In the Spring of 2026, a paper published by the MIT Sloan School caused a bit of a stir among those who take an interest in the long-term economic implications of artificial intelligence. Although “Some Simple Economics of AGI” fails its titular promise of simplicity, it elucidates the problems of AI applied at scale with remarkable clarity. It is simultaneously an economic model, a scenario document, a civilisational warning, and a set of recommendations to avoid an AI-powered societal meltdown.

Most economic models, the authors argue, misunderstand AI completely by not paying sufficient attention to the bottleneck that decides its true potential – human beings. Whether or not we can build a truly automated and augmented future depends less on the evolution of machine intelligence, and more on the ability of humans and machines to achieve a kind of symbiotic alignment. Someone needs to check and audit the output of AI, and these capabilities are embodied, built up over time, and difficult to synthesize. Our finite capacity to validate and underwrite machine work complicates the idea of an imminent economic AGI revolution while also highlighting the plausible counter scenario: a slop future where understaffed firms drown in endless piles of useless machine work.

We spoke to one of the paper’s authors, Christian Catalini, to learn more. Catalini is the Founder of the MIT Cryptoeconomics Lab and a Research Scientist at MIT Sloan School of Management.

---

This paper has a grand scope, but it makes the problem very clear. As agentic labour grows, human verification remains bound by the biological constraint of embodied experience. This pushes us towards a “hollow economy”, characterised by endless volumes of unverified and useless machine output. You contrast this prospect with the “augmented economy”, in which human verification and underwriting is sufficiently scaled with machine productivity, leading to a historic economic transformation.

What does this augmented economy, and the human scaling required, look like?

The augmented economy is, at its core, a verification infrastructure project. The Measurability Gap – the distance between what humans can understand, audit, and stand behind, and what machines can produce – is the variable we have to manage. The hollow path lets that gap widen until trust collapses. The augmented path closes this gap by building the tools, institutions, and data infrastructure that let us safely scale agentic output without losing track of whether that output is actually serving human intent.

Closing that gap has two parts. The first involves building outward: public ground truth registries, standardized incident reporting, audit infrastructure, liability frameworks that price unverified deployment, and observability tools that compress agent behavior into actionable human signals. This is the institutional work, and it determines whether the agentic economy runs on trust or on hope.

The second is building inward by augmenting humans themselves. As models get smarter, we will need to augment our own capabilities to remain peers with what we have built. That includes synthetic practice environments and AI-native verification stacks for individuals and firms. Eventually it includes direct neural augmentation – brain-computer interfaces that let us keep up with systems whose pace of cognition will outstrip the unaided human bandwidth. The point is not to merge with the machine. It is to ensure that when we delegate, we still understand what we have delegated, and that our verification capabilities scale alongside the systems we deploy.

If scaling human verification requires technological augmentation, does this not merely create new issues of verification and risk, further adding to the biological burden?

Yes, that concern gets to the heart of the Measurability Gap. Layering more technology onto the problem can absolutely introduce new vectors for risk and error. But the solution is not to reject augmentation. It is to be precise about what kind of augmentation is needed. The goal is to give humans leverage.

Three things have to come together. The first is observability as verification leverage – tools that don’t just generate more data for humans to review, but compress agent behavior into signals humans can actually audit. Think of a cockpit for a skilled pilot: high-dimensional behavior summarized, anomalies surfaced, auditable traces preserved. Used well, these systems reduce the biological burden of oversight rather than adding to it. The second is what we call verification-grade ground truth – a class of data that includes not just successes but edge cases, near misses, failure logs, audit trails, and postmortems. This is the data that teaches both humans and machines what to reject and how to verify, and it directly addresses the new risks that emerge with each new layer of augmentation. The third is a regulatory environment in which the new risks must be priced and internalized by the market. Standardized incident reporting, auditable logs, and clear disclosure formats make the risk legible. Once a risk is legible, it can be priced.

