AI Is Too Important to Be Left To Its Believers

I.

Artificial intelligence has become the latest refuge of people who want the future to absolve them of precision. It is no longer only a technology. It is now a permission structure. A budget becomes strategic if AI is attached to it. A product becomes inevitable. A founder becomes visionary. A consulting deck becomes transformation. A government program becomes sovereignty. A layoff becomes efficiency. A valuation becomes destiny. The word has grown so large that it can now hide almost anything inside it.

This is usually how a serious technology begins its descent into bad faith. First, something real appears. Then the priests arrive. They do not invent the miracle; they monetize it. They wrap it in inevitability, sell urgency to the anxious, sell superiority to the early, sell fear to the late, and sell vocabulary to everyone. A few years later, nobody is discussing the thing itself. They are discussing the social obligation to believe in the thing.

AI is now at that point. The question is no longer whether the models are powerful. They are. The question is whether the civilization talking about them has become capable of ordinary judgment. On that, the evidence is less encouraging.

II.

The models are real. That must be said before anything else, otherwise the argument becomes too cheap. These systems can write, summarize, translate, classify, compare, generate code, inspect documents, create images, produce drafts, simulate expertise, search across messy knowledge, and compress certain kinds of cognitive labor with astonishing force. To call this a fad is not skepticism. It is laziness with a pipe in its mouth.

But it is equally lazy to treat every capability as destiny. This is the grand error of the present AI narrative. It takes the fact that a machine can produce something intelligent-looking and quietly converts that fact into a theory of society. Because the model can answer, the organization will change. Because the model can code, the profession will collapse. Because the model can reason inside a prompt, the enterprise will become autonomous. Because the demo worked, the future has already voted.

This is childish. A demo is not a deployment. A deployment is not adoption. Adoption is not transformation. Transformation is not value. Value is not achieved until the old world of incentives, accountability, workflow, trust, regulation, fear, and politics has been forced to move. That world moves slowly. It has survived stronger sermons than this.

III.

There is also too much money in the room for the conversation to remain honest. Recent estimates put AI-related capital expenditure at staggering levels. Goldman Sachs’ baseline model projected annual AI capex of about $765 billion in 2026, rising toward $1.6 trillion by 2031; McKinsey projected that data centers may require $6.7 trillion worldwide by 2030, with $5.2 trillion tied to AI workloads. Reuters has also reported expectations of AI-related Big Tech spending crossing hundreds of billions in 2026, alongside a surge in AI-related debt issuance.

At this scale, belief is no longer innocent. It has payroll. It has debt schedules. It has equity stories. It has national policy behind it. It has warehouses of GPUs waiting to be justified. It has consulting practices to feed, public markets to soothe, employees to frighten, and boards to impress. The AI story is no longer merely a forecast. It is collateral.

This is why maximalists sound so certain. Not always because they have seen farther, but because they have already bought too much land in the promised country. Their conviction has become part of their balance sheet. The founder needs the story. The investor needs the story. The hyperscaler needs the story. The consultant needs the story. The executive who just announced the strategy needs the story. The analyst who upgraded the stock needs the story. Once enough people need a belief, evidence begins to arrive already domesticated.

Motivated reasoning is not an accident in the AI age. It is the atmosphere.

IV.

Software engineering is the strongest argument for AI and also the source of the worst confusion. Code is unusually hospitable to these models. It is language under discipline. It has patterns, syntax, examples, tests, errors, logs, compilers, documentation, and repeatable feedback. A model can generate code, fail, read the failure, regenerate, and improve. One could explain this technically. That would be useful, and also beside the point. The machine’s advantage in software is not merely that code resembles language. It is that code forgives. Most of life does not.

That last sentence matters. Code can be regenerated. A patient cannot. A credit decision cannot. A termination letter cannot. A regulatory breach cannot. A court filing cannot. A production outage cannot. A public policy cannot. A child’s education cannot. A safety incident cannot. The world outside software is full of acts that do not become harmless because they began as drafts.

So yes, AI will change software deeply. In many engineering teams it already has. It makes prototyping cheaper, boilerplate less annoying, debugging faster, translation between intent and implementation easier. It rewards people who know what they are trying to build and punishes people who only know how to ask for code. But software is a flattering domain. It lets AI look like a general solution because software itself is already a language machine. To generalize from code to life is to mistake the cleanest room in the hospital for the whole body.

V.

Most enterprises do not fail because they lack answers. They fail because answers have nowhere to go. Everyone who has worked inside a large organization knows this. The deck may be clear. The strategy may be correct. The dashboard may show the problem. The consultant may have named the gap. The pilot may have proved the concept. Still nothing moves, because the system is not waiting for intelligence. It is waiting for permission, ownership, courage, budget, alignment, governance, or someone senior enough to absorb blame.

AI enters this world and is immediately misunderstood. Leaders imagine they are adding intelligence to a process. Often they are adding speed to confusion. The workflow is unclear, the data is poor, the approval chain is political, the risk boundary is vague, the user has no reason to change behavior, and nobody wants to own the final decision. Then the organization wonders why the expensive model has produced only theatrical improvement.

This is why so many implementations fail in the most respectable way possible. No scandal. No dramatic collapse. Just pilots that remain pilots, assistants that become ornaments, knowledge bots that work until the question matters, productivity claims that cannot survive measurement, and transformation programs that quietly become license management exercises. Most organizations say it is becoming AI-first. The employees continue copying outputs into the same old workflows, asking the same old managers for the same old approvals. The machine has not failed; the organization has confessed.

