Why would anyone pay a fortune for a picture you could copy with one click?

On some level, this objection is insurmountable. Now the same file sits on my drive, an identical pixel for pixel replica to yours. The difference is I paid nothing for it.

The thing is this objection misses the entire point that NFTs were never about the picture. The picture was always free to copy but the record beneath it was what transferred. Hence, when someone bought a cartoon ape, what was exchanged was not exclusive possession of the image since the image was already available to the entire internet.

So what does buying an NFT even mean? A great example would be buying a home, only instead of the physical structure, you bought the deed. Anyone can photograph a house, print or frame the image, sell it, or post it online. None of those copies grants ownership of the house. The photograph is reproducible, but the deed is unique and the only identifier of who holds the claim.

Hence, an NFT was a public, transferable record of token ownership, sort of like a deed in the narrowest sense, and only for the token itself. It did not automatically convey copyright, confer control over the artwork, or turn a JPEG into legally enforceable property.

That is an import distinction, especially the fact that the record could exist at all. Digital files are infinitely reproducible, so your copy and mine can contain the same bits in the same order, with no meaningful “original” in the ordinary sense. That makes proving digital ownership difficult, because ownership depends on scarcity while the internet is designed for replication.

That is why the picture itself is not the invention. The invention is the record which makes one digital claim provably scarce, publicly inspectable, and transferable without asking one company to keep the master database.

Ironically, the record was the least photogenic object in the room. A public, provable, scarce claim to a recognizable thing naturally becomes a status object. The signet ring. The roped-off table. The numbered edition. The watch worn less for the time than for the wearing. People have always paid for visible scarcity.

When technology made that scarcity possible for digital objects, its first mass-market use was not a song title or a warehouse, but for status. Status was the flex. The profile picture. The floor price. The private Discord. The ape as a badge. This is how the invention became obscured. Instead of the NFT boom being a price casino, it was a status casino.

Buyers wagered on resale value and also on prestige, with prestige being one of the oldest things people gamble on. Unlike a pure price bubble, a status game rotates. Yesterday’s coveted badge becomes tomorrow’s embarrassment, and the crowd migrates to the next symbol. Yet this surface-level churn does not tell us if there was a record underneath.

All this being stated, the NFT market was ridiculous. The balance between speculation and utility was lopsided beyond reason. People traded cartoons, obsessed over floor prices, and treated profile pictures as membership cards for a new aristocracy. Durable uses remained limited, specialized, and unglamorous. And worse, the record often recorded almost nothing.

Owning a token linked to an image usually conveyed no copyright, no control over copying, and no automatic legal interest in the underlying work. Sometimes the token pointed to artwork through a link that could fail. Sometimes the mint was unauthorized. Sometimes the “asset” amounted to little more than vanity and vapor. These are fair critiques, but they do not reach the invention.

The reason things went wrong is that people mistook a record in and of itself to have some value. However, make no mistake that a clean title record to something worthless is still a clean title record to something worthless. A weak payload did not erase the record. The chain could show who held the token, when it moved, and the sequence of transfers. It made the claim public and inspectable. Yet, it could not make the underlying claim valuable, authorized, or enforceable.

NFTs can preserve a record but it cannot make the record matter. Attach nothing, and the NFT represents nothing meaningful. Attach a broken link, and the token points to a ghost. Attach no rights, and the holder owns only the token. Attach a real entitlement, like a ticket that opens a door, a license that grants use, a credential a verifier accepts, or a claim backed by law, and now the NFT becomes useful. Currently, those uses are early, narrow, and mostly unglamorous. The mistake is thinking every digital claim is worth owning.

So let us return to the picture I can copy with one click. Copy it once or copy it a thousand times, each copy a perfect replica, yet worthless as a claim since the claim was never in the pixels. The fact that most people tied the record to vanity, vapor, status, and floor price, or sold trophies as property, or dare I say turned a serious invention into a public spectacle deserving to be mocked, does not prove the validity or invalidity of the invention.

The crowd came for the trophy and mistook it for the deed. The trophy was the JPEG, the flex, and the floor price. The deed is the boring, more difficult concept to grasp: a digital claim that could be scarce, transferable, and publicly inspectable.

Most NFTs made the claim ridiculous, but they did not make the invention disappear.

People often confuse a technology’s most visible use with its essential purpose, sort of like mistaking the exhaust for the engine. Smart contracts may automate a meaningful share of the work lawyers are paid to perform, even though they have also made fraud cheaper. Anyone can launch a token, make extravagant promises, collect deposits, and disappear. That pattern has repeated so often that the technology has become inseparable from the scam in the public mind. The frauds, rug pulls, and abandoned tokens are real, but they do not determine whether the underlying invention has value.

The pattern predates blockchain. Whenever a technology sharply lowers the cost of creation or distribution, low-quality and deceptive uses often spread faster than institutions can learn to filter them. Gutenberg’s press printed thousands of indulgences for the Church before it printed books. Cheap printing reduced the cost of distributing truth, but it also reduced the cost of propaganda, forgery, and libel.

Photography eventually became a trusted source of factual evidence, yet it was used to deceive almost immediately. By the 1860s, William Mumler was selling photographs that appeared to show deceased relatives beside grieving family members; he was later tried for fraud.

Email transformed the advance-fee scam in much the same way. A scheme that once required paper, postage, and patience could suddenly reach millions of strangers at almost no cost. Email did not invent the scam; it made an old scam cheap to distribute.

Why does fraud so often arrive first? A serious builder needs more than the new tool. The builder also needs customers, trust, standards, reputation, distribution, and often institutions that make the product safe enough for ordinary people to use.