That shifts the burden from an unquantifiable biological load on individuals to a concrete, economic cost for deployers – and it creates powerful incentives to build the verification and observability tools above.

So while new technologies do create new risks, they also create new possibilities for managing those risks. The solution is not to slow down. It is to shift the basis of competition from raw, unverified output to verifiably safe and effective performance. That transforms the verification problem from an infinite, biological one into a finite, economic, and technological one.

You highlight what you see as the barriers to getting us there, in how the automation mechanisms that maximise short-term efficiency also erode the capacity for long-term human verification.

There’s the “Missing Junior Loop”, which is the automation of entry-level jobs and the collapse of traditional apprenticeship pathways. We see this today in firms not hiring as many juniors due to AI. Then there’s the “Codifier’s Curse”, where experts feed their tacit knowledge to the machine and in turn contribute to their own automation. Finally, there’s the “Alignment Drift” – as AI systems take on more tasks, humans lose the capacity to verify what they’re actually doing.

How can trustworthy (human) verification layers be scaled successfully in the face of these strong counterforces?

The Missing Junior Loop and the Codifier’s Curse are two of the most significant barriers to a future where AI is safely integrated into our economy. They create a vicious cycle: automation erodes the human expertise needed to keep automation in check.

Scaling trustworthy human verification is possible, but it requires a shift in two dimensions: in our economic incentives, and in how we develop human capital. On the incentive side, we need to shift from raw, frictionless scale to what we call “verified network scale”. In an economy filled with AI agents, the most valuable networks will not be the largest. They will be the most trustworthy. Authenticated, high-trust activity becomes the moat, and that creates real economic demand for human expertise.

On the human capital side, we need to consciously rebuild the expertise supply chain that automation threatens. If the market no longer provides the reps necessary to build experience through junior work, we have to create them. We advocate for synthetic practice environments – high-density, simulated environments where people can be exposed to a wide range of scenarios, edge cases, and feedback loops, building the tacit knowledge that was once the byproduct of entrylevel jobs.

This applies as much to seniors as to juniors. Junior talent uses synthetic practice to compensate for the bottom rungs of the career ladder that have been automated away. Senior talent uses it to escape the Codifier’s Curse – to push their judgment into novel territory faster than the next model can absorb their existing expertise. In both cases, AI itself becomes the medium through which humans build the intuition needed to verify AI. Same tool, different stage of career.

In short, we scale human verification not by trying to preserve the old system, but by building a new one. We create an economy that pays for trust, and we use AI to help build the next generation of human experts – at every level – who can provide it.

You treat AGI as an imminent and almost assumed development, writing that “over the next 12 to 24 months, the global economy will be transformed”. Does this assumption imply that agents and models are already at the necessary level of development for this transformation to occur – granted human verification can be scaled to keep up?

Yes. And part of the reason we have not yet seen the full effects of current capabilities is that for many end-to-end tasks, verification is the last mile and the final bottleneck. The model can do the work; the friction is in checking that the work is right and standing behind it.

That bottleneck is what shapes the timeline. As exponential improvements in capability continue, two things will happen in parallel. More jobs will become fully automatable, and far fewer individuals will be needed to deliver the throughput of an entire team. Both pressures will reshape the labor market, and the speed of that reshaping creates the sense of imminence in the paper.

It is also, of course, a massive opportunity. Every prior wave of automation eliminated existing roles and created new ones we could not have predicted in advance, and there is no reason to think this wave will be different. The harder question is whether we build the institutions to define and capture those new roles in time. And as capabilities continue to compound, our ambitions will compound with them. We will want to do more, attempt harder things, and expand the scope of what we expect from ourselves and the systems we deploy. The verification problem does not get smaller. It gets more important, because the stakes of getting it wrong grow with our reach.

The paper has an implicit emphasis on the logic and processes of tech work. The Codifier’s Curse seems especially true for coding and IT engineering. ‘Verification capacity’ may be a universal constraint, but one that looks very different across sectors and fields. What’s the case for this being a generalisable prospect for the automation of knowledge work, also in fields with more elusive measurability?

