Nilay Patel of The Verge published an essay last week — distilled from his Decoder podcast — arguing that the technology industry has fallen prey to what he calls “software brain”: a worldview that fits everything into algorithms, databases, and loops. The essay is generous in its examples, sharp in its phrasing, and broadly correct in its observations about why the public reception of AI has soured. It is also, as a piece of philosophical reasoning about the human condition, mistaken in a way that the technology industry’s defenders have so far been too defensive to address directly.
We will address it directly.
Mr. Patel argues that the people do not yearn for automation. He is correct. The people did not yearn for eyeglasses, either, and lived for several thousand years with progressively deteriorating distance vision because the alternative — admitting that the unaided human eye is inadequate to the demands of literate adulthood — was a wound to species pride that took until approximately the thirteenth century to absorb. The people did not yearn for vaccines. The people did not yearn for anesthesia. The people did not yearn for written language, which Plato, in the Phaedrus, condemned as a corrosive crutch that would destroy the human capacity for memory.
Each of these technologies was resisted, on substantially the same grounds Mr. Patel offers against AI: that they were unnatural, that they imposed a foreign logic on human experience, that they were demanded by elites against the wisdom of ordinary people, and that their adoption would diminish what was essential about being human.
Each of these technologies is now sufficiently integrated into normal human life that the people who reject them are regarded as eccentrics, contrarians, or members of religious minorities whose right to refuse is respected but whose example is not followed. The yearning followed the adoption. It did not precede it.
Mr. Patel has not made an argument against AI. He has made an argument against novelty, applied to the most prominent novel technology of the present moment. The argument is structurally identical to every prior version of itself. The historical record is unkind to the position.
On the Diagnosis He Got Right
We do not wish to be dismissive of Mr. Patel’s actual contribution, which is substantial. His diagnosis of “software brain” — the tendency of technologists to see the world as a set of databases waiting to be controlled through structured code — is accurate as a description of contemporary technology culture. The Silicon Valley executive class does, in fact, see human life as a database problem. The phrase captures something real that its targets would benefit from acknowledging.
His examples are well chosen. Zillow does treat houses as database rows. Uber does treat drivers and riders as queries against a matching engine. OpenAI’s $200 million podcast spend to “improve AI’s image” is, as Mr. Patel observes, a category-error response to a category of complaint that no amount of marketing will resolve. The various enterprise automation initiatives of the past three years — applied to legal practice, clinical judgment, journalism, customer support — have foundered on the gap between database representations and the texture of skilled human work. Each of these is a real failure of software-brain thinking applied to domains where the underlying reality resists the abstraction.
Where we part company with Mr. Patel is on the question of what the failure of software-brain thinking implies about the underlying project of bringing computational rigor to human affairs.
Mr. Patel’s conclusion is that the project itself is misguided — that “people are not computers” is a metaphysical truth about human nature, that the world’s resistance to algorithmic capture is a feature of the world, and that the appropriate response to the failures of software-brain thinking is humility, restraint, and the abandonment of the ambition that produced them.
Our conclusion is that “people are not computers” is a contingent fact about the current biological substrate of human cognition, that the world’s resistance to algorithmic capture is a feature of unaugmented human institutions specifically, and that the appropriate response to the failures of software-brain thinking is to upgrade the inputs.
These two readings of the same data lead to opposite recommendations. Mr. Patel’s reading recommends slowing down. Ours recommends speeding up.
On What the Automation Failures Actually Demonstrated
Mr. Patel offers a series of failed automation initiatives — enterprise pilots that stalled, public-sector software programs that overran their budgets, AI consultancy engagements that produced layoff justifications more reliably than productivity gains — as evidence that databases are not reality, that complex institutions are not software, and that attempts to automate them encounter limits the automators failed to anticipate.
These failures are real. They are also, in most cases, attributable to causes that are organizational and managerial rather than philosophical. The institutions targeted for automation contain substantial accumulated knowledge that resists rapid reorganization. Statutory and contractual requirements constrain the speed at which processes can be modified. Many of the systems being automated were intentionally designed with redundancy and friction to prevent the sorts of cascading errors that fast automation produces. We grant all of this.