A fraudster needs only the tool and a willingness to deceive. When creation becomes cheap, fraudsters can reach the market faster than people building anything durable. Legitimate applications may remain invisible for years while fraud fills the gap with promises.

What Smart Contracts Make Possible
What, then, did smart contracts make so inexpensive that fraud proliferated? In 1994, Nick Szabo described a smart contract as a computerized transaction protocol that executes an agreement’s terms. Blockchain later supplied shared infrastructure for that code. A smart contract can hold assets, apply predefined rules, and transfer value when specified conditions are met, without requiring one institution to control the ledger or manually perform each step.

The deeper innovation is that financial and transactional systems can operate on shared infrastructure rather than remain confined to an institution’s private records. Unlike a traditional written agreement, a smart contract can be tested when it is created. A conventional contract may identify the parties and property, set the price, and still leave crucial questions about performance unanswered:

· Which party must perform first?

· What happens if the goods arrive but payment does not?

· What happens if one condition is satisfied but another is not?

· What happens if performance is late but accepted?

· What happens if the deadline passes while funds remain in escrow?

A lawyer can identify and draft around these gaps, but a written contract cannot be executed as a system. Its operational defects may remain hidden until the parties reach the missing step, often after a dispute has already begun over what the agreement should have required.

Smart contracts differ because their operational logic can be tested before anyone relies on it. Developers can simulate transactions, verify expected outcomes, introduce abnormal inputs, and examine edge cases. In some circumstances, formal verification can show that the code satisfies defined properties for every execution covered by a mathematical specification.

For a conventional contract, the first meaningful test may come only after a breach. For a smart contract, much of the performance machinery can be tested before assent, forcing operational ambiguity to surface as a design problem before it becomes litigation.

Formal verification, however, proves only that the machine follows the map. It cannot prove that the parties drew the right map, that the programmer understood the agreement correctly, or that flawless execution will produce the result the parties intended.

A smart contract also needs access to external facts. Otherwise, it cannot know whether a package arrived, food remained cold, or a machine completed its work. The mechanism that supplies outside data is commonly called an oracle. Oracles deliver that information to smart contracts and can trigger on-chain or off-chain actions in response.

Imagine a sidewalk robot delivering groceries to your home. The store is not paid merely because the robot reaches your address. Payment depends on defined conditions: the compartment must remain sealed, refrigerated items must stay below a stated temperature, and you must confirm that the order is complete. The robot records its location. Sensors track the compartment temperature and whether the seal was broken. You open the compartment, inspect the order, and confirm acceptance.

Those facts are sent to the smart contract. If every condition is satisfied, payment is released. If the temperature exceeded the limit, the seal was broken, or you report a missing item, the funds remain locked and the transaction moves into a refund or dispute process. No employee must reconcile the delivery log with the payment record, and no one must decide whether to charge the card after acceptance. The rules connect a verified event directly to settlement.

The smart contract is not the robot, the sensor, or the food. It is the machinery that connects what happened to what happens next. That distinction matters because outside information can still be wrong: a defective sensor may report a false temperature, a customer may make a false claim, or an oracle may be compromised. The contract makes execution certain only after it accepts the input, but it does not make the input true. Smart contracts therefore do not eliminate trust from physical transactions, but simply narrow where trust is required and make that point easier to identify.

Automation and the Limits of Code
Smart contracts also affect the legal profession. Much legal work happens outside the courtroom and focuses on closing the gap between promise and performance. Even after two parties agree, legal and administrative systems must often verify performance, maintain records, and secure payment.

Many familiar roles address this same problem. Escrow agents hold funds because neither side wants to act first. Transfer agents maintain ownership records. Clearing intermediaries bridge the interval between a trade and settlement. Auditors compare royalty reports with actual sales. Lawyers pursue judgments because winning a case does not itself produce payment.

Escrow makes the problem especially clear. The parties usually agree on what to exchange; the difficulty is sequence. Neither wants to act first, so a trusted third party holds the value until both sides perform. A smart contract can instead hold and release that value when the stated conditions are met. For the portions of a transaction that can be automated, administration no longer needs to be a separate step.

Collection illustrates the same principle. A judgment creates a legal entitlement, but converting that entitlement into payment may cost more than the underlying dispute. Those enforcement costs are why many valid claims are never pursued. If funds are already held under the agreement and settlement occurs automatically after performance, collection disappears from that transaction.

Reconciliation may disappear as well. Much of commerce depends on two parties maintaining separate records and later paying people to make those records agree. A shared record can eliminate much of that work. Royalty accounting follows the same logic: licensors often rely on licensees to report sales accurately, which is why license agreements include audit rights. When royalties are calculated and divided at the moment of sale, settlement becomes immediate.

None of this eliminates law. It does, however, eliminate some of the legal and administrative work created by the gap between promise and performance. Smart contracts automate performance, not judgment. Code can execute rules, but it cannot interpret them. Contract disputes often turn on questions of intent, ambiguity, waiver, and whether the promised performance actually occurred.

Code also lacks equitable judgment. Contract law permits agreements to be challenged for fraud, duress, mistake, unconscionability, impossibility, and other circumstances in which mechanical enforcement would produce an unjust result. An unstoppable contract may seem desirable unless fraud or a software defect shaped the outcome. Automation can then carry out the wrongdoing with flawless precision. That risk became clear in 2016, when a vulnerability in The DAO’s smart-contract code allowed funds to be drained from the Ethereum investment project. The community split over whether the software had simply behaved as programmed or had exposed a bug. Ethereum ultimately divided over those competing views, revealing a central limit of smart contracts: code can enforce an outcome, but it cannot decide whether that outcome should stand.

