Permanent Underclass: a socio-economic group trapped at the bottom of society with virtually no path to upward mobility, separated from the mainstream economy by deep structural, historical, or technological barriers.
The Permanent Underclass meme has been circulating around AI (and to a similar extent, crypto) discourse for at least a year now. The crux of its origins is that in the midst of an economic cycle where a few winners provide exorbitantly above market returns because of their anticipated future impacts on society, the owners of these assets will generate extreme wealth while even some of the highest-paid workers struggle to convert earnings into ownership quickly enough to catch up.
That’s what makes the framing of the ‘underclass’ so jarring in this context. A second-year Investment Banking analyst in New York earning $150,000 base may have been the highest achiever in all of their classes growing up, ticked off all of the extracurricular boxes and seemingly looks on track for a bright future both financially and in the profession. Even before any displacing effects, this individual is staring down the barrel of rent, living and student loan expenses that make retaining any savings to reinvest in the market a test of discipline.
Compound this effect with the possibility that models will gain the trust of the bank to perform such tasks as pulling company statements, spreading financials and drafting first-pass pitch materials. At this point, the analyst is surplus to requirements and their route into the asset-owning class is inhibited by the likely loss of the one thing underwriting their path – their wage.
So what then happens to this class of what was previously overachievers?
Many solutions have been proffered to this problem, all previously acknowledged ideas but this time with their own AI flavour: Universal Basic Income, mass re-training in non-knowledge-work trades and curriculum overhaul lie among them. Universal Basic Compute has even been offered up as a means to redistribute intelligence to the populace.
However, the above would all require either substantial outlays at the expense of the taxpayer (of whom many may still feel their employment is at threat) or lead times that aren’t close to matching the pace of development in AI.
If you are of the belief that AI will ultimately drive substantial productivity gains and, as a result, increase GDP (whilst not simultaneously holding extreme doom scenarios), these displacement scenarios represent short-to-medium term dislocations that will be remedied by the growth in the wider pie, creating new opportunities for human hands to fill.
In this case, market-based solutions for handling this dislocation offer quicker relief for those affected and cheaper bills to the taxpayer. With the rising acknowledgement of the risks to the knowledge working class, this essay serves to examine the supernormal profits that can be made by those that structure and originate insurance products to help allay the potential threat to professional service career paths as a result of AI displacement.
Reintroducing Livelihood Insurance
One such combination of ideas, as introduced by Robert Shiller in 1993 and 2003 respectively, is the confluence of Macro Markets and Livelihood Insurance.
Livelihood Insurance refers to a kind of contract that pays out when the earning power of an occupation or field declines, protecting most individuals’ largest asset, their future income. I will also explore in this essay the possibility of having more binary livelihood insurance options that allow individuals to manage the risk of being made redundant as a result of technological dislocation.
Rather than insuring an individual’s ability to work (the role of General Income Protection insurance) or the security of their current job (government unemployment schemes), Livelihood Insurance instead aims to solve the gap in protecting an individual against the long-term earning power of their profession.
Macro Markets are the bedrock on which this livelihood insurance would exist. These markets are theoretically expressed as liquid, tradable securities whose values track large economic aggregates such as national income, GDP, occupational wages, so that widespread societal risks can be priced, hedged and spread among investors.
This essay will focus solely on the potential market-based application of these concepts by entrepreneurs or incumbent insurance firms to combat the rising threat of widespread unemployment or loss of income as a result of increasing deployment of AI and robotics.
That being said, there are plenty of interesting considerations that were first introduced at the time of these ideas’ publication regarding their application in combatting any of a wide range of circumstances that negatively impact an individual’s livelihood and are out of their control. These include home equity depreciation, sovereign risk and loss of income as a result of career risk taking (e.g. pursuing further education, novel fields or starting a startup).
These ideas were first introduced over 30 years ago, and developments relating to both the market need for these instruments and the technology necessary for their implementation have improved its candidacy as an idea whose time has come.
An Idea Whose Time Has Come
In 2025, Anthropic CEO Dario Amodei issued a warning that artificial intelligence could eliminate up to 50% of entry-level white-collar jobs within one to five years. It was estimated that the carry-on effects of such displacement would thus drive overall unemployment into the teens for the knowledge work sector.
