The last human in the room

When economists and technologists discuss which jobs AI will displace, the conversation usually centers on capability: can AI do this task better than a human? But there's a second, less discussed question that determines whether a profession survives: even if AI can do the job better, does someone still need to be legally responsible for it? For a surprising range of high-stakes professions, the answer is yes, and that accountability requirement may protect those jobs far longer than raw capability comparisons would suggest.

Call it the "throat to choke" theory. Somewhere at the end of every consequential professional act, a legal opinion, an audited financial statement, a structural engineering stamp, a medical prescription, there needs to be a person. Not because the person necessarily did the work. Not because the person is better at the underlying task than an AI. But because when something goes wrong, society needs an identifiable human being who can be fired, sued, disbarred, or prosecuted. AI systems, as currently constituted, cannot be any of those things.

What "Throat to Choke" Actually Means

In the context of professional services, legal accountability has two components. The first is competence certification: you need a licensed attorney or CPA or physician because their license is evidence of demonstrated ability, regulated training, and ongoing professional standards. That part, AI is already eroding, GPT-4 passed the bar exam in 2023 at roughly the 90th percentile; frontier models today score higher.

The second component is accountability structure: when the legal opinion is wrong, when the audit misses fraud, when the diagnosis is incorrect, someone must face enforceable consequences. Consequences that matter, lost livelihood, legal liability, potential imprisonment. Consequences that, crucially, require a body. A corporation can be fined. A license can be revoked. A person can go to jail. An AI model cannot be any of those things in any legally meaningful sense today.

This is what Phil Trammell, Head of Economics at Epoch AI, called "transitional" on a recent episode of the Dwarkesh Podcast. And Alex Imas at Google DeepMind added a layer: "With lawyers particularly, you need some entity to back up the product. You need ownership of the product. You need somebody to be able to fire or hire, and there are licensing issues. There's a lot of regulatory layers that are also going to be keeping (even if there's no relational element) humans in the loop."

The key phrase is "even if there's no relational element." This is a different argument from the relational sector thesis, which holds that humans retain value because people want human involvement. The accountability argument holds that humans retain value because the legal system requires their presence, full stop. It doesn't matter whether the client wants a human lawyer. Regulation says one must sign the brief.

The Professions Most Protected Today

The clearest examples are the licensed professions, where a legally recognized credential is required to perform specific acts:

• Law: Bar admission is required to practice law in every US jurisdiction. An AI cannot represent a client in court, sign a legal brief, or provide formal legal advice without a licensed attorney taking responsibility for it. When a court document is wrong, the attorney is sanctioned, not the software they used.
• Accounting: A Certified Public Accountant license is required to audit a public company's financial statements under SEC rules. AI can perform the analysis; a CPA must sign the opinion. When the audit misses fraud, the CPA loses their license and faces civil liability.
• Engineering: Licensed Professional Engineers must stamp drawings for most commercial and public construction projects. That stamp means the engineer has personally verified the design. When a bridge fails, the PE of record faces consequences, regardless of whether AI generated the original design.
• Medicine: A medical license is required to prescribe medication, perform surgery, or formally diagnose disease. AI diagnostic tools are already more accurate than physicians in many domains, but the prescription still requires a licensed physician's signature. When the prescription harms a patient, the physician is liable.

In each case, the same structure applies: AI does an increasing share of the underlying cognitive work, but a licensed human must accept accountability for the output. The human has become, in a meaningful sense, primarily an accountability vehicle rather than a cognitive contributor.

The Near Term: The Signatory Era (2026–2032)

In the next several years, the most likely trajectory is what might be called the signatory era: AI does more and more of the work, but the human sign-off requirement stays in place. The licensed professional's role evolves from doing the work to supervising, reviewing, and certifying AI-generated work, and, crucially, accepting liability for it.

This is already happening. Law firms are deploying AI for contract review, legal research, and due diligence at scale. The associate hours have dropped; the partner signatures have not. Accounting firms use AI for audit sampling and anomaly detection; the CPA sign-off remains. Radiology AI reads CT scans faster and more accurately than humans; radiologists still sign the report.

The economics of this arrangement are uncomfortable for the professions involved. If the cognitive work is done by AI, the human's primary contribution is their license and their willingness to accept liability. The question then becomes: what is that worth? And who captures the value, the AI company that did the work, or the licensed professional who signed off?

In the near term, the answer will likely be determined by professional associations and regulators moving slowly to preserve existing structures. Bar associations have been notably resistant to allowing non-lawyers to own equity in law firms. Medical licensing boards have been cautious about expanding scope of practice. These organizations have strong institutional incentives to maintain the human requirement, and they have the regulatory authority to do so.

But the pressure will mount. If AI-assisted legal services can provide an adequate legal opinion for $50 that currently costs $500 for human review, there is an enormous access-to-justice argument for reducing the human requirement. Nearly 80% of Americans currently cannot afford legal help for civil legal problems, according to the Legal Services Corporation's 2022 Justice Gap Report. If AI could close that gap, the case for regulatory reform becomes harder to dismiss.

The Medium Term: New Accountability Structures (2032–2045)

The more interesting question is what happens when the signatory arrangement starts to break down, when there are too few licensed professionals to sign off on the volume of AI-generated work, or when it becomes obvious that the human signing the document isn't actually verifying anything meaningful.