VI.

The dirtiest trick in the AI narrative is the substitution of output for value. AI produces outputs beautifully. That is its narcotic. The organization sees abundance and mistakes it for progress.

But output was never the deepest scarcity. In many places, we already had too much output. Too many things that look complete because they are written down. AI now threatens to multiply this debris. It can make mediocrity fluent. It can make confusion presentable. It can give laziness the manners of competence.

The real scarcity is judgment. What should be done? What should be ignored? What is safe? What is true enough to act upon? Who is accountable? What should not be automated? Where does speed create fragility? Where does a human have to remain in the loop not because of sentimentality, but because someone must be answerable when reality bites?

AI reduces the cost of producing options. It does not remove the burden of choosing. In some places, it makes that burden heavier. People often confuse the ability of models to reason with judgment. These are different gifts.

VII.

The romance around agents is where the conversation becomes most unserious. The language is almost religious. The agent will understand the goal, plan the task, call the tools, remember the context, execute the workflow, correct itself, and report back. A tireless digital, synthetic colleague and a servant with APIs.

This fantasy survives because people speak about action as though action were only execution. It is not. Action is liability entering the world. A chatbot that hallucinates is embarrassing. An agent that changes a firewall rule, approves a loan, denies a claim, alters a medical record, triggers a refund, escalates a security event, or sends a legal notice is not embarrassing. It is dangerous unless the boundaries are exact.

Serious AI pays a safety tax. Access control, audit logs, escalation paths, human review, policy constraints, red teaming, rollback, monitoring, abuse testing, data boundaries, exception handling, responsibility assignment. These are not bureaucratic decorations. They are the price of letting probabilistic systems touch real consequences. The maximalist calls this fear. Adults call it engineering.

VIII.

The deeper philosophical error is the belief that intelligence is the master bottleneck of civilization. This is the superstition of clever people. They assume the world is broken because it lacks enough cognition. Add cognition, the world improves. Add more cognition, the world transforms. Add artificial cognition at scale, the world reorganizes.

But human systems are rarely so innocent. They are not merely intelligence systems. They are incentive systems, status systems, fear systems, memory systems, legitimacy systems. A company may know exactly what must be done and still not do it. A government may possess the data and still lie. A hospital may have the protocol and still fail. A leader may understand the risk and still choose the politically survivable option. A team may have the correct answer and still lose to hierarchy.

AI can generate answers. It cannot automatically create the conditions under which answers become action. It can produce analysis. It cannot make a coward brave. It can summarize risk. It cannot make an institution honest. It can recommend decisions. It cannot carry moral responsibility for deciding.

This is why the slogan that AI will replace humans is too crude. It will replace tasks. It will compress workflows. It will destroy some roles. It will create others. It will cheapen first drafts of almost everything. But the human problem will not disappear. It will move. From production to judgment. From drafting to deciding. From searching to verifying. From doing to owning. Scarcity does not die. It migrates.

IX.

The priests of inevitability hate migration. They prefer apocalypse. It sells better. Either AI replaces everyone or fools deny the future. Either you are early or dead. Either you automate or vanish. This binary is useful for frightening buyers, but useless for understanding history.

No serious technology arrives in the shape promised by its loudest believers. It is bent by law, absorbed by incumbents, distorted by markets, resisted by professions, domesticated by bureaucracy, cheapened by vendors, weaponized by politics, and misunderstood by almost everyone for the first decade. The internet was real. It still produced a sewer. Cloud was real. It still created new dependencies. Crypto had real ideas. It still became a carnival of leverage and vocabulary. Social media connected the world and then taught it to perform pathology for engagement

AI will be no purer. It will do magnificent work. It will also launder responsibility, flood organizations with polished nonsense, automate bad processes, create new security failures, reward shallow operators, deepen surveillance, and give incompetent leaders a new language for old avoidance. This is not an argument against AI. It is an argument against innocence.

A technology does not become safe because it is powerful. Power only increases the consequences of stupidity.

X.

The serious position is therefore not balance. Balance sounds too polite. The serious position is disciplined belief. Believe the capability. Distrust the sermon. Use the tools. Refuse the mythology. Measure the value. Inspect the workflow. Demand ownership. Separate software gains from civilizational claims. Ask where the model is strong, where the process is weak, where the risk sits, who benefits from the story, and who carries the cost when the story fails.

The skeptic who calls AI a fad is unserious. The maximalist who treats every objection as fear is worse. The skeptic is avoiding the burden of learning. The maximalist is avoiding the burden of precision. Precision is the enemy of hype because it asks small humiliating questions. Which task? Which data? Which user? Which workflow? Which approval? Which failure mode? Which metric? Which owner? Which consequence?

That is where AI becomes real. Not in prophecy. Not in keynote language. Not in the fantasy of a machine that dissolves every institution it touches. It becomes real in the narrow places where work is redesigned, risk is bounded, and value survives measurement.

AI is not fake. Its impact will not be small. But the story around it has become swollen with money, vanity, fear, and the old human desire to mistake a tool for salvation. The people who benefit most from AI will not be those who worship it first or mock it early. They will be those who can look at it without needing it to become a god.

Originally published on LinkedIn