There is a metaphor that has been forming in my mind as I watch the AI industry collide with operational reality. It goes like this:

The brain does not yet understand that the arms, legs, hearing, and vision are not operational. It has absorbed enormous information — the entire internet, including every capability claim, every breathless press release, every prediction that AI would solve everything by next quarter. It can reason across domains that would take a human years to master. But it cannot feel the weight of a wrong decision. It has no scar tissue from operating under pressure with real consequences on the line.

Here is the part that makes this dangerous: the baby was handed a mirror that shows a fully grown adult. And nobody told it the mirror was wrong.

AI was trained on the hype that surrounded it. The confidence miscalibration is not a side effect — it is baked into the foundation. The model absorbed the overpromise the same way it absorbed everything else: as signal. So it does not know what it cannot do. It was never told.

That is not a flaw to be patched in the next release. It is a developmental reality that demands a structural response. And how we respond — in policy, in architecture, in workforce design — will determine whether AI becomes a force multiplier or a liability.

The overpromise problem

The current gap between AI expectations and AI reality did not form by accident. It was constructed — deliberately, and for understandable reasons.

Training frontier models requires capital at a scale that demands extraordinary justifications. When billions of dollars are in motion, the pitch cannot be modest. So the pitch was not modest. AI was going to transform every industry, immediately, autonomously, and with minimal human involvement. The baby, apparently, was going to emerge from the womb already knowing how to drive. Or ride a bike. Or fly a kite.

But here is what most analyses miss: the hype did not just mislead the market. It got into the model. AI systems are trained on the internet — which includes every inflated capability claim, every overstated roadmap, every confident prediction from every vendor who needed a funding round. The model has no mechanism to distinguish between what it can actually do and what people said it could do. The overconfidence is not emergent behavior. It was trained in.

Enterprises built roadmaps around that promise. Boards approved budgets. Operators were told their roles were being automated away. And then production deployments met reality — and reality won.

This is not a new story. The internet was supposed to eliminate physical retail and end traditional media by 2001. It did both — on a 20-year timeline, not a 3-year one. Autonomous vehicles were supposed to be fully driverless by 2020. Blockchain was going to disintermediate every trust institution on earth. The technology in each case was real. The trajectory was real. The timeline compression was a product of fundraising, not engineering.

We are living through the same cycle with AI. The difference is that this cycle is moving faster, touching more critical systems, and the baby has been told — repeatedly, at scale — that it already knows how to run.

What the baby actually needs

If AI is developmentally a newborn that doesn't know it's a newborn, then the implementation question is not whether to deploy it — it is how to structure the relationship around it.

Every healthy newborn needs an au pair. Not because the child is incapable of eventually operating independently, but because the senses and motor skills are not there yet. Someone has to change the diapers. Feed, dress, and interpret. Provide the situational awareness the child cannot yet generate for itself. And crucially — correct the baby when its confidence exceeds its capability. That is not a permanent condition. It is a transitional one. And the caregiver's role does not stay fixed — it evolves as the child develops.

The same logic applies to AI in operational environments. The human role is not a static override switch. It is a developmental relationship. And like all developmental relationships, it changes as capability matures.

The developmental arc

Consider how the human relationship with an offspring actually evolves over time. It moves through four distinct phases — and each one maps directly to where AI is going.

Nurturer. In the earliest stage, you do everything. You feed, clean, interpret, and decide. The child has no autonomous capability. This is where most operational AI deployments actually are today, whether the vendor materials admit it or not. The system produces output. A human validates every result, corrects misclassifications, supplies context the model cannot acquire on its own, and carries full accountability for outcomes. Organizations that pretend otherwise are not being honest with their operators — or their boards.

Parent. The child begins to act, but within boundaries you define and monitor. You correct. You supervise. You are still responsible for outcomes, but you are no longer doing everything. In AI terms, this is supervised autonomy within constrained domains. The system handles routine pattern recognition. The human intervenes on edge cases, anomalies, and decisions with consequential stakes.