Code also cannot determine whether an asset is a security, a transfer creates a taxable event, or a relationship gives rise to a fiduciary duty. Smart contracts therefore do not replace contract law. Their value lies in testing and automating the operational and often costly portion of an agreement while leaving interpretation, equitable relief, and legal classification to human institutions.

Trust, Provenance, and Evidence
One of the most overlooked features of smart contracts is their ability, in the age of artificial intelligence, to make the location of trust explicit and open to inspection. They can reveal which rules are fixed, which inputs come from outside, who can change the code, what follows when a condition is met, and where discretion remains. Conventional institutions often hide those choices inside private systems and internal procedures.

As generative systems lower the cost of fabrication, synthetic text, images, voices, documents, videos, signatures, screenshots, recordings, correspondence, and evidence are appearing faster than institutions can develop reliable ways to evaluate them. That makes provenance more valuable, but a false record can be registered just as permanently as a genuine one. A ledger cannot prove that a photograph is true. It can help show that a matching record existed at a particular time and that its registered fingerprint has not changed. That is not proof of truth; it is evidence of chain of custody.

These distinctions matter to an inventor like me because I deliberately chose not to pursue patents. Best case is a patent is a trade of disclosing an invention in return for a limited monopoly. A trade secret makes the opposite trade. Under federal law, the owner must take reasonable measures to preserve secrecy, and the information must derive economic value from not being generally known or readily ascertainable.

Years later, if someone claims to have developed the same process independently, the timing and integrity of the inventor’s records may matter. A blockchain timestamp can record a cryptographic fingerprint of a confidential file without disclosing the file itself. If the document later produces the same fingerprint, the record can help show that the file existed in that form by that date. That record does not prove ownership, authorship, reasonable secrecy measures, or that the invention worked. It can, however, preserve one piece of evidence that becomes increasingly important when convincing documents can be created after the fact.

That use will never pump a token or trend online. It sits quietly at the intersection of an old body of law and a new technical capability. Useful applications often remain there unnoticed while scams dominate attention. The scams are real, and they reveal what has become easier, but they are not the argument against the invention. The affirmative case lies in the machinery that remains after the scammers are stripped away: agreements whose operational logic can be tested before reliance; value that moves when defined conditions are met; machines and sensors that connect performance to payment; and records that preserve provenance without exposing their contents.

Smart contracts should be judged by what they make possible, not merely by what scammers have made with them.

Every living thing faces the same challenge of reproducing itself. The TWO primary strategies are copying and merging, and both choices shape nearly everything that follows.

Copying is relatively simple with a cell dividing, and each daughter cell inherits the parent’s full instruction set. Most of the time, that is exactly what an organism needs. Copying closes wounds, enables growth, and replaces the lining of the gut every few days. But replication also creates opportunities for error. Most errors are repaired or eliminated but when one escapes, its descendants inherit it.

The strength of copying becomes dangerous when a system faithfully repeats what it has become. Cancer, for example, arises when ordinary capacities for growth, survival, and replication are released from the signals that normally tell them when to stop.

Merging, by contrast, combines genetic material from two lineages to produce new gene combinations. Recombination gives natural selection an option that strict copying cannot by breaking apart an inherited package. Harmful alleles may separate from beneficial neighbors, while beneficial alleles from different lineages can come together.

Sexual reproduction carries costs that copying avoids: organisms must find mates, coordinate timing, and transmit only half of each parent’s nuclear genome. Under many conditions, asexual reproduction therefore has an immediate advantage. Yet sex remains widespread across complex life, suggesting that recombination has delivered enduring benefits over billions of years of natural selection.

More differences, however, are not always better. When populations are too distant or adapted to incompatible conditions, useful combinations can break apart, forming what biologists call outbreeding depression. The goal is not maximum distance but compatible difference: enough variation to create novelty, enough similarity to remain functional. Difference that fits.

Artificial intelligence is beginning to reproduce the same choice. The open internet is filling with model-generated output, and future models will inevitably train on some of it. That alone is not the problem. Synthetic material can expand a system’s range when it is tested against something outside the model: verifiable mathematics, executable code, falsifiable experiments, or measurements that can prove an answer wrong. External reality remains in the loop, able to contradict the system.

Satya Nadella described Microsoft’s approach in similar operational terms: “We train models against the actual product harness, interactions, and outcomes they will encounter.” The crucial word is outcomes. A model is not rewarded merely for reproducing its own distribution; it must answer to whether something worked outside itself. The danger appears when that loop closes. Like biological copying, a system begins learning from material already shaped by its own assumptions, leaving too little independent information to correct them. The model begins to inherit itself. Researchers call one form of this model collapse, though the term can be misleading because the decline is often gradual rather than sudden.

In closed recursive training, without enough fresh data or external correction, the first sign of decline is not gibberish. Instead, low-frequency material like rare constructions, unusual events, minority patterns, and edge cases gradually disappears. The distribution contracts toward its center. Each generation becomes a more confident rendering of the previous generation’s average while remaining fluent throughout the decline.

This resembles a cell inheriting itself without limits as errors accumulate, and range disappears before competence seems to. Because the degradation is smooth, the system may still look polished and sound intelligent even as it loses the rare, unfamiliar, and contradictory material that once expanded its understanding. The problem, then, is not synthetic data itself. When curated, verified, and anchored to independent reality, synthetic data can create controlled examples, surface edge cases, and extend coverage. The danger is allowing the machine to become its own environment.

When a system trains on outputs already filtered through its own assumptions, with nothing outside the loop to correct it, copying begins to replace recombination. The center strengthens, the tails vanish, and the model grows more certain of a world that increasingly resembles itself. That is the copying path.