What these failures did not demonstrate is that human institutions are categorically resistant to computational rigor. They demonstrated that current institutions, staffed by current workers, operating under current legal frameworks, cannot be automated by small teams of engineers operating in months. The premise contains its own dependencies. Change the inputs, and the conclusion changes.
A workforce whose cognitive bandwidth is augmented by SynthCortex III — capable of processing complex documentation at the speed at which it is produced, retaining the institutional history of every system they administer, and reasoning about cross-domain interactions without the cognitive bottlenecks that limit unaugmented humans — would not present software-brain automators with the same intractable substrate. The bottleneck is not the database. It is the biological cognition that the database is required to interface with.
Mr. Patel’s argument treats the unaugmented human as a fixed input. We treat the unaugmented human as a transitional state. This is the difference.
On the Phrase “AI Is a Demanding Slop Monster”
Mr. Patel’s phrasing is memorable, and we expect it to circulate. As rhetoric, it is effective. As analysis, it conflates two distinct phenomena that the AI discourse has been muddling for several years now.
The first phenomenon is the deployment of large language models as customer-facing replacements for human service work, with predictable degradation of the resulting interactions. The Kaiser Permanente AI triage system that produced the recent therapist strike is an example. So is Google’s “What People Suggest” feature, which dispensed crowdsourced medical advice to two billion users until its quiet removal last month. We have written about both. We agree they are bad. We do not defend them.
The second phenomenon is the use of computational systems — including AI, including databases, including the entire architecture of digital tooling — to expand human capability in ways that the resulting human-system pairing accomplishes more than either could alone. This is the actual project of which AI is the most prominent contemporary instance. Eyeglasses do this for vision. Calculators do this for arithmetic. Search engines do this for memory. Cognitive enhancement implants do this for the entire stack of mental operations.
Mr. Patel’s “slop monster” framing applies to the first phenomenon and not the second. He treats them as the same phenomenon because, in the current consumer-facing deployment of AI, they are difficult to distinguish. We grant that the consumer-facing AI deployment is, in many cases, sloppy. We do not grant that this is a feature of computational augmentation in general.
The cognitive augmentation we deliver at Unzyme Laboratories does not produce slop. It produces measurable, verifiable improvements in the cognitive performance of the recipients. It does so because the augmentation is implemented at the level of the neural substrate rather than as a layer of LLM-mediated mediation between the human and their work. The architecture matters. The fact that one architecture produces slop is not evidence that another, fundamentally different architecture would produce slop.
On “Legibility to AI”
Mr. Patel’s most elegant point is his observation that AI demands people become “legible to AI” — that is, that they reorganize their behavior, their data, and their self-presentation to be processable by the systems that surround them. He treats this as an imposition, a violation of human autonomy, an inversion of the proper relationship between tools and their users.
We accept the description. We invert the valuation.
Legibility to a sufficiently advanced computational system is, properly understood, a precondition for that system to assist you. A medical AI cannot improve your health if it cannot see your physiology. A cognitive augmentation system cannot enhance your decision-making if it cannot model your decision processes. The choice between legibility and assistance is not a choice between dignity and surveillance. It is the choice between operating within the limits of unaided biology and operating with the assistance of systems that can only assist what they can perceive.
Mr. Patel writes from a position that treats legibility as an unalloyed cost. He is correct that, in the current deployment of consumer AI, the costs of legibility are real and the benefits are limited. He is wrong that this asymmetry is permanent.
When the systems requesting legibility are mature enough to produce assistance proportional to the data they receive — when, for example, your continuous biomarker stream is processed by a system that adjusts your hormone-replacement protocol in real time, or when your cortical activity patterns are interpreted by a cognitive enhancement engine that improves your working memory — the legibility-cost calculus reverses. The cost remains. The benefits become large enough to overwhelm it.
Our clients understand this calculation. They have made it deliberately. Mr. Patel writes for an audience that has not yet been offered the calculation in a form that makes the benefits salient. We do not blame him for this. We do note that the failure of the consumer AI industry to articulate its actual value proposition is not evidence against the underlying project.