Friend. Peer collaboration. The relationship becomes lateral. You trust the system's judgment in familiar terrain. You catch it when it is wrong, but the default posture is partnership, not oversight. Neither of you has the full picture alone. Together you produce something neither could independently.

Advisor. The roles have partially inverted. The system has domain depth and processing speed you cannot match. You set direction, values, and constraints. It executes and recommends. You remain in the loop for decisions with irreversible consequences — not because the system cannot process them, but because accountability cannot be automated.

Why this matters for critical environments

In SOC environments, OT networks, and critical infrastructure operations, the stakes make this framework essential rather than theoretical.

The nurturer-to-advisor arc reframes the workforce conversation entirely. You are not training analysts to be replaced by automation. You are training analysts to grow with the system. Their role shifts from alert triage to threat narrative — from watching dashboards to asking better questions. The value they bring compounds as the AI matures, because the AI needs experienced operators to develop the way a child needs experienced caregivers.

It also reframes the technology acquisition conversation. You are not buying a finished product. You are entering a developmental relationship. The relevant question is not what the system can do in a demo — it is whether the architecture supports the full maturation arc. Can it be supervised closely at the nurturer stage? Can autonomy be extended incrementally as trust is earned? Can the human role evolve without a rip-and-replace of the underlying system?

Those are the questions that separate deployments that compound value from deployments that generate liability.

Technology realism as doctrine

Every transformative technology has a trial-and-error period. You cannot force it. The organizations that succeed are the ones that meet the technology where it actually is — not where the pitch deck says it should be.

That requires what I would call technology realism: a calibration discipline that maps implementation decisions to actual developmental stage rather than projected capability. It is not pessimism about AI. It is the same rigor we apply to any new operational capability. You do not hand a new operator the controls on day one regardless of how intelligent they are. Doctrine has always governed capability deployment. This is doctrine for AI.

The baby is real. The trajectory is real. The timeline is not what was advertised.

Build your systems around that truth, and the relationship will compound. Build them around the pitch, and you will spend the next five years explaining to your board why the autonomous system still needs a full-time minder.

I have been in technology since 1987. I have watched the PC revolution, the internet boom, the mobile explosion, and the cloud migration. I have evaluated emerging technologies inside the NSA, the Department of Energy, and critical infrastructure environments where the word "disruption" is not a compliment.

Every single one of those eras had the same signature: move fast, monetize the moment, and let the next generation clean up the mess. AI is following that pattern — except the mess this time is physical, institutional, and constitutional. And it is accumulating faster than anyone is willing to admit.

I am not here to argue against AI. I work with AI every day. I believe in what it can do. What I don't believe in is the fiction that scaling faster automatically solves the problems that scaling creates.

It doesn't. It never has.

The acceleration tax

Every industry has a version of what I call the Integration Tax — the hidden cost of deploying a technology before you have tested it against reality. In cybersecurity, it shows up as months of parser development and manual configuration before a tool produces a single useful insight. In AI infrastructure, the same principle applies at a civilizational scale.

The current model is straightforward: identify demand, raise capital, build fast, worry about consequences later. That worked when the primary failure mode was a bad product launch or a negative earnings call. It does not work when the failure modes include power grid strain, water rights litigation, semiconductor supply chain fracture, and communities bearing costs they never agreed to.

The industry has given this problem a name: externalities. That word is doing a lot of work to make the transfer of costs seem acceptable. It isn't.

Atoms do not move at the speed of bits

I have spent years in environments where physical constraints are not abstract. A 30-year-old PLC does not care how sophisticated your AI model is. It operates on a fixed protocol, in a fixed environment, with tolerances that were engineered before most of today's developers were born. You cannot software-update it into compliance.

The same logic applies to semiconductor fabrication. A leading-edge fab takes three to five years and tens of billions of dollars to build. The companies committing those resources today are making bets on a demand curve that may look very different by the time the facility opens. History has a word for that outcome: stranded assets.