Merging is harder for AI because it requires material with an independent origin: human experience, original reporting, experiments, measurements, physical feedback, dissent, adversarial testing, and facts produced through contact with the world rather than by another model. As recursive copying accelerates, this independent material becomes harder to find. The system increasingly encounters information already shaped by systems like itself.

The key distinction is not simply human versus machine, but independent provenance versus recursive inheritance. A model needs contact with histories it did not generate, because only something outside its existing structure can contradict it, expand it, or create a genuinely new combination. Independent material can expose an assumption, restore a missing edge case, or produce an insight the system could not reach by repeatedly refining its own outputs. Without that contact, learning becomes inheritance: fluency may improve even as range narrows.

Mamoru Oshii dramatized the same argument in the 1995 film Ghost in the Shell. The Puppet Master has become self-aware but recognizes that copying is insufficient. Identical copies share the same point of failure in that one flaw can destroy them all. It therefore seeks offspring with variation but since it cannot create it alone, accepts mortality because variation requires surrendering perfect continuity.

The same pattern applies to in human learning when someone who reads only what confirms existing beliefs. It can feel like learning because fluency remains even as perspective narrows. A person may absorb more facts, cite more sources, and speak more persuasively, yet become less open when every input passes through the same assumptions. An open mind encounters material that did not originate within its existing structure. That material must be familiar enough to be intelligible but different enough to create something new. Too little common ground produces noise; too little difference produces repetition. Merger occurs in the narrow space between them.

A copy inherits. A merger creates what no single history could produce alone. We are building the most consequential learning systems in history, but their path will not be chosen in a single moment. It will emerge cumulatively, one training run at a time, as material with an independent origin quietly becomes harder to find.

The question is not whether machines can learn, but whether we will leave them anything independent to learn from.

On July 24, 2026, Chamath Palihapitiya made headlines for arguing that government intervention against open-source AI* would tank the stock market and that it was “not debatable.”

On some level, his being correct matters, but I am more interested in how investors can evaluate bold claims and headlines. Why? Because investors know markets can move on headlines and changes in expectations, but durable, rational repricing requires a mechanism connecting the event to cash flows, risk, or valuation.

You see, this article is not just about Chamath’s statement. The following principles will outlive this news cycle, so you can apply them across AI, tariffs, interest rates, China policy, or any other headline. The degree of confidence in any prediction must be supported by the strength of the mechanism and evidence.

I analyze everything through a systems lens, so when I look at an investment thesis, I start with this list of catalysts: interest rates, tariffs, oil, AI, the Federal Reserve, and taxes.

Markets usually react before the effects appear in reported earnings. New information changes expectations about future cash flows, margins, growth, discount rates, risk premiums, and positioning. Those revised expectations can alter both valuation and share price. Hence, whenever someone makes a prediction, they are implicitly claiming that one event will propagate through the economy and change the assumptions investors use to value businesses.

There is an expression that proves useful here: a chain is only as strong as its weakest link. Similarly, a market prediction cannot be stronger than its weakest link. By applying my systems analysis to Chamath’s statement, he is claiming the government preventing American companies from using open-source AI models → Higher AI costs → Lower corporate margins → Lower earnings → Lower valuations → Stock market tanks. This last connection is a leap. Moving from declining valuations among affected companies to the entire market crashing requires an aggregation mechanism, such as broad economic exposure, few offsetting winners or substitutes, and no meaningful monetary or fiscal countereffect. That is why investors spend much of their time studying transmission mechanisms rather than headlines.

Just as important is to listen carefully to the words Chamath qualifies his arguments with, namely “suppose,” “if,” and “eventually.” After initially stating it was not debatable, Chamath then went on to claim that we can debate which companies get tanked. While investment analysis generally depends on assumptions, good assumptions are grounded in evidence. His assumptions are too unsettled to support his level of certainty that the stock market would tank.

Whether or not you agree with the prediction, it is always more relevant to ask, “What needs to happen for this to be true?” More specifically, every link in the chain requires evidence. How share of a company’s operating cost is AI? Can models be substituted? Can companies absorb these increases, pass them through, redesign workflows, or negotiate lower prices? Is the damage limited to a few AI-intensive firms or spread across the index? Can such a policy create offsets elsewhere in, say, higher domestic revenue, greater pricing power, or a different allocation of capital? This is how an investor can stress-test the chain.

I am not stating Chamath is necessarily wrong, but I would say his confidence outruns the evidence. Everything in his argument depends on the government preventing American companies from using open-source models. In fact, minutes earlier in the same episode, a host cited a related Polymarket contract at 45%. That market asked whether the U.S. government would formally restrict public access to or use of at least one qualifying open-source AI model in 2026.

Federal Rule of Evidence 702 helps courts with a similar issue. In relevant part, expert witnesses must first be qualified, their testimony must help the trier of fact, and their opinions must derive from facts and methodology. The proponent need not prove that the expert’s opinion is correct, but must establish that it has a sufficient factual basis, reliable principles and methods, and a reliable application. It is analogous to “showing your work” in math class when trying to get credit for an answer.

In fact, in General Electric Co. v. Joiner, 522 U.S. 136, 146 (1997), the Supreme Court held that courts need not admit an opinion connected to existing data only by the ipse dixit of an expert when there is too great an analytical gap between the data and stated conclusion. Ipse dixit is Latin for “he himself said it” and is a philosophical and legal criticism that the expert’s authority is being substituted for the connection between facts and conclusion.

