Machines may exceed human intelligence, yet every output remains bounded by the domains of architecture, data, sensors and objectives.
Until science can detect first-person experience beyond behavioural imitation, machine consciousness remains performance — not presence.
The decisive experiment is not whether a machine can say “I am,” but whether there is an inner witness to whom anything appears at all.
Seven years ago, on this blog, the claim was that creativity would become the new productivity as automation absorbed the routine work underneath it — that the organisations and minds who survived the transition would be the ones who stopped competing on throughput and started competing on origination. In 2019 that was a forecast, built on Christensen’s observation that adoption is blocked less by the merits of an innovation than by the anxiety and inertia surrounding it — what he called the circumstances of struggle. In 2026, with a large language model sitting behind every dictation app, every meeting summariser and half the drafting tools sold to knowledge workers, the forecast reads less like a prediction and more like a description of what already happened. It also reads like a warning about something that almost didn’t survive the transition intact: the muscle that does the originating in the first place.
Performance, Not Presence
Start with the epigraph, because the argument underneath it is the same argument underneath everything that follows.
In May 2025, one of the most rigorous tests neuroscience has ever run on the subject of consciousness reported its results. An adversarial collaboration — proponents of Global Neuronal Workspace Theory and Integrated Information Theory pre-registering their competing predictions with a neutral consortium before a single result came in — put 256 human participants through fMRI, MEG and intracranial EEG while they viewed stimuli of varying duration. The point of the exercise was to find out, empirically, where and how consciousness shows up in a brain we can directly instrument. The result, published in Nature, was genuinely mixed: real signal for conscious content in visual and prefrontal cortex, but no clean verdict for either theory. We still do not have a settled, working test for the neural correlates of consciousness — in a biological system we can put in a scanner and ask questions of directly.
That is the actual scale of the problem being waved away every time someone claims a chatbot is “basically conscious” because it can describe its own reasoning. If neuroscience cannot yet locate the inner witness in a brain we can study from the inside, there is no principled basis for inferring one inside a system whose entire output is the statistically most probable next token, however fluently that output performs the vocabulary of interiority. Performance of the words “I think” is not evidence of a thinker. It’s evidence that the training data contained a great many humans saying “I think.” The gap between imitation and presence is not a rhetorical flourish — it’s the actual, current, unresolved frontier of consciousness science, and it should make everyone more careful with the word “intelligent,” not less.
The Priming Problem
Here is the part that matters for anyone who writes, and the part the LinkedIn comment this article grew out of was actually about: almost everyone reading this has, in the last twelve months, been primed — repeatedly, invisibly — by AI-generated content, AI-assisted dictation, and AI-generated summaries of their own thoughts. The voice note becomes a transcript becomes a draft becomes a “cleaned up” version of something you never actually wrote. Somewhere in that pipeline, the part of the process that used to do the actual thinking got quietly disintermediated.
This isn’t nostalgia for pen and paper. It’s neuroscience. A 2023 high-density EEG study in Frontiers in Psychology put students through handwriting and typing tasks side by side and found that handwriting — but not typing — produced widespread theta-band synchronisation across temporal, parietal and central brain regions, the specific frequency band associated with memory encoding and the formation of new information. Typing, by contrast, engages a narrower, more mechanical circuit — recognisably less brain doing recognisably less integrative work. The advantage isn’t the pen. It’s the friction: handwriting is slow enough, and effortful enough, that the brain is forced to organise and prioritise the material rather than simply transcribe it.
That friction has a name in cognitive psychology, and it predates the EEG evidence by four decades. In 1978, Norman Slamecka and Peter Graf ran an experiment that has since been replicated so many times it has its own textbook entry: one group of participants read word pairs like “hot–cold”; another group had to generate the second word themselves from a fragment, “hot–c___”. On a later memory test, the group that generated the word remembered it substantially better than the group that merely read it — a reliable, moderate effect (d ≈ 0.40 in the modern meta-analytic estimate) known simply as the generation effect. The leading explanation is that self-generation demands more cognitive effort than passive reception, and that effort is precisely the thing that deepens encoding.
Put those two findings together and the LinkedIn comment’s intuition stops being an intuition. Writing by hand, and writing by effortful original composition generally, isn’t a stylistic preference — it’s a documented mechanism for building the kind of memory and understanding that dictating a thought and letting a model tidy it up simply does not exercise. The authenticity people sense missing from AI-smoothed prose isn’t sentimental. It’s the absence of the theta-band, generation-effect work that produces a thought worth having in the first place.
Why This Is Mathematically, Not Just Aesthetically, True
The LinkedIn comment made a stronger claim than “AI writing feels hollow” — it said AI is, by mathematics, destined to fail where the creativity paradigm does not. That claim now has real evidence behind it, and it’s more precise than the rhetoric suggests.
A 2025 meta-analysis pooling 28 separate studies and 8,214 participants found that people using generative AI do produce more individually creative output than unaided people — but at the population level, GenAI use substantially reduces the diversity of ideas produced across everyone using it. Everyone gets a little better; everyone converges on the same handful of good answers. A companion strand of research on model collapse shows the mechanism driving this at the model level: models trained recursively on AI-generated data drift toward “bland central tendencies” and lose their grip on rare, unusual, or minority patterns — the exact material genuine originality is made of. A 2026 paper in Trends in Cognitive Sciences extends this from the models to the humans using them, documenting a measurable homogenising effect of large language models on human expression and thought itself, as AI-shaped text re-enters the culture, gets learned from, and narrows the next generation’s range of expression a little further.
This is what “destined to fail by mathematics” actually means, precisely stated: a system trained to predict the statistically most likely continuation of existing text is, by construction, a regression toward the mean of what has already been said. It can interpolate brilliantly inside that distribution. It cannot originate outside it, because originating outside the distribution is definitionally the one thing the objective function penalises. Creativity, in the sense that matters, is not a especially fluent point inside the distribution of existing human expression. It’s the thing that moves the distribution.
The Substrate Creativity Actually Runs On
There’s a specific neural mechanism for this, and it was only nailed down causally in 2024. The default mode network — the set of brain regions most active during rest, daydreaming, and internally directed thought, engaged in something like half of a person’s waking hours — has long been correlated with creative output. What a 2024 study in Brain established for the first time is causation: using direct cortical stimulation, researchers found that suppressing default-mode-network regions specifically reduced the originality of participants’ responses on a divergent-thinking task, without touching their fluency or how much their minds wandered. Originality has an identifiable neural substrate, and it is not the same substrate as simply generating a lot of options or letting attention drift. It is a specific, internally directed, effortful mode of cognition — the same register that handwriting activates and dictation short-circuits.
Current AI architectures have no analogue to this. There is no idle, internally directed state in which a language model free-associates against its own accumulated experience and returns something genuinely novel to itself. There is only the forward pass: input, weighted probability, output. The absence isn’t a current limitation awaiting a bigger model. It’s architectural, in the same sense the epigraph means it — bounded by architecture, data, sensors and objectives, and creativity’s specific neural mechanism lives outside all four.
The Actual Challenge, Restated Properly
This is the real problem for the software industry building on top of these models, and it’s worth stating as precisely as the LinkedIn comment first raised it: the dominant product pattern right now is stimulus-on-demand — dictate, summarise, generate, polish — which is optimised to remove exactly the friction that the EEG data, the generation effect and the DMN-originality research all independently identify as the mechanism creativity runs on. A product built that way doesn’t make its users better writers. It makes them faster producers of averaged prose, which is a different and, on the present evidence, actively degrading skill.
The company that wins the next phase of this isn’t the one with the most fluent autocomplete. It’s the one that builds tools shaped like the generation effect — that interrogate persona, purpose and context before producing a word, that create productive friction rather than dissolving it, that measure success by whether the user’s own thinking got sharper, not by how few keystrokes it took to ship a paragraph. That’s a genuinely harder product to build than a summariser. It’s also the only version of this technology with a defensible claim to making anyone more creative rather than more average.
So Do That
None of this is an argument against the tools. It’s an argument for using them with your eyes open about what, specifically, they can and cannot do for you — and for treating the parts of writing that still require friction, generation and an internally directed mind as the parts worth protecting, not the parts worth automating away first.
Machines can now do a large share of what used to constitute a working day, often for less money and, on narrow tasks, with fewer errors. The economically rational response to that isn’t panic about displacement, though the displacement is real. It’s recognising what’s left standing once the routine work is gone: the one capability that a regression toward the mean cannot produce, that a system without a default mode network cannot generate, and that a century of neuroscience keeps finding is inseparable from the effortful, self-generated, occasionally illegible act of thinking something through yourself, on the page, before anyone — human or otherwise — smooths it out for you.
We are finally freed up, and disproportionately rewarded, for doing the one thing only a mind can do.
So do that. I will be extending this further in my forthcoming essay
This piece extends the argument first made in Creativity Is The New Productivity (2019).