Software is special. It has the fastest feedback loops, most pre-existing formalization, densest infrastructure of automated tests, so the dynamic surfaces there first. But software is just the beginning. The underlying force is measurability, and measurability cuts across every form of knowledge work.

The vulnerable task is not the simple one. It is the one whose inputs, outputs, and feedback can be tracked and graded – even when the work historically required elite credentials. Legal due diligence, financial modeling, medical triage, journalism, parts of scientific research. The Codifier’s Curse runs in all of them: every time a senior expert corrects an AI output, they generate a high-fidelity training signal that converts their tacit knowledge into machine-readable form. The expert rationally monetizes their scarcity right up until the moment the market clears it.

A note on the usual rebuttals. When people invoke scope, taste, or judgment as the categories that protect their work from AI, those are mostly cope. The frontier of what can be measured is not static, and AI is the engine that expands it. Modern models are exceptionally good at converting unstructured material – text, images, voice, complex behavior – into structured signal. A field that looks elusive today is one telemetry stack away from being legible tomorrow.

That said, some domains do genuinely escape measurability – and those are the places where the highest-leverage human work will concentrate. The paper highlights three. The first is deep tech and frontier research, where there is no existing ground truth to train against and progress depends on hypothesizing the right experiment in the first place. The second is fields built on social consensus and coordination – politics, regulation, large organizational change. The third is what we call meaning-makers – the people who define what is worth doing and what is worth caring about, in domains where value is anchored in shared human attention rather than in anything an algorithm can score.

These are the fields where measurability bends rather than breaks. They are also where the next generation of high-value work will increasingly live.

You believe that human intuition and our ability to navigate “unknown unknowns” is provisional. Sooner or later, spatial intelligence and large world models will synthesize this final redoubt for human cognition. What do you think the timeframe for this is?

Less than a decade for the machine side of that question. The capability curves on spatial intelligence and large world models don’t give me a strong reason to bet against the timeline, and the track record of betting against AI capability progress in any specific domain over the past few years has been pretty bad.

But the framing slightly understates what is actually happening. The interesting development is not just on the machine side. It is also on the human side. Before models fully close the gap from above, we will start closing it from below – through synthetic practice environments, AI-native verification stacks, and eventually direct neural augmentation. The boundary between “machine intelligence” and “augmented human intelligence” will get blurry well inside that decade. By the time models are mature enough to plausibly replicate the kind of cognition we currently treat as uniquely human, we will be a different kind of cognitive entity ourselves.

So I would not describe human cognition as a “final redoubt” that gets synthesized away. I would describe it as the part of ourselves we will most want to augment. The real question is whether we do that quickly and thoughtfully enough to remain peers with what we have built – not whether the machines arrive first.

Once we reach this point, you claim, “by exhaustively simulating counterfactual realities, agents will increasingly internalize the capacity to deduce optimal policies in entirely novel environments.”

This brought up multiple questions for me that I am interested in your response to. Granted that agents will graduate to this level of thinking, the prospect of applying machines to policy formulation seems like a dead end to me for two reasons.

Let’s start with the fact that politics is inherently conflictual – not a puzzle with optimal solutions but a power struggle. Is the implication that powerful agents will somehow convince us that this isn’t so, or are you imagining a technocratic ‘world brain’/planned economy scenario with ‘optimal solutions’ imposed from above?

Neither. The right framing is not “machines deciding policy.” It is “machines making it possible to decide policy with better information.” For complex societal coordination – which is most of what policy worth talking about actually is – the realistic scenario is that these systems help inform better decisions while humans stay firmly in the loop, weighing trade-offs, brokering interests, and absorbing accountability for the outcome. Politics is a power struggle, not a puzzle, and there is no “optimal solution” the machine is going to compute and impose. There is just a slightly more legible map.