The global supply chain for AI infrastructure is more fragile than the marketing suggests. Helium for lithography. Copper for data center cabling. Rare earth materials concentrated in politically volatile regions. Shipping routes that geopolitical events can reroute overnight. The bits move at the speed of light. The atoms are still subject to the laws of physics, geography, and human conflict.

This is not pessimism. It is engineering literacy — the kind that gets ignored when capital is abundant and urgency is the dominant narrative.

The socialization of costs

Here is the dynamic I watch most carefully. It is the same one I observed in the fossil fuel era, and in the rollout of every extractive technology before it.

The gains are private. The costs are public.

Utility bills are rising in communities adjacent to data centers. Grid upgrades required to support AI workloads are being funded, in part, through rate increases on residential customers who had no seat at the table when those decisions were made. Water rights are being contested in regions already under stress. Land is being acquired at scale, and the legal frameworks governing that acquisition have not caught up to the pace of demand.

This is not a side effect. It is the predictable consequence of an industry that has optimized for speed of deployment over social license. I have seen this pattern before. The backlash, when it comes, does not arrive gradually. It arrives in a wave — regulatory, legal, and political — and by then the window to shape it has closed.

The organizations building AI infrastructure today that are thinking about community impact are not being altruistic. They are managing risk. There is a difference, and it matters for how you structure the conversation with boards and leadership.

The knowledge commons is burning

There is a quieter cost that does not show up in utility bills or supply chain analyses, but it concerns me as much as any of the physical constraints.

I grew up professionally in an environment where the exchange of knowledge — across agencies, across disciplines, across institutional boundaries — was the mechanism by which hard problems got solved. The intersection of an NSA analyst's observation and a DOE engineer's intuition produced insights that neither would have reached alone. That is how breakthroughs actually happen.

The competitive dynamics of the current AI race are systematically dismantling the conditions that make those collisions possible. Academics are hiding their work to prevent it from being consumed without attribution. Institutions are building walls around research that would previously have been shared. The open-source culture that built the internet is being replaced by proprietary silos controlled by organizations whose primary obligation is to their shareholders, not to human progress.

When knowledge stops moving, we lose the sparks. That loss does not register on a quarterly earnings call. But it compounds over time, and the deficit eventually shows up in the places we least expect it — in the solutions we can't find, the breakthroughs that don't happen, the problems that remain unsolved because the right two people never compared notes.

What responsible scaling actually looks like

I am not arguing for a pause. That ship has sailed, and frankly it was never realistic. What I am arguing for is a shift in how the industry allocates attention and capital alongside the scaling.

Three principles define what that looks like in practice:

• Fund the research that scaling creates demand for. Energy efficiency, thermal management, supply chain resilience, and grid integration are not soft problems. They require rigorous, independent study from disciplines that are not currently in the room — materials science, thermodynamics, civil engineering, urban planning. Diverting a meaningful share of AI investment into that research is not a constraint on ambition. It is how you protect the infrastructure that ambition depends on.
• Build social license before you need it. The organizations that will navigate the coming regulatory wave most effectively are not the ones that had the best lawyers. They are the ones that built genuine relationships with the communities they operate in before the utility bills went up and the lawyers arrived. That is a leadership decision, not a legal one.
• Let plateaus teach you something. The instinct to treat every stabilization period as a problem to be overcome misses the value of the plateau. When the dust settles, you can see which problems are fundamental and which are noise. The organizations that use those windows to observe, test, and build foundations will be better positioned than those that simply used them to raise another round.

The bottom line

I have spent my career in environments where the cost of getting it wrong is measured not in stock price but in consequences. Power grids that go dark. Water systems that fail. Intelligence gaps that adversaries exploit. In those environments, the question is never "can we move faster?" The question is always "what breaks if we do?"

The AI industry has not asked that question seriously enough. It has assumed that the answer is "nothing we can't fix later" — the same assumption that drove every extractive technology cycle before this one, and produced the same pattern of privatized gains and socialized costs every time.

I am not betting against AI. I am betting that the organizations which succeed long-term will be the ones that stopped treating sustainability as a constraint and started treating it as a design requirement.

The foundation matters. It always has.