You could be forgiven for thinking that the firm (and its peers) may be overestimating the efficacy of their product merely as a marketing ploy … particularly in the lead-in to an IPO where such claims may serve to jack up prices.
But this piece was not inspired by the investor relations theatrics at play, but instead the very real possibility that these products are functional and that even if they do not completely wipe out junior employment across knowledge work, they have reached a level of efficacy where prompting one of their models produces better returns on investment than hiring or retaining employees in training.
In the case that foundation models continue to improve on functional tasks at a similar rate to what they have done since their public releases, it seems feasible that this shock to the labour market will expand beyond graduates and juniors to mid-level employees as well. If this case is realised, even in the event of Amodei’s 50% figure not being realised it would still materialise as a significant shock to the broader labour market as a whole.
Even observing the exhibit above, it is clear that it is not just knowledge workers who feel the anxiety. The growing promise of new technologies such as autonomous vehicles, humanoid robots and vocation-specific applications and models have triggered heightened perceived threats to non-knowledge-work occupations than previously experienced.
As a result, career anxiety has gone mainstream.
The good news? Whilst AI has unleashed a batch of fears among university students and experienced professionals alike, simultaneous developments in information technology, data availability and markets may provide the basis for its redress.
Why Now?
When the Macro Markets and Livelihood Insurance ideas were first proposed by Shiller, much of their basis was presciently predicated on the information age offering far improved labour market data, a prediction that has materialised.
The availability of far richer market data today than existed at the time of publication of these concepts provides a measurement layer for introducing livelihood insurance securities. In the 90’s, the labour market was measured primarily through a collection of monthly surveys, often published with a lag.
Nowadays, the migration of much of the hiring process to firms such as Indeed or Glassdoor offers real-time data on wages in very specific subsets of the economy, with very deep information on the knowledge work industry. This is complemented by independent surveyors such as Revelio Labs and ADP now covering over one-fifth of nonfarm private employment.
Relationships with these data providers offer a basis for occupation and paygrade-level income statistics that can be used as baseline metrics against which policyholders could feasibly be insured.
To simplify how much easier independent firms might make this task, consider that we were introducing a livelihood insurance for Investment Analysts (1-3 years experience) in New York City. In order to get a basic gauge on the changes in the market, Indeed could provide time-series data on the number of postings for such a role from verified firms and the aggregate salaries included therein to provide a basic proxy for the livelihood threat to current Investment Analysts.
Secondly, where many older academic papers cite prediction markets and event contracts as means for insuring against macroeconomic risk, the liquidity and sophistication of prediction markets was not at a stage to support broadspread risk transfer. This then provides a basis for a more sophisticated execution layer than was available for livelihood insurance markets than twenty years ago.
We are now at a stage where the efforts of the market have gotten us to the point where these are commercially normalised and increasingly liquid. Event contracts on these platforms could settle on real-time indices of early-career employment in certain fields provide a market-underwritten probability of displacement. These same event contracts could then theoretically serve as the venue where the insurer lays off their risk.
Outside of the prediction markets duopoly of Kalshi and Polymarket (both of whom have overwhelmingly seen much of their volumes concentrated in short-term-settled sports markets), the introduction of Hyperliquid’s HIP-4 provides another potential venue for bounded outcome markets. Under the Hyperliquid design, broader dislocation markets could potentially be introduced with a contract settling on probability of early-or-mid career displacement across a range of industries. There is a liquidity challenge here in the short-run, but it offers promise and exhibits the way in which market and technological innovations have contributed to the relative ease of introducing new, experimental financial instruments.
Further to the above two points, capital is now better placed to carry this risk than it was previously. The catastrophe bond market today stands at around $65bn in outstanding capital, with record issuance in 2026 already. Whilst catastrophe bonds are typically used by insurers to transfer their risks linked to acts-of-god (earthquakes, hurricanes, fires) to investors, the risk transfer model fundamentally doesn’t operate dissimilarly to potential concentrated dislocations in certain segments of the employment market.
Why did it stall, and what has changed?
Outside of the technological advances that have occurred in the intervening period between the ideas’ introduction and today, there are some other barriers that may have contributed to their actioning at scale.