The harder issue is steering. The risk we are trying to manage in the paper is the widening Measurability Gap – the distance between what these systems can do and our ability to verify that what they do is aligned with human preferences. That alignment problem is the same problem at the policy layer as at the agent layer, just with higher stakes. Without serious effort, we end up with systems that look like they are serving us and are actually optimizing for something else. It is also already obvious that these systems will be used to shape opinion and influence – exactly the way social media algorithms have been for the past decade, only more capable, more targeted, and harder to detect. So the future of AI in politics is not technocratic central planning. It is hybrid, contested, and adversarial – informed decision support on one side, persuasion machinery on the other, and democratic institutions trying to do what democratic institutions have always done, which is keep up.

My second question is this: aren’t we engaging in utopian thinking when we expect policymakers to defer to machine judgement when they already wilfully ignore the suggestions and warnings of scientists and other experts?

The optimistic case – and I am aware it might be too optimistic – is that AI makes it harder for policymakers to ignore evidence, because it lowers the friction of surfacing ground truth and getting it into circulation. A society that can extract, verify, and disseminate the actual state of affairs faster, in ways harder to suppress and easier to audit, may be marginally better at holding power accountable than the one we have now.

Whether that translates into actual deference is another question. Realistically, it will be the usual adversarial cat-and-mouse game between capabilities. The same tools that surface ground truth will be used to manufacture plausible counterevidence. The same models that make scientific synthesis cheaper will make synthetic confusion cheaper too. So I would not predict a regime change. I would predict a faster cycle, with the side that invests harder in verification infrastructure ending up with more leverage in any given moment.

You provide an operation playbook for individuals, companies, investors, and policymakers in the AGI economy. Your advice to the missing juniors with no access to the traditional career ladder seems especially crucial since, in a ‘lost generation’ scenario, we will eventually run out of skilled human verifiers. You recommend they start acquiring intuition via synthetic means. What does this mean in practice, and how should juniors who are feeling the squeeze today start doing this?

The practical answer is more direct than the framework might suggest. Use every AI tool you can get your hands on. Follow what you are actually interested in. Get on the frontier – you now have the ability to learn about basically anything, on your own schedule, at a pace that would have been unimaginable a generation ago. Then try to build and execute. Iterate. Pick a thing you care about, ship a version of it, find out what is wrong, ship another one. Get the reps under your belt.

The point that needs to land for juniors is this: you can do many times more than your seniors could at your age, with a fraction of the support they had. The traditional career ladder is being pulled up, but the leverage on the new ladder is genuinely much higher. It is a superpower if you embrace it. The juniors who are going to do well are the ones who treat AI not as a threat to their entry-level role, but as the tool that lets them skip three rungs ahead of where the career path used to start.

There is no shortcut to experience, but there is a much faster path to it than there used to be – if you are willing to actually use the tools to build, fail, adjust, and try again.

You also highlight some suggestions for experts to augment their work and escape the Codifier’s Curse. Experts must become “directors of agentic swarms” and they can also mirror the junior strategy of augmentation, you write. But looking further out to the emergence of LWMs, their only defence may be a “paradoxical gamble” where they use the models trained to replace them to generate “synthetic entropy”. What does this look like in practice?

For experts as for everyone else, the ultimate answer is to identify the new jobs and roles the technology enables, rather than to defend the old ones. Most of the pain in any labor-market transition is in the transition itself – between when the old roles erode and when the new ones become legible enough to migrate into. The synthetic-entropy gamble is, in that sense, a bridging move: a way for senior experts to keep their judgment ahead of the models long enough to make the migration on their own terms.

A point worth making here. Even when we reach AGI in the formal sense, we will not become redundant by design. We will simply be a different kind of intelligence – embodied, slower in some ways, faster in others, with our own tradeoffs and our own structural relationship to meaning, attention, and value. That asymmetry is what enables meaningful gains from trade between humans and machines, in roughly the same way trade has always worked between parties with different capabilities and constraints.