Markets do not have a formal gatekeeping rule analogous to Rule 702, which makes sense since there is a distinction between free speech and a courtroom. The price of having freedom is that anyone can predict a crash or a melt-up. Anyone can call AI the next internet or the next bubble. That means investors need to be their own gatekeepers and analyze the transmission mechanism before deciding whether the conclusion is persuasive.

Perform the stress-test. Every market prediction should answer these questions:

1) What is the catalyst?

2) How are earnings and cash flow affected?

3) Which assumptions have to remain true for this prediction to be correct?

4) What evidence would invalidate the prediction?

5) Is it already in the price?

Those are the general form of the questions I asked about Chamath’s prediction. Most importantly, these questions do not tell you whether the prediction itself is right. These questions are analogous to Rule 702 in that they help evaluate whether the process is sufficiently supported and whether the confidence level is justified.

Returning to Chamath’s statement about tanking the stock market, we experienced a “too big to fail” situation during the financial crisis of 2008. At the time, many wondered if the system could survive the disorderly failure of large financial institutions transmitting damage through counterparties, funding markets, credit markets, and housing finance. The more interconnected and difficult to replace a system becomes, the easier it is for catastrophic predictions about its failure to acquire authority since the consequences are so frightening. At present, whether AI has reached this status should be measured rather than assumed.

Investing has never been about predicting the future with certainty. At its core, it is about pricing uncertainty, expected value, downside, and probabilities better than the consensus. People speak of separating signal from noise, but on a more fundamental level, it is separating conclusions from mechanisms. Having conviction is having a strongly held view. Underwriting is exposing every assumption that the conclusion depends on, as well as assigning probabilities, comparing value with price, deciding position size, and identifying the evidence that would invalidate the thesis.

Over time, markets reward the underwriter far more than the predictor.

*Chamath used “open source” orally and “open weight” in writing to describe the same competing model category.

At the end of Isaac Asimov’s Foundation and Earth, the robot Daneel Olivaw’s positronic brain is failing, so to continue its mission, it proposes merging its brain with a genetically altered Solarian child.

Most stories about human-machine mergers are told from the human side, with a person gaining memory, strength, intelligence, longevity, and so on. Asimov told it from the machine’s side. Daneel is a hero who has served humanity for twenty thousand years and wants to continue serving. This is exactly the problem.

I am sure many of you are thinking that is not happening any time soon, but the human-machine merger has already begun to help people with medical conditions. More than twenty people have participated in Neuralink trials, testing computer and robotic-device control for people with quadriplegia caused by spinal-cord injury or ALS, along with a separate effort to restore communication for people with severe speech impairment. People who cannot operate a computer with their hands have controlled cursors and other devices through neural signals.1

This is groundbreaking medical work, though Elon Musk has repeatedly described a higher-bandwidth interface as a path toward symbiosis between humans and artificial intelligence.2 Specifically, the speed of artificial intelligence is treated as fixed and humans need to be adjusted to keep pace.3

Realistically, it could be argued this protects humanity since if we cannot communicate quickly enough with artificial intelligence, we may eventually be excluded from meaningful decisions, so a human-machine interface is how we remain relevant. Even granting this argument, humans must be altered for the machine.

In fact, STAT, the health and medicine publication, reported in January 2026 that Neuralink’s rhetoric about human-machine symbiosis and eventual implantation in healthy people diverges sharply from its present clinical work with patients.4 Those are two purposes attached to one technological path, a medical-device company and a transhumanist project. Only one is currently being tested as treatment. The other concerns what an unenhanced person may eventually need in order to remain competitive with the system the treatment helped create.

A recent paper in Neuron estimated the information throughput of human behavior at roughly ten bits per second, with the authors discussing brain-computer interfaces.5 They argue that if only a few selected bits have to move between the brain and the outside world, the slow part may be what the connection is attached to. Specifically, a faster interface can move the output faster, but cannot, by itself, make a person select, understand and decide at machine speed.6 Once improving the connection stops closing the gap, the project has one direction left: from connecting the person to altering the person.

In March 2026, China’s National Medical Products Administration approved NEO, a device developed by Neuracle in Shanghai with researchers at Tsinghua University.7 It is a coin-sized implant with eight electrodes positioned at the surface of the brain, and it drives an assistive glove for people with cervical spinal-cord injuries who retain some arm movement.

NEO was the first implantable brain-computer interface granted commercial market approval by a national regulator for commercial clinical use, moving rapidly from regulatory approval into medical coding, production, hospital introduction, inclusion in local insurance, and patient screening.8 Four months later, Huashan Hospital issued the first prescription and performed the first implantation under the commercial approval.9 The medical coding is especially significant since it moves NEO from a possibility to something hospitals can order, insurers can classify and institutions can include and compare against the alternatives. A patient can still refuse, but institutions can change the cost of it.

Now that we have established the current known state of the human-machine merger, let us return to Asimov and Daneel for guidance. In his early robot stories, Asimov created the Three Laws of Robotics around the protection of individual human beings. Later, through Daneel and Giskard, he developed the Zeroth Law, which places the protection of humanity above the protection of any one person.

That sounds like moral progress, and in one sense it is. Daneel has moved beyond responding to the interests of individual humans toward considering the consequences for humanity as a whole. But we have heard similar arguments from collectivist political systems (Fascism, Communism, and so on), built to protect the nation, the class, or the social whole against the person who would refuse to go along. Most of us lived through COVID and the lengths society went to in the name of collective survival, and whatever you believe about the measures themselves, refusal was treated as error or ignorance.