Several alternative accountability structures are being seriously discussed:

Corporate liability for AI outputs. Rather than requiring a licensed human to sign off, the company deploying the AI assumes direct legal liability for errors. If a legal AI produces a wrong opinion and a client is harmed, the AI company pays. This model already exists in some forms, pharmaceutical companies are liable for drug harms regardless of whether a physician prescribed correctly. It's not a large conceptual leap to apply similar logic to AI professional services.

AI insurance as a substitute for licensure. Instead of requiring a licensed professional to backstop AI work, require the AI deployer to carry malpractice-equivalent insurance. The accountability function of the license (ensuring someone bears financial consequences for errors) is replicated by an actuarial mechanism rather than a professional one. Insurance markets already price risk; they could price AI malpractice risk.

Limited AI legal personhood. Some legal scholars, including Shyamkrishna Balganesh at Columbia and others working on AI governance, have proposed frameworks under which AI agents could hold limited legal personhood, enough to be named in a lawsuit, to have assets attached, to face enforceable consequences. This would be a significant legal innovation, essentially creating a new category of legal entity below the level of a corporation but above pure property. It would require legislative action rather than common law evolution, and would face enormous political resistance from existing professional lobbies.

None of these are imminent. Legal and regulatory systems change slowly, and the professions most affected by these changes have substantial political influence. The American Bar Association, the AICPA, and the AMA are not institutions that welcome frameworks that reduce the value of their credentials. But the medium-term window (roughly 2032 to 2045) is where the most interesting regulatory battles will play out, particularly as AI capability advances make the cognitive justification for human sign-off increasingly difficult to defend.

The Long Term: Is It Really Transitional? (2045+)

Trammell called the regulatory protection of human professions "transitional." That's probably right as a long-run prediction. The argument that a human must sign every legal opinion because humans are uniquely capable of legal reasoning has already been empirically undermined. The argument that a human must sign because humans need accountability structures that currently only apply to humans is a legal design choice, not a physical necessity. Legal design choices can be changed.

The deeper question is whether the accountability function of professional licensing can be separated from the human requirement. Today, we require a licensed human because that's the only mechanism we have for creating enforceable personal accountability in professional services. It is conceivable (not certain, but conceivable) that by 2045, we will have developed legal frameworks where corporate AI liability, mandatory AI insurance, or some form of AI legal personhood provides equivalent accountability without requiring a human body in the chain.

What would drive that transition? Three things seem most likely to accelerate it:

First, access pressure. The cost of human-mediated professional services is already a significant driver of inequality. As AI makes it technically possible to provide high-quality legal, medical, and financial services at a fraction of the current cost, the case for maintaining regulatory structures that keep those services expensive and scarce becomes harder to sustain politically.

Second, AI reliability demonstrated at scale. The current caution about removing human accountability requirements is partly reasonable, AI systems do fail in unexpected ways, and the consequences in legal, medical, and financial contexts can be severe. If AI systems accumulate a long track record of high reliability in these domains, the risk calculus changes. A system that is wrong less often than the average human practitioner, at a fraction of the cost, eventually challenges the accountability rationale for human sign-off.

Third, generational change in regulatory institutions. The people running bar associations, medical boards, and accounting standards bodies today trained before AI was a serious competitive threat to their professions. A generation of regulators who grew up with AI as a normal professional tool will make different decisions about how accountability should be structured.

What This Means for People in These Professions Today

If the long-run trajectory is toward reduced human accountability requirements, what does that mean for a lawyer, accountant, engineer, or physician in 2026?

The near-term answer is more reassuring than the long-term one. The signatory era will likely last at least a decade, long enough for a mid-career professional to build a practice around AI oversight, AI quality control, and the genuine judgment calls that AI handles poorly. The professions that survive the transition best will be those that use the signatory era productively: building expertise in supervising AI, developing deep specialist knowledge that AI still struggles to replicate, and positioning themselves as the human judgment layer for genuinely high-stakes decisions.

The more difficult long-term question is whether new entrants to these professions today are making a sound investment. A 25-year-old starting law school in 2026 will be in peak career years in 2040–2055, right in the window where the accountability structures that currently protect the profession are most likely to be challenged. That's not a reason not to pursue law or medicine. But it is a reason to think carefully about which parts of those professions are durable versus which parts rest primarily on regulatory protection that may not last.

The Honest Uncertainty

The accountability argument for human professional presence is real, and it is currently holding. Regulatory change in licensed professions is slow, and the institutions that would need to change have strong incentives to resist. The near term is reasonably clear: humans will remain in the accountability chain for high-stakes professional services for at least another decade, probably longer.

The long term is genuinely uncertain. Whether the "throat to choke" requirement evolves into new accountability structures that don't require a human body, or whether society decides that human accountability is intrinsically valuable and worth maintaining even as AI capability grows, is as much a political and philosophical question as a technical one. It will be decided not by AI developers but by bar associations, legislatures, courts, and the public, in a process that will be messy, slow, and contested.

That process is one of the most consequential ways in which our relationship with AI will be shaped over the next two decades. Whether it happens deliberately, with careful attention to the tradeoffs between accountability and access, or reactively, driven by crisis and political expediency, will matter enormously for how the transition unfolds. Being intentional about it, thinking through what we actually want accountability structures to accomplish, and whether current regulatory frameworks are the best way to achieve that, is exactly the kind of engagement with AI that deserves more attention than it gets.

Frequently Asked Questions