausible way. Surfaces behave consistently. Materials appear to possess the right density, texture, wear, and reflectance. Shadows agree with form. Scale feels coherent. The object sits in space with the expected confidence. Restraint helps. So does proportion. So does the absence of noise. When enough of these signals converge, conviction stabilises.

This is why synthetic images can persuade with such force. They do not need to be true; they need to present the right signals of truth. What the viewer grants is not factual endorsement but perceptual agreement. The image behaves correctly, and so belief forms. It is not a matter of gullibility. It is simply how seeing works. Human perception does not passively record the world; it predicts it. It makes rapid inferences from incomplete information, assembling certainty from coherence rather than from direct access to reality. A synthetic image enters this system not as an alien object but as something that has learned to speak its language fluently. It understands, statistically and visually, what wood should look like when burnished, what glass should do with light, how flesh should carry weight, how a shadow should anchor a stem against a black void. The image appears credible because it supplies the eye with what the eye expects.

In this sense, synthography does not invent deception so much as expose the mechanics of persuasion that have always been present in image-making. The synthetic image makes the process easier to see because it removes the comfort of origin. Once nothing has stood before the lens, nothing has been touched by a hand, and nothing has been witnessed in the world, the viewer can no longer appeal to capture as the source of authority. What remains is appearance itself. Surface, structure, behaviour, proportion, optical logic: these become the real grounds of conviction. A generated image of a glass tulip, a ceramic vessel, a male portrait, or a forged metal flower may never have existed materially, yet the eye responds to it with the same habits of reading it would bring to a photograph, an object, or a still life. The image does not need to prove itself; it needs to behave.

This has consequences beyond the medium itself. We are used to thinking of truth and fiction as opposites, but in visual culture the relationship is often more slippery. Fiction can be persuasive without being factual. Documentary can be misleading while remaining indexical. Advertising, cinema, fashion, editorial photography, museum display, architectural rendering, political image-making — all operate through selective constructions of reality. Synthetic images intensify this condition, but they do not create it from nothing. What they do create is a clearer test case. They force us to confront how much of our trust in images depends not on what happened, but on how correctly something appears to have happened. They reveal that visual certainty is often an effect of style, coherence, and surface discipline rather than of truth alone.

My own work has increasingly taken shape within this tension. I am less interested in fantasy for its own sake than in the authority of things that never existed. Across projects involving imagined ceramics, fabricated flowers, synthetic drawings, sculptural propositions, and photographic fictions, I return to the same question: how can an unreal thing acquire the weight of a real one? The answer is never simple realism. On the contrary, realism alone is often too blunt. What matters is material behaviour, restraint, proportion, and the quiet language of craftsmanship. A chrysanthemum in blackened steel persuades not because it is merely detailed, but because the petals seem heat-shaped, the stem seems hand-worked, the finish seems worn in the right places, and the whole object appears makeable. A synthetic portrait persuades not because a face is present, but because posture, light, skin, lens logic, and atmosphere align. Belief enters through coherence.

This is why I do not see artificiality as a flaw to be concealed. The synthetic condition is not an embarrassment; it is the material of the work. The goal is not to trick the viewer into thinking that the image has an origin it does not possess. It is to investigate what happens when an image acquires authority without one. In that sense, a synthograph is not a failed photograph, nor a substitute for an object, nor an imitation waiting to be forgiven. It is a different kind of proposition altogether: an image that asks what we are really responding to when we say that something looks real. It suggests that realism may be less about evidence than about agreement, less about fact than about the successful performance of visual truth.

We do not believe images because they are true. We believe them because they present the right signals of truth. Synthetic image-making does not weaken that insight; it clarifies it. In removing the world as guarantor, it makes the act of seeing more visible to itself. The question is no longer simply whether the image lies, but why its lie is so persuasive, and what that persuasion reveals about the habits of perception we call certainty. What synthography offers, at its best, is not the collapse of truth into fiction, but a sharper understanding of how visual belief is made.

Apples carry a lot of baggage. There is the apple in the Garden of Eden (temptation, knowledge, trouble); the poisoned apple in Snow White (beauty, vanity, collapse); Cézanne's apples (form, repetition, the slow dismantling of painting itself). There is the Big Apple, and there is Apple™ — a logo, a company, a device in your pocket. There are thousands of apple varieties worldwide, yet most of us encounter only a small handful in UK supermarkets. They are familiar, neutral, quietly perfect. We recognise them without thinking, which is precisely why apples are useful here. Because these are not apples. And this is not a photograph. What you are looking at is a synthograph — an image made using generative artificial intelligence. No camera. No fruit bowl. No light bouncing off real apples. Just language, probability, and intention, assembled into an image that behaves like a photograph. If they look convincing, that is the point. If they look boring, even better. Sometimes the most radical thing an image can do is sit calmly and refuse to announce itself. Synthography is a form of image-making that uses generative AI to create synthetic images rather than capture the world. Like photography, it can produce images that appear convincingly real. Unlike photography, it does not record light from a scene, nor does it document an event or a moment in time. It produces synthetic media: images that look like evidence while remaining unbound from any single, actual occurrence. In simple terms, a synthographic image is not taken; it is authored into being. It begins not with a camera and a subject, but with intention, articulated through language, structure, and constraint.

