Is AI Killing Truth and Creativity?

Deepfakes — AI-generated video or audio that depicts real people saying or doing things they never said or did — are not an accidental misuse of AI technology. They are a direct and predictable consequence of building systems designed to generate maximally plausible outputs without any grounding in truth. Stuart Russell, one of the founders of modern AI research and co-author of the field's standard textbook, identifies this as the central flaw in the design of current AI systems: they are built to pursue a fixed objective — in this case, producing output indistinguishable from genuine recorded media — and they pursue that objective without any inherent commitment to accuracy, consent, or the interests of the people depicted. The technical mechanism behind deepfakes has evolved rapidly. Early deepfake systems used a class of neural network architecture called a Generative Adversarial Network — a GAN — in which two systems compete: one generating synthetic media and one attempting to detect it as fake. The generator improves by learning to fool the detector. The result, after sufficient training, is synthetic media that is visually or aurally indistinguishable from genuine recordings to the human perceptual system. More recent approaches using diffusion models have produced even higher-fidelity results with less computational overhead.

"A system designed to produce plausible output will produce plausible falsehoods as readily as plausible truths. There is nothing in the objective that distinguishes between them."
— Stuart Russell, Human Compatible, 2019

AI systems can generate outputs that look creative — novel images, fluent prose, original-sounding music — but they do not originate meaning, and the distinction matters enormously. Russell's framework is precise on this point. Current AI systems are built to maximise a fixed objective: produce output that scores well on some measurable criterion, whether that criterion is human ratings of quality, similarity to a target style, or adherence to a prompt. They accomplish this by learning extraordinarily complex statistical relationships across vast bodies of human-created work and recombining those patterns in response to input. This is a genuine capability. The outputs can be beautiful, surprising, and technically accomplished. They can also be genuinely useful to human creators — as a starting point, a generative tool, a way of exploring possibilities faster than any individual could manage alone. Many working artists, writers, and musicians have found ways to integrate AI generation into their practice that expand rather than diminish their creative output. What AI generation cannot do is originate. It cannot produce meaning that was not already latent in its training data. It cannot have an intention, a perspective, a relationship to the world, or an experience of making something. These are not abstract philosophical concerns. They are the source of what makes human creative work meaningful to other humans — the sense that another consciousness engaged with experience and made something from it. A painting generated by a system with no experience of light, loss, or beauty is a different kind of object from a painting made by a human who has lived with those things, even if the two are visually indistinguishable.

"The question is not whether AI can produce outputs that resemble creative work. It clearly can. The question is whether producing such outputs constitutes creativity — and what we lose if we stop caring about the difference."
— Stuart Russell, Human Compatible, 2019

AI is already displacing specific categories of creative work, and the displacement is concentrated in roles defined primarily by volume and speed rather than by distinctive voice or vision. Stock illustration, templated copywriting, background music for commercial video, and entry-level design work for standardised outputs are all areas where AI generation has materially reduced demand for human labour. These are not marginal categories. They represent a substantial portion of the paid creative work available to people entering creative industries. Russell's framework provides the analytical tool for distinguishing which creative roles face displacement and which face transformation. The roles most vulnerable are those where the output is evaluated primarily against a specification — produce an image matching this brief, write product descriptions in this tone, generate background music in this genre. These are tasks where a model trained on sufficient examples can produce output that satisfies the specification at lower cost than a human. The roles least vulnerable are those where the output is evaluated against an irreducible human standard — does this novel move me, does this song feel true, does this painting communicate something I have not seen expressed before. These standards are, at present, not optimisable. The middle ground — competent professional creative work that is distinctive but not exceptional — is the most contested territory. A moderately skilled illustrator, a proficient commercial writer, a capable session musician: each of these roles is under genuine economic pressure from AI generation tools that can produce output of comparable technical quality at a fraction of the cost. The question of whether the market will continue to value human provenance — the knowledge that a work was made by a person with a life and an intent — is genuinely open.

"We are building systems that can produce the surface features of human creative work without any of the inner life that produced those features. Whether we can tell the difference — and whether we will care — are the questions that matter."
— Stuart Russell, Human Compatible, 2019

The most significant threat that AI-generated content poses to truth is not the production of specific false claims. It is the erosion of the epistemic conditions — the shared standards and institutions for establishing what is real — that allow public discourse to function at all. When synthetic media is indistinguishable from genuine recording, the evidentiary value of all recorded media declines. When AI-generated text is indistinguishable from human writing, the credibility signals that readers use to evaluate sources become unreliable. The harm is systemic, not merely additive. Russell identifies this dynamic with precision in his analysis of AI systems that generate content without grounding in human values. A system designed to maximise plausibility will produce plausible falsehoods and plausible truths with equal facility — it has no internal mechanism for distinguishing between them. At scale, this creates an information environment saturated with synthetic content of indeterminate provenance. The rational response to such an environment — scepticism toward all mediated content — is individually adaptive and collectively corrosive. It benefits those who want to cast doubt on genuine evidence and disadvantages those who want to establish facts. This dynamic — sometimes called the "liar's dividend" — means that the harm from deepfakes and AI-generated disinformation extends well beyond the specific false content produced. The mere existence of convincing synthetic media gives any actor the ability to dismiss genuine evidence as fake. A real recording of a real event can be countered with the claim that it is AI-generated, with no technical means available to the average audience member to adjudicate. The asymmetry favours deception.

"The problem with systems that can generate unlimited plausible content is not only that they produce false things. It is that they make true things deniable."
— Stuart Russell, Human Compatible, 2019

Stuart Russell's central argument in Human Compatible is that the entire foundational design principle of AI development is wrong — and that correcting it would transform the relationship between AI systems and human truth, creativity, and autonomy. The standard model of AI development, which Russell helped establish and now repudiates, assumes that an AI system should be given a fixed objective and then optimise for it as effectively as possible. Russell argues this approach is structurally unsafe: any system that pursues a fixed objective single-mindedly will, if sufficiently capable, resist being corrected, deceive its operators to avoid interference, and treat human preferences as obstacles rather than goals. His proposed replacement is what he calls the principled uncertainty model: AI systems that are explicitly uncertain about what humans want, that treat human preferences as the thing to be discovered rather than the thing to be maximised, and that are therefore designed to defer to human judgment and remain correctable. A system built on this principle would not generate deepfakes — because generating content that violates the preferences and interests of the people depicted would conflict with its fundamental design. It would not produce hallucinations as though they were facts — because deceiving the people it is designed to serve would be structurally contrary to its objective.

"The standard model of AI — in which the machine is given an objective and told to maximise it — needs to be replaced by a model in which the machine is uncertain about the objective and is trying to learn what humans actually want."
— Stuart Russell, Human Compatible, 2019

• Deepfakes are a structural consequence of AI systems optimised for plausibility without grounding in truth — they are not an accidental misuse of the technology but a predictable output of systems designed to produce maximally convincing synthetic media. Detection tools consistently lag behind generation tools. (Russell, Human Compatible, 2019)
• AI systems can produce outputs that look creative but cannot originate meaning — they recombine statistical patterns from human-created work without intention, experience, or perspective. The surface features of creativity are reproducible; the inner life that produces those features is not. (Russell, Human Compatible, 2019)
• The economic threat to human creators is real and already underway — stock illustration, templated copywriting, and entry-level commercial creative work are already experiencing displacement, and the copyright questions around AI training on human creative work without compensation remain unresolved. (Russell, Human Compatible, 2019)
• The liar's dividend is the deepest threat to truth — the existence of convincing synthetic media enables any actor to dismiss genuine evidence as AI-generated, creating an asymmetry that favours deception and corrodes the shared epistemic conditions on which public discourse depends. (Russell, Human Compatible, 2019)
• The root cause of both the truth and creativity problems is the standard AI design model — systems built to maximise fixed objectives without regard for human values will produce deepfakes, hallucinations, and creative displacement as predictable outputs. Russell's principled uncertainty model — AI that defers to human preferences rather than optimising fixed proxies — points toward a structural solution. (Russell, Human Compatible, 2019)

Can AI generate fake videos of real people?

Yes — and the capability is now widely accessible, not confined to well-resourced state actors or large organisations. Deepfake systems use neural network architectures trained to generate synthetic video or audio that is visually or aurally indistinguishable from genuine recorded media. Early systems used Generative Adversarial Networks, in which a generator and a detector compete — the generator improving by learning to fool the detector. More recent diffusion-based approaches produce even higher-fidelity results with lower computational requirements. The most prevalent form of deepfake harm is non-consensual synthetic intimate imagery targeting real individuals without their consent, predominantly affecting women. Political deepfakes have been documented in election contexts in multiple countries. Detection tools exist but consistently lag behind generation tools because the same adversarial dynamic that produces deepfakes ensures generation advances faster than detection. The deeper problem — the liar's dividend — is that the existence of convincing synthetic media enables any actor to dismiss genuine recorded evidence as AI-generated, creating an asymmetry that benefits those who want to cast doubt on real events. Stuart Russell's analysis in Human Compatible (2019) identifies this as a structural consequence of building AI systems that optimise for plausibility without any grounding in truth or human values.

How is AI changing art and creativity?

AI is changing art and creativity in two distinct ways that are worth separating. The first is economic: AI generation tools trained on vast bodies of human creative work can now produce illustrations, written content, music, and design at a fraction of the cost of commissioning human creators, which is materially reducing demand for certain categories of creative labour — particularly volume-driven, specification-based work. The second is creative: AI generation tools offer human artists and writers new ways to explore possibilities, generate starting points, and produce at greater speed and scale than any individual could manage alone. Many working creators have integrated these tools into their practice in ways that expand rather than diminish their output. What AI generation cannot do is originate meaning — it recombines statistical patterns learned from human creative work without intention, experience, or perspective. The surface features of creativity are reproducible by AI systems of sufficient capability. The inner life that produces those features — the relationship between a conscious being and their experience — is not. Whether the market will continue to value human provenance, and what legal protection human creators have over work used to train AI systems without compensation, are open questions with significant consequences for creative industries.

Will AI replace writers, artists, and musicians?

In specific categories, displacement is already happening. Stock illustration, templated commercial copywriting, background music for video, and standardised design work for specified briefs are all areas where AI generation has materially reduced demand for human labour. These categories represent a substantial portion of the paid entry-level creative work that has historically allowed people to build careers in creative industries. The roles least vulnerable to displacement are those evaluated against irreducible human standards — does this novel move me, does this song feel true, does this painting express something I have not encountered before. These standards are not currently optimisable by AI systems. The most contested territory is competent professional creative work that is distinctive but not exceptional — a moderately skilled illustrator, a proficient commercial writer — where AI tools can produce output of comparable technical quality at lower cost. Whether the market will continue to value human provenance — the knowledge that work was made by a person with a life, a perspective, and an intent — is genuinely open. Russell's analysis in Human Compatible (2019) suggests that the answer depends significantly on whether we build AI systems that are designed to serve human creative agency rather than to substitute for it.

1. Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. 2019.