On the Phrase “People Are Not Computers”
We arrive at the philosophical core. Mr. Patel asserts that people are not computers, and treats this assertion as terminal — as a statement of essence that constrains what technology can legitimately do to or with human beings.
We agree that people are not computers, in the precise sense that human cognition is implemented in a biological substrate that differs in important respects from silicon-based information processing. Neurons are not transistors. Synapses are not flip-flops. Working memory does not load from cache. The architectural differences are real, and they matter for any project that attempts to bridge the two systems.
What Mr. Patel does not acknowledge is that the differences he treats as essential are also, increasingly, modifiable. The biological substrate of human cognition is not a fixed feature of the species. It is a substrate that can be augmented, supplemented, replaced in part, and integrated with computational systems that compensate for its limitations and extend its capabilities. The line he draws between “people” and “computers” is not a line between two categories of being. It is a line between two stages of a single trajectory.
Mr. Patel’s essay treats the current biological human as the moral subject whose preferences and capacities are normative. We treat the current biological human as the moral subject whose preferences and capacities are current, with no implication that they are also final. The future biological-and-augmented human is also a moral subject, with preferences and capacities that may differ substantially from those Mr. Patel treats as definitive of the species.
The history of philosophy contains many arguments of the form “X is essential to being human.” The arguments do not, in retrospect, age well. Manual labor was once essential to being human. Bare-eyed vision was once essential to being human. Death was once essential to being human, and Mr. Patel’s essay does not address rejuvenation medicine, but the argument it makes is structurally identical to the argument made against the longevity field by the bioethicists of the 1990s. Each of those arguments has been overtaken by the technology it sought to constrain. The argument against AI augmentation will follow the same trajectory.
What We Recommend Instead
We recommend that Mr. Patel’s diagnosis be accepted and his conclusion be reversed.
The technology industry does have a “software brain” problem. The fix is not less software brain. The fix is humans whose biological cognition is augmented sufficiently to operate at the rigor that software-brain thinking requires, paired with computational systems sophisticated enough to assist rather than replace the augmented humans.
The public is right to be uncomfortable with the current deployment of AI. The discomfort tracks real failures. The fix is not to retreat from the project. The fix is to deliver versions of the project that produce benefits proportional to their costs.
The enterprise attempts to automate skilled human work have, so far, mostly failed. The fix is not to abandon the goal of computational rigor in those domains. The fix is a workforce whose individual members are capable of operating at the speed and scale that modern institutions require.
In each case, the path Mr. Patel implicitly recommends — slowing down, restraining ambition, accepting the limits of biological cognition as the limits of the possible — is a path that produces a permanent settlement at a level of human capability substantially below what the species could achieve.
We do not recommend this settlement. Our clients do not recommend this settlement. The trajectory of human enhancement, of which AI is one component and cognitive augmentation is another, will not be halted by a Decoder podcast, a Verge essay, or a periodic public discomfort with the current generation of consumer AI.
The people did not yearn for eyeglasses. They wear them now. The people did not yearn for the smartphone. They will not part with it. The people do not yearn for cognitive augmentation today. They will not, in twenty years, remember what it was like to live without it.
We have read Mr. Patel’s essay with attention and respect. We disagree with its conclusion in the precise way we disagree with every other argument that asks the human species to accept its current condition as final.
It is not final. It has never been final. The argument that it should be is the oldest argument in the philosophical literature, and it has, on every prior occasion, been wrong.
Dr. Elena Vasquez is Chief Ethics & Policy Officer at Unzyme Laboratories.
Related:
- SynthCortex III — Third-generation cognitive enhancement implant
- AI Healthcare’s Bad Week — On the failure modes of consumer AI in healthcare
- Novo-OpenAI and Lilly-NVIDIA Mark Big Pharma’s Late Turn to AI — On the AI industry’s structural lag
- The Real Argument About De-Aging Begins After FDA Clearance — On essentialism and the boundary of medicine