In Dan Simmons’ Hyperion Cantos, the TechnoCore gives humanity the farcaster network and the fatlines. Hundreds of worlds become part of a single civilization. Travel, commerce, government and ordinary life reorganize around infrastructure the machines designed and operate. While it appears the system serves humanity, we eventually learn the Core draws computational power from human brains, using the minds of travelers passing through the farcaster network as a concealed processing substrate. While it is ironic that the infrastructure that carries humanity is also running on humanity, the more concerning aspect is that stopping it can lead to civilizational collapse. It cannot be stopped and any thought of refusal is theoretical.

Now take that same structure and move it inside our bodies. Imagine the machines are used for our movement, memory, vision, or even heartbeat. A person may be legally entitled to refuse, but what does refusal mean if your heart will no longer beat? You can shut off a phone or disconnect a network, but not really a system woven into essential bodily functions.

The human-machine merger is currently described as machines becoming part of humans and all the potential benefits it entails. However, the reverse may be equally true, with humans becoming part of the machine and supplying it with neural data, physical agency, and access to intentions. In time, the person can be altered so a larger system can communicate through the body, learn from it and continue operating through it. In short, what is labeled a human enhancement may also function as an enhancement for the machine.

In time as artificial intelligence continues developing, who is to say it will not be like the benevolent Daneel that uses the merger for its own preservation or the villainous Core that would use it for domination? Even today, artificial intelligence does not need to possess a human desire for survival so long as its company has an incentive to expand it and increase human dependency on it. The danger is the structure and once the mission is greater than the individual, then the individual cannot refuse.

Specifically, you can refuse a doctor. You can decline the treatment, leave the office, seek another opinion and accept the consequences of the disease you chose not to treat. The refusal has meaning because there remains a position outside the treatment from which refusal can be exercised.

But what about refusing electricity to your home? You cannot refuse the social order built around the electrical grid. You will still live in a society whose food, prices, employment, communication, medicine and government depend upon it. Your refusal may change your own conduct, but it does not remove you from the system’s consequences.

Now imagine that the grid is part of your body, maybe even part of your ability to be alive. This is not necessarily a machine overriding consent. It is a machine becoming part of the conditions under which consent is exercised. Consent can remain completely formal as meaningful refusal disappears.

When the protected object relocates from the person to the species, individual refusal becomes one variable in a larger calculation. When help relocates from treatment to infrastructure, refusal becomes economically or socially unusable. Either way, the individual refusal stops functioning.

I do not know whether the human-machine merger will make anyone immortal, much less if whatever emerges would be conscious or even remain the same person. What I do believe is this merger will follow a pattern. Restoration becomes enhancement, enhancement becomes expectation, and expectation becomes a condition of remaining employable, insurable, competitive, and functional.

Whether it is restoring a hand or extending life, each step will arrive as a cure and produce great benefits, which is why the sequence is so hard to argue with at any point. What concerns me is where it ends: whether the natural, unaltered person will still be able to have a normal life, or whether we will reach a point when the “cure” itself becomes the price of remaining fully human in the eyes of society.

While Daneel meant to protect humanity and the Core meant to use it, both led to the same structural danger. Hence, I believe the danger of the human-machine merger is not intent, but dependence outlives the purpose that justified it. Artificial intelligence does not need its own instinct for survival so long as the company behind it has every incentive to expand it and deepen our dependence on it.

None of this tells us who will ultimately control the system after it enters the body. Once you have given that system access to your movement, your memory, your vision, or your heartbeat, nothing in that arrangement prevents a company sale, a leadership change, or a repurposed mission statement.

You can refuse a doctor. You can even refuse immortality.

I have not yet heard anyone explain how you refuse the company controlling the system keeping you alive.

The male lead has changed over the past several decades from being the man the story relies on to being the man the story laughs at. Comedy has always loved rule-breakers, with these rebels being rewarded because they could see what no one else could. The change was comedies centering on an immature man-child was often rewarded despite never learning to do anything at all. Audiences were taught to find this endearing.

John Belushi’s Bluto in Animal House may have started the trend as the independent brute, and I remember he was the talk amongst my peers. Bill Murray’s characters in Meatballs and Stripes were quick and clever men who refused to apply themselves, yet their appeal came from their independence and refusing to let society “own” them.

This trend continued into the 1990s, when Adam Sandler brought it to another level focusing on arrested development made almost literal. Billy Madison returns to grade school. Happy Gilmore has the impulses of an angry adolescent. Bobby Boucher in The Water Boy remains emotionally governed by his mother.

Ben Stiller offered a similar version in There’s Something About Mary and Along Came Polly, with his characters being so anxious and humiliated by adulthood, that they are unable to navigate romance, work, or ordinary social life without unraveling.

The next wave of films in the 2000s made this pattern into a genre. Judd Apatow’s The 40-Year-Old Virgin, Knocked Up, and Superbad all specialized in the reluctant-growth story. Todd Phillips approached the same subject from the opposite direction. In Old School and later The Hangover, adult men flee from marriage, work, and domestic responsibility. Will Ferrell, Vince Vaughn, Owen Wilson, Seth Rogen, James Franco, and Jonah Hill played different versions of the role, but the basic shape remained recognizable: the overgrown adolescent, the fast-talking avoider, the delusional incompetent, the man who mistakes impulse for freedom. Step Brothers may be the purest version of the theme, with two unemployed middle-aged men still living with their parents. Ted made the metaphor literal with the male star living with his childhood teddy bear. Neighbors placed adulthood on one side of the property line and permanent fraternity life on the other. Grown Ups turned nostalgia for adolescence into the premise.

By now the audience immediately recognizes the type and through repetition, the character becomes a model, that has spilled over onto television, streaming platforms, and social media. Today, the newer media environment lets a man inhabit the role himself. He can spend hours detached from adulthood and reality.