The first issue is demand – whilst individuals will on occasion voluntarily purchase insurance, much of the time it is a big step to expect individuals to purchase novel financial products that they may not understand. Add to this the high probability that outside of a certain discrete outcome, they will bleed an annual premium and it becomes a tough sell. The buyer has little means of judging whether the price is fair, because the probabilities of the career risk that they would be insuring are still widely debated by experts with limited historical precedent for the specific risk (in this case AI-related displacement).
The second issue is the vast amounts of capital that are required to support a novel insurance initiative. Because livelihood insurance in its initially proposed form is designed to support correlated shocks among entire occupations, the insurer lacks the possibility of pooling risks across the policyholder base.
This is a large argument for insuring certain individuals rather than the group as a whole under a new model, and one that would be justified by the outcome of AI-related unemployment leading to a bifurcation of high-leverage junior employees who are good with tools shooting up the ladder and those left behind being those that are dislocated and filing livelihood insurance claims. Time will tell but based on the priorities of the firm and certain reported trends in financial firms this seems like a decently likely outcome.
Related to the above is the significant anticipated cost of being the first-mover in a market such as this. Early sellers bear the costs of educating the market and likely selling policies at a net loss in the short-run in order to establish themselves as a trusted player in the market.
Most importantly, and again related to the capital quandary, is that given the substantial amounts of funding that these occupation-wide policies would require very robust, well-defined policy designs will need to be established before rolling out at scale in the market.
For example, cover paying out based on an individual’s circumstance invites moral hazard where employees could be incentivised to underperform, whilst cover that pays occupation-wide based on an index may leave workers unprotected compared to their needs when their own income falls.
However, returning to some of the above points, we are now at a point in the market where some of these challenges may be more tractable within the specific window of AI-related employment threats.
Government awareness (and growing expenditure related to) these issues gives the public sector more incentive to sponsor these schemes through a variety of potential machinations including, but not limited to: dollar matching insurance contributions, sponsoring some of the insurer’s spread to ensure individuals are covered, or acting as a market maker for these policy transactions.
Institutional investors have exhibited far more appetite than two decades ago for trading index-linked risk, whilst the rest of the private sector (i.e. employers) provide a natural route for group enrolment into these products as it can either complement or replace parts of defined contribution or defined benefit plans (DBPs / DCPs are helpful in the case that the employee continues to earn, but fall off afterwards – by sponsoring livelihood insurance for employees, employers can demonstrate goodwill in supporting them post-employment if the actions of the firm trigger this change in circumstances).
One example design for Livelihood Insurance to address AI risk
To ground all of the above in an investment case, I will propose one design and evaluate their feasibility and potential returns.
Keep in mind that this is a very narrow window into the broad variety of implementations or designs that could be applied to the AI labour market problem by itself, let alone the broader applications of livelihood insurance against other risks.
The selected design is structured to insure against the case where AI makes firms more productive and thus creates a bifurcation among early-career employees where a small few may see outsized rewards whilst a majority or at least a significant minority are left without work.
Where Shiller’s work cites index-based payouts as an option for insuring against employees losing income as a result of declines in an industry-wide earnings index, for this particular circumstance we deem the selected alternative more adequate for providing backstop earnings in the event of unemployment.
This is then designed as a model where:
> The proposed policy will solely cover Investment Bankers in New York City with 1 – 3 years of experience, with a conversion option upon promotion to a new package tailored for an index of 3 – 5 year investment bankers and so on;
> The policy will be wholly owned by the employee, but the employer maintains the option to sponsor these policies and introduce them as part of perks packages etc.;
> The payout trigger will be a verifiable involuntary termination as a result of the firm’s shifting priorities – this may require validation from the firm that it was not for underperformance or other reasons;
> The reference salary that will be used around which to centre the distribution of claims for this experience range will be $150,000.
> Payouts will thus be broken down into two stages:
>>>>>>> (1) A 1st bridge payment of $15,000 lump sum (10% of index average) if the employee is still without work 30 days post-termination;
>>>>>>>(2) A 2nd track exit payment of $20,000 at t + 12 months if the analyst is not back in a qualifying banking role.
This leaves the maximum payout available to an analyst at $35,000 – