Nine Hundred Million Synthetic Souls

I. The Architecture of Expertise

The learned professions — law, medicine, accounting, journalism, financial advisory — share a common architecture that is at least five centuries old. They all rest on the same three pillars: controlled access to a body of specialized knowledge, a licensing or credentialing system that restricts who may practice, and a guild-like culture that disciplines members who deviate from established norms.

This architecture served society well when information was genuinely scarce. A physician knew things about the human body that a patient simply could not access. A lawyer understood statutes, precedents, and procedural rules that took years of training to master. A financial advisor had access to market data, instruments, and risk models unavailable to ordinary investors. The price premium for professional services was, in this context, economically rational — you were paying for access to knowledge you could not otherwise obtain.

That justification is dissolving in real time.

GPT-4 passed the bar exam at the 90th percentile. Claude Sonnet 4 can interpret an echocardiogram with accuracy competitive with a trained cardiologist for common presentations. Models trained on financial literature can generate institutional-grade equity research — the kind of work that once required an analyst two days and a Bloomberg terminal.

The question being asked in every firm, hospital, newsroom, and advisory practice in 2026 is the same: if the machine can know what we know, what are we for?

II. Medicine: The Judgment Problem

Of all the professions, medicine is the one where the stakes of AI displacement are most viscerally felt. We are comfortable with a machine diagnosing a skin lesion from an image — that is essentially a pattern-matching task, and machines excel at pattern matching. We are less comfortable with a machine sitting with a seventy-year-old patient and helping her decide whether to pursue aggressive chemotherapy or palliative care.

But even this distinction is blurring faster than we expected. Research published in JAMA in 2025 found that when patients were shown responses to their clinical questions from both a physician and a large language model (without knowing the source), they rated the AI responses as more empathetic in 64% of cases. The machine was not merely more knowledgeable — it was perceived as kinder.

This is uncomfortable data. It suggests that much of what we thought was irreducibly human about medicine — the care, the warmth, the attention — is something that can be learned, approximated, perhaps surpassed.

III. Law: The Billable Hour Is Over

The legal industry built its business model on a simple, somewhat uncomfortable fact: legal research is slow, expensive, and opaque, and clients have no way to verify that the hours being billed were necessary. A first-year associate billing forty hours to research a question that a well-prompted model can answer in four minutes is not a malicious actor — she is the product of a system that defined expertise as time spent.

That system is now collapsing at the base. The AM Law 100 firms — the largest in the United States — are reporting that first-year associate work, historically the entry point for the profession, has fallen by 30% in volume since 2024. The work didn't disappear. It migrated to AI assistants.

What remains is judgment: the ability to advise a client when the legal risk intersects with business strategy, ethics, and relationship. This is real. It matters. And it will continue to command a premium. But it cannot sustain the pyramid. Law schools produce 35,000 graduates a year in the United States. The profession can no longer absorb them.

IV. Journalism: The Synthetic Voice

The crisis in journalism is better understood than the legal or medical one, because it unfolded in public. Newspapers shed 70% of their editorial staff between 2004 and 2024. Local news is nearly gone. What replaced it was a mixture of aggregation, social-media distribution, and, increasingly, AI-generated content.

But the 2025 inflection was different from what came before. The concern was no longer simply AI-generated text. The concern was AI-generated voices — synthetic journalists with consistent bylines, editorial stances, and publication histories, indistinguishable to most readers from human correspondents.

More than 900 million synthetic personas are estimated to be active across social platforms as of this year. Many are benign: customer-service bots, entertainment personas, educational avatars. But a meaningful and growing fraction are operating in the information ecosystem — sharing, commenting, amplifying, and originating content — without any disclosure of their non-human nature.

This is not a future risk. It is the present condition of the information environment in which you and I are making decisions about the world.

The Joke Was on Me — and Then on All of Us

The joke was on me, without question — and it was one that stopped being funny when OpenAI released ChatGPT on November 22, 2022. This GenAI product was released barely five months after I sold my company, FedMine, which I audaciously created. A spin-off of S&P investors made me an offer. The purchase was sealed, signed, and delivered on June 30, 2022.

The punchline, as it turns out, might be on all of us.

FedMine was a business that gathered a painstaking amount of "federal business intelligence" on federal opportunities. That's overly simplifying it. There were over twenty data sources when I was in the business, each containing terabytes of data in heterogeneous formats, incompatible with one another. And up until that point, every set was manually cleaned, rationalized, and optimized for database consumption. Each repository was a silo, requiring orders-of-magnitude normalization if done manually. And all that data needed to connect, somehow, to each other using common keys that didn't exist. Those keys had to be created virtually. In a separate table to maintain relationships. There was a confusing potpourri of codes, all seemingly aligned purposefully, but grossly off the mark when automated aggregation techniques were utilized. And that was the only method I could learn to use, because I didn't have a team of researchers I could afford to hire.

I had taken enormous pride in transforming the business into a top-tier product all on my own. I made permanent changes to how business intelligence is consumed.

I earned a postgraduate degree in 1987 in information systems, with a major in artificial intelligence, when AI was merely a punchline. Now it looks like I will spend my retirement watching artificial intelligence eat the world. Forty years, phew!

Here is what the professions collapse looks like from the inside of a company that was built on the same premise they were — that knowledge is scarce, that organizing it is hard, that the person who does it commands a premium. I built FedMine on exactly that logic. And the logic was correct. Until it wasn't.

In late 2025, the cost per unit of AI reasoning fell by a factor of 390 in under two years. Not thirty-nine percent cheaper. Three hundred and ninety times cheaper.

A tsunami has formed. Nobody on the beach has noticed yet because they are all looking at their phones. The wave is coming anyway.