{"id":1404,"date":"2026-10-21T10:12:01","date_gmt":"2026-10-21T10:12:01","guid":{"rendered":"https:\/\/blog.franke-ip.com\/en\/?p=1404"},"modified":"2026-10-21T10:16:19","modified_gmt":"2026-10-21T10:16:19","slug":"creative-but-not-an-author-what-ai-can-do-in-art-and-what-the-law-says","status":"publish","type":"post","link":"https:\/\/blog.franke-ip.com\/en\/creative-but-not-an-author-what-ai-can-do-in-art-and-what-the-law-says\/","title":{"rendered":"Creative, but Not an Author: What AI Can Do in Art \u2013 and What the Law Says"},"content":{"rendered":"
\n
\"\"<\/a><\/figure>\n<\/div>\n\n\n

Can a machine be an artist? The debate is usually fought between two camps. One holds that human creativity, too, is nothing more than the clever recombination of what has been learned, and sees AI well on its way to replacing designers, musicians, actors and writers. The other counters that an AI can never leave the tracks laid out for it. Neither position fully survives critical scrutiny. And while cultural commentators and the tech industry are still arguing, courts and legislators have long since given an answer of their own.<\/em><\/p>\n\n\n\n

The brain as a language model? A seductive analogy<\/h2>\n\n\n\n

The strongest argument of the AI optimists comes from neuroscience. Creativity researchers such as Roger Beaty show that\u00a0remembering and imagining draw on the same mechanism: the flexible recombination of experience \u2013 people, places, events. Geoffrey Hinton, Nobel laureate and one of the fathers of deep learning, goes further. Humans, he argues, are\u00a0above all analogy machines, with only a thin layer of logical reasoning on top.<\/p>\n\n\n\n

If that is true, the difference between a person drawing on their store of experience and a model drawing on its training data is one of degree rather than of kind. Hinton concludes that large language models are\u00a0highly creative\u00a0\u2013 precisely because the only way they can squeeze vast amounts of knowledge into their networks is by recognising analogies between the most diverse things.<\/p>\n\n\n\n

The analogy has a gap, however, and it lies exactly where things get interesting. The same research shows that merely retrieving memories rarely produces the most original ideas. Creative achievement arises from the interplay of associative memory and cognitive control \u2013 the ability to deliberately discard the obvious. Anyone who equates the brain with a language model leaves out this second half. And it is precisely the half where today\u2019s models measurably fall short.<\/p>\n\n\n\n

\u201cCountless loops\u201d: the missing-feedback objection<\/h2>\n\n\n\n

One objection that often comes up in discussions about AI creativity deserves closer scrutiny: the brain, it is said, has countless feedback loops, whereas AI merely performs a simple one-way \u201cdecoding\u201d. The technical terms are feedback or recurrent processing. The objection contains a kernel of truth \u2013 though not quite where it is usually located.<\/p>\n\n\n\n

In the brain, feedback is the norm<\/h3>\n\n\n\n

The visual system alone shows how heavily the brain relies on feedback. The relay station in the thalamus that passes signals from the eye to the cerebral cortex receives only\u00a05 to 10 per cent of its synapses from the retina. Roughly a third each comes from the cortex itself and from the brainstem. The few retinal synapses are particularly powerful, but most connections serve to steer the flow of information from above.<\/p>\n\n\n\n

The Amsterdam neuroscientist Victor Lamme developed an influential theory of consciousness from findings like these. According to it, sensory information first passes through the brain areas in a rapid forward sweep. It is only consciously perceived once\u00a0higher areas feed signals back into earlier ones. Predictive coding theory goes further still: higher areas continuously send predictions downwards, and what is reported upwards is mainly the deviation from them.<\/p>\n\n\n\n

Creativity itself is also a cyclical process. An fMRI study by Canadian researchers showed that\u00a0generating ideas\u00a0mainly engages memory regions in the medial temporal lobe, whereas evaluating them engages a combination of the executive control network and the default network. Creative work oscillates between the two modes: generate, test, discard, start again.<\/p>\n\n\n\n

In AI: what \u201cdecoding\u201d actually means<\/h3>\n\n\n\n

On the AI side, \u201cdecoding\u201d is not the wrong word, but it needs to be made more precise. Today\u2019s widely used large language models are mostly based on so-called decoder-only transformers. Their text is produced by autoregressive decoding: in a single pass through all of its layers \u2013 the forward pass \u2013 the model computes a probability distribution for the next fragment of a word (a token). It selects one, appends it to the text so far and starts again. Within such a pass, there is indeed no feedback. The transformer architecture was deliberately designed in 2017 to\u00a0dispense with recurrence entirely. That made it highly parallelisable \u2013 and thus trainable at today\u2019s scale in the first place.<\/p>\n\n\n\n

Yet AI is not entirely loop-free. Each generated token becomes the input for the next step \u2013 that is feedback, albeit through a narrow bottleneck. What the model\u2019s top layers have computed reaches the lower layers in the next step essentially only via that one selected token. That even this outer loop achieves a great deal is now mathematically established. A transformer that has to answer immediately\u00a0provably cannot solve certain simple tasks\u00a0\u2013 under the assumptions standard in complexity theory \u2013 such as determining whether two points in a network are connected. If it is allowed to write out intermediate steps, its computational power grows with the number of those steps. This is exactly what reasoning models rely on when they produce long chains of thought before answering.<\/p>\n\n\n\n

Research is also working to widen the bottleneck. At Meta, a model was made to\u00a0feed back its last internal state instead of a word\u00a0\u2013 to keep thinking without words, so to speak. The authors point to findings that the brain\u2019s language network remains largely silent during many reasoning tasks. So-called recurrent-depth models\u00a0run through the same block of layers as often as needed\u00a0before outputting a token. A model with 3.5 billion parameters was thus able to perform as much computation at inference time as a model with around 50 billion parameters, in some cases improving markedly. A preprint published in July 2026 reports that such\u00a0\u201clooped\u201d transformers\u00a0can overtake conventional models for the same training effort.<\/p>\n\n\n\n

The case of the Hierarchical Reasoning Model from 2025 is particularly instructive. It explicitly imitated the brain: two recurrent modules, a slow one for abstract planning and a fast one for details. With only 27 million parameters, it\u00a0outperformed far larger language models\u00a0on reasoning puzzles. However, a re-examination by the ARC Prize team found that\u00a0the brain-like hierarchy contributed little. What mattered was an outer refinement loop in which the model repeatedly revises its own answer. A successor model with only\u00a07 million parameters\u00a0dispenses with the hierarchy altogether and still scores better on the ARC test than most large language models. The lesson: what counts is not the biological form of feedback, but that there is a loop of trial and correction at all.<\/p>\n\n\n\n

What AI really lacks<\/h3>\n\n\n\n

Put precisely, then, AI does not lack \u201cthe feedback loop\u201d as such, but three things:<\/p>\n\n\n\n