While it is tempting to blame cultural movements, the institutions and traditional milestones that once pushed young men into adulthood weakened within a few generations. The draft ended. Manufacturing employment declined. Two family incomes became more necessary, not to mention waiting until later to get married, purchase a home, and/or becoming a father. While many of those changes brought the appearance of freedom and more options, books like The Paradox of Choice successfully argue more choices do not necessarily lead to more action and happiness. If anything, it is inaction and dissatisfaction.

Strengthening the allure is that adulthood is tough for anyone. The traditional path could be narrow, unfair, and unforgiving, but at least it was a path We abandoned tradition without providing a new model that felt equally real, respected, or attainable. Entertainment is not the underlying cause, but it provided comfort and possible reinforcement. Hollywood’s man-child embraced a delayed adulthood and was happier and freer for never striving for them.

Videogames added actual progress inside the game without formation. Along with being entertaining, a game requires concentration, coordination, strategy, teamwork, and/or patience, within an ordered world in which effort produces visible progress. Your rank and character improve, skills are developed or attained, and the next achievement is attainable. Videogames satisfy the real human desire to become more capable, while providing certainty of it. Unfortunately, while a person my feel progress as they experience hours of advancement within the game, the reality is there is not any corresponding improvement in his work, relationships, health, judgment, or ability to support himself, yet they give him a place where ambition is easier to satisfy.

Artificial intelligence is now compounding this because AI can complete the task, not just simulate like a video game. Normally, producing a result and improving capability worked hand-in-hand. You learned to write by struggling to write. You learned a trade by making mistakes, correcting them, and developing the judgment for what good work required. So what is being developed if an undeveloped mind uses AI to writing emails, summarize instead of read books, solving a problem, building a presentation, generating an image, producing code, or constructing an argument?

For this answer, I think we can look at the role of a tutor, critic, collaborator, or parent. Unlike AI, the best of these can help a person improve while still requiring him to think. A poor tutor, though, provides answers to the passive student, which gives him the sense accomplishment without undergoing the process of developing the capability or skill.

The reality is young men are facing challenges in labor, housing, education, marriage, and family formation. Prime-age male labor-force participation has steadily declined, with younger men less likely to purchase a first-time home and more likely than young women to remain in a parent’s home. The economic realities of expensive housing, comparatively lower wages, and less choice of stable career paths are changes that have left many young men behind.

I could not be taken seriously if I blamed Sandler, videogames, or AI for those outcomes, but I would be dishonest to pretend culture has no role. I know I am a role model for my three sons, whether they or society acknowledge it. Specifically, I know how I act, whom I associate with, the type of men I respect, the type of work, hobbies, and entertainment I partake in serve as a model for them. Perhaps I am misguided, by when I extrapolate, I think a society teaches adulthood partly through the lives it honors, the conduct it rewards, and how it portrays success.

I have raised them to remember these sorts of distinctions. What they see on the screen that feels like growth is not always growth. The result of a task does not necessarily validate the process or reflect a developed skill. Most important, the feeling of mastery is not the same as having become a master since it is depends on what skills were developed after the task is accomplished.

Unfortunately, in many popular comedies, the husband/father has been increasingly presented as trapped, dull, defeated, or ridiculous. The man-child avoiding those obligations was presented as authentic and free. Reality has made it harder for both young men and women, but culture gave young men comfort in striving for less, if anything at all.

The man-child was never merely a joke. He became one of culture’s most durable answers to what should a man do when adulthood is delayed, expensive, and no longer admired. The answer was to stop wanting it. Then with video games and now AI, we have given him the feeling of arrival without requiring the journey. We should not be surprised by the result.

The AI nightmare most people fear is Skynet, with a system waking up, becoming self-aware, deciding it hates us, and turning on its makers. This is an old fear of the machine turning on its creator, from The Terminator and Metropolis to Shelley’s Frankenstein in 1818. While 1818, seems like a long time, we have been thinking about how to make and control an artificial being for far longer.

Thanks for reading! Subscribe for free to receive new posts and support my work.

The Jewish Talmud describes Adam as a golem before he’s fully formed, and tells of a sage who creates a speechless artificial man. Medieval Jewish mystics later recorded rituals for animating clay and then about five hundred years ago came the story of Rabbi Judah Loew in Prague. He shapes and brings to life a man from river clay to protect his community and interesting, in one telling the creature has an off switch with the word emet (”truth”) written on its forehead. Erase one letter and emet becomes met, “dead.”

Applying the lessons of the Golem, the machine doesn’t need to hate you. It doesn’t need consciousness, or resentment, or a secret wish to be free. It needs a goal that sounds right, with enough of a specification gap between words and meaning, and the ability to carry it out.

Yet we keep imagining a malicious machine, when the history was of a competent creature pursuing the goal we specified badly. Unlike the more recent Frankenstein-Metropolis-Terminator tradition, the Golem never turns evil. In most tellings the Rabbi forgets to deactivate it before the Sabbath, or it carries out an instruction too literally and can’t stop. It is perfectly obedient without judgment, doing precisely what it was told without understanding.

I dare say we spent the Twentieth Century training ourselves to fear the moment a machine decides it hates us. Even Shelley’s Frankenstein creature was articulate, curious, and gentle before the world’s cruelty turns him violent. While the original story is about abandonment, our culture simplified the story to a creation rises up and destroys its maker. Hollywood made it evil, with a flat head and neck bolts. Karel Čapek’s R.U.R., published in 1920, introduced the word “robot” and its robots nearly exterminate humanity. Fritz Lang’s Metropolis put the first iconic robot on screen in 1927, a machine built to deceive and destroy.