The word synthography follows the same linguistic structure as photography. Where photography means writing with light, from the Greek phōs, meaning light, and graphē or graphein, meaning writing, drawing, or inscription, synthography may be understood as writing with synthesis. The term combines synthesis — from the Greek sense of putting together or composition — with -graphy, a suffix associated with writing, drawing, recording, or describing. It names images written through synthetic generation rather than exposed by illumination: images assembled by computation rather than captured by a camera. The result may resemble a photograph — sometimes uncannily so — but its relationship to reality is fundamentally different. A photograph is indexical: it bears a physical trace of something that occurred. A synthograph does not. It is not evidence, nor a record of presence. It is a plausible artefact, designed to behave like an image we recognise while remaining untethered from the world it appears to depict. This distinction becomes clearer when compared with illustration and painting. Illustration is an act of direct translation: an image is intentionally drawn, painted, or rendered through deliberate mark-making. Even when mediated by software, the illustrator or painter constructs form through gesture, stroke, and explicit visual decisions, working materially on a surface. In synthography, the artist does not execute the image piece by piece. Instead, they define the conditions under which an image may emerge. Language initiates the process, systems generate possibilities, and the artist responds — selecting, refining, redirecting, or refusing outcomes. For this reason, synthography is sometimes described as painting with words: not because it resembles painting materially, but because it shares a similar posture of indirect control, where intention shapes form without dictating it outright. The image is not executed; it is negotiated.

What a synthographic image resolves from is not a scene, an object, or a memory, but a latent space: an abstract field shaped by patterns learned from training data drawn from vast numbers of images. This space does not contain pictures as such, but compressed visual possibilities — tendencies, correlations, and statistical relationships without fixed referents. When an image is generated, it emerges gradually, as structured noise resolves into form under conditions set by language, reference, and probability. The image feels photographic because it inherits the visual grammar of photography, yet it depicts no moment that was ever witnessed by a lens. What becomes visible is not a captured instant, but a synthetic probability made legible. In practice, synthography most often begins as text-to-image: written language used to initiate an image. From there, images are refined through selection and iteration, sometimes guided by existing images in image-to-image workflows or reference-based processes that steer composition, material, or tone. These approaches form a continuum of control rather than discrete stages, allowing the artist to move between broad instruction and precise adjustment. The tools vary — including systems such as DALL·E, Stable Diffusion, Midjourney, and Flux — but the underlying principle remains constant: language initiates a field of synthetic possibility; judgement shapes outcome. For this reason, "AI art" is an imprecise label. It names a technology rather than a practice. Synthography instead names a specific image-making medium, defined by process, decision-making, and intent.

Synthography is collaborative by nature, but it is not automated. The machine generates images without understanding meaning, coherence, context, or restraint. Those responsibilities remain human. The artist works through writing prompts, setting parameters, refining direction, and evaluating results. Authorship resides in selection, refusal, iteration, and curation — and, increasingly, in the craft of post-production, where synthetic images are edited, stabilised, and made consistent with a larger body of work. This distinguishes synthography from novelty-driven generation or one-click production. Like photography, illustration, or printmaking, it rewards patience and discipline. Images are tested, rejected, adjusted, and contextualised; series are constructed over time rather than accumulated. Generative image-making did not appear overnight. It developed gradually through decades of research in computer vision, procedural art, neural networks, and machine learning, long before it entered public awareness. What has changed in recent years is not the existence of generative systems, but their accessibility and visual fluency. The current moment marks a point of maturity, where synthetic image generation has become stable enough to support sustained artistic practice rather than isolated experimentation. As a result, synthography is already in active use across the creative industries — in fashion, advertising, editorial, and brand-led image-making — where precision, consistency, and cultural literacy matter as much as visual impact. In these contexts, generative AI functions not as an autonomous artist, but as a studio instrument: a camera for ideas, not for light. To use an analogy, synthography can be understood as a form of digital alchemy: where alchemists once sought to turn base matter into gold, synthographers transform language into images.