In practice, I expect the simulated environments we build for synthetic practice will become much more than training infrastructure. They will become the mediumthrough which we explore our own creativity, prototype new artifacts, and build experiences for others to engage with. The line between practice, work, and creative expression will get blurrier than people currently expect.

The harder, and more uncertain, prediction is what happens to society when this is unevenly distributed. There is a real possibility that we split into two cultures: those with the agency, urgency, and motivation to keep building – using the tools to extend themselves, to make new things, to push frontiers – and those happy to consume what is now broadly available without much effort.

Both are legitimate ways to live. But they will produce very different relationships to economic value, and very different futures. That divergence, more than the technology itself, may turn out to be the defining cultural feature of this period.

There is a broad risk of market failure in the agentic economy – what you call the “Trojan Horse Externality” – where the accumulation of unverified risk and counterfeit utility consumes resources without serving human intent. Companies and individuals capture some gains but the losses, which are things like security debt and erosion of institutional trust, are socialised.

To counter this prospect, you say that governments must “shape the direction of technical change” to where agents are verifiable, liabilities are clear, and economic activity is robust against the systemic fragility that will plague low-verification jurisdictions.

Where would be the best way for policymakers to start working to avoid the meltdown of a hollow economy?

Things are moving extremely fast, and the honest answer is that policymakers should resist the temptation to design comprehensive frameworks for a target that will not stand still long enough to be regulated. The highest-leverage moves are a small number of structural ones, and they look more like enabling infrastructure than like rule-writing.

Four priorities, roughly in order. The first is education and retraining at scale. The labor market is going to shift faster than any prior wave of automation, and the social cost of getting this wrong falls hardest on people who never had the option to opt in. Investing seriously in retraining infrastructure – including AI-native retraining infrastructure, since that is what works at the speed of the disruption – is the single most consequential public-sector move available. The second is supporting open source models, because in a world where capability concentration is the default outcome, open source is the structural counterweight that keeps the field competitive, auditable, and accessible to research that proprietary labs will not run. The third is public ground truth infrastructure, i.e. the foundational data layer that makes risk legible enough for insurance markets and downstream regulation to actually function. The fourth is creating an environment that is genuinely favorable to new startup formation and business experimentation: tax, immigration, regulatory predictability, capital access. The structural answer to “how do we capture the upside of this transition” is to make sure the people most willing to take risk and build are not blocked from doing so.

Democracies must also work together, you say, by enacting legislation to curtail “uncertified agentic services” and leveraging their collective purchasing power to force better standards and practices that make safety a prerequisite for profit.

Aren’t you, at least to an extent, describing the thrust of EU tech policy – much more cautious and user-centric than the US – that major AI builders (and the US government itself) are in strong opposition to? If so, what’s the feasibility of uniting around such a common project?

I would actually argue the opposite. The EU’s tendency to regulate new technology early and comprehensively has been one of the most consistent forces stifling the enormous talent base it has across academia and the private sector. Europe produces world-class researchers, engineers, and entrepreneurs, and then routinely makes it harder for them to build at scale than for their counterparts in the US, China, or elsewhere. AI is just the most recent and most visible example.

The deeper problem is that you cannot regulate this technology properly in the way the EU’s instinct prefers. It is too much of a moving target. Capabilities that did not exist eighteen months ago are now table stakes, and the rate of change is accelerating, not slowing. Any regulatory framework specific enough to be useful will be obsolete by the time it is implemented, and any framework general enough to be durable will not actually constrain the failure modes anyone is worried about. That is not a flaw in any particular regulator. It is a structural property of trying to govern a technology that is changing faster than legislative cycles.

The most important policy dimension, I think, is creating the conditions for the private sector to take on verification as the central bottleneck and the central commercial opportunity. Verification is the part of the agentic economy that is least solved, most economically valuable, and best suited to entrepreneurial experimentation. If governments – on both sides of the Atlantic – focus on making verification easier to build, easier to fund, and easier to monetize, they will accomplish far more than any safety mandate written in advance.

---