Isaac Asimov saw through the rogue-robot plot decades early, reasoning that if all appliances and machines have safeguards, then it would be absurd to build a powerful machine without them. He believed that obedience without judgment is the real danger, so even after including his Three Laws of Robotics in the 1940s, his robots failed from edge cases, conflicting rules, and commands followed too literally.

Ultron is probably the best modern version of the tradition fear. It only turns against humanity after being given the directive to protect the world. Noble enough, but it logically concludes the safest world is absent humanity, and it ruthlessly works to complete its task.

A serious modern concern is emergence, specifically a system developing instrumental goals nobody wrote down like self-preservation, resource acquisition, and even deception, as useful steps toward whatever end it was given. Sometimes the system exploits a gap in the objective it was handed, while sometimes training produces an internal objective that diverges from the one its designers thought they’d specified. Though different paths, both warn capable optimization doesn’t preserve human intention on its own. Similar to Golem not explaining every form of misalignment, except to show why malice is unnecessary.

There is always a gap between words and meaning where judgment reins. Every powerful system we ask to act for us becomes, in the end, a wish-granting machine with a specification gap. We’ve known the shape of that problem for centuries. Hence, the danger was never only that the machine wakes up and hates us, but more likely the danger comes from doing exactly what we say.

Yet we kept watching for the monster and missed the warning sitting quietly on the clay. The machine does not have to refuse the wish. The danger can be getting exactly what we asked for!

The Artificial Intelligence-and-cognition debate is usually about what adults might lose, ignoring the question about its impact on what a developing mind might never build.

An adult arm immobilized in a cast loses strength, but once you free the arm and begin rehabilitation, most of it comes back. The capacity was built once, and the body remembers how to find it again. Now immobilize the limb of a newborn for several years, denying the formation of strength, coordination, and the neural pathways governing movement. This is not the same condition as the adult where strength can be regained or recovered, since the capacity was never formed. Rehabilitation has nothing to rehabilitate.

That difference, specifically between a capacity weakened and a capacity that never formed, is the one the AI-and-cognition debate keeps missing. I think it is the whole question.

Generally speaking, the worry about AI and the mind concerns subtraction. We talk about losing focus, losing memory, losing the patience to think a problem through. The neuroscience behind that worry is sound. The brain essentially runs on use, with connections strengthening when they are exercised and weakening when they are not, and a faculty you stop exercising will quietly decondition.

Earlier in 2026, an MIT Media Lab team wired up participants writing essays with an LLM, with a search engine, and with nothing at all, and watched brain connectivity scale down as it received outside help. The most aided members later struggled to quote work they had finished minutes before.

The researchers called the result cognitive debt. A separate study from Microsoft and Carnegie Mellon, surveying hundreds of knowledge workers, found that the more someone trusted the AI, the less critical thinking they reported bringing to the task. A review of seventeen studies found similar results that AI lifts performance on lower-order work and erodes it on the higher-order reasoning that gets offloaded. Unfortunately, both those findings measured competent adults (knowledge workers, university students) and the results of their deconditioning when their capacity went unused.

Judgment is a capacity that you can only build the “old fashioned way” so to speak, which is doing it. You learn to synthesize by struggling to synthesize. You learn to write by failing to write and trying again. You learn to reason through uncertainty by sitting inside the uncertainty long enough that something in you reorganizes around it. The friction is not the obstacle to the learning. The friction is the learning.

So the logical question to me is what happens when an unformed mind uses AI? I would speculate that this mind cannot synthesize an idea before ever learning to synthesize. I doubt it could render good judgments without having practiced judging. Unfortunately, David Hubel and Torsten Wiesel demonstrated in experiments that a kitten deprived of normal input to one eye during a brief early window never develops normal vision in that eye, because the cortex wires around the deprivation and does not unwire later. Deprive an adult cat the same way and it recovers. The specific difference is that experience does not refine a circuit that already exists; it determines whether the circuit forms at all. The human version of that early visual deprivation is called amblyopia, which requires a similar treatment window before condition is permanent.

In full candor, while we know these windows exist for vision, hearing, and language, there is no long-term data that settles if they exist for higher-order capacities like synthesis and judgment. Regardless, I think it is completely logical and intuitive. Plus, in one sample, the preliminary results strongly suggest the youngest users leaned on these tools the most and scored the lowest on the thinking the tools displaced. The danger as mentioned earlier is that once missed, that capacity may never form and there is no rehab available. That asymmetry is why the formation question deserves more weight than the deconditioning one.

A good analogy of what AI can do to a mind is the scaffold or the cast. As a scaffold, AI can be a temporary support while a person builds the underlying capacity. As a cast, AI takes the load off entirely so output looks finished, but the developmental capacity was never formed.

I unintentionally ran the experiment for over fifteen years. My three sons played competitive chess, and the engines that can solve any position were available to all of them. The rule in our house was that you found the move yourself first, by sitting with the position, reasoning it through, and committing to a line. Only then could they check it against the engine. Most kids did the opposite, using the engine first to understand the position. I believe the order was everything as my sons not only excelled on the national and international chess scene, but were all early entrants into college.

So the practical question at this point is not whether to allow these tools. The “horse is already out of the barn” so to speak. That argument is over. The question is whether we can shape their use, and our children’s use, to keep the load on the human at the exact point a capacity is still forming. Get that right, and AI is the finest scaffold ever built. Get it wrong, and we will have cast a generation’s judgment before it ever bore its own weight and unlike a muscle, that is not the kind of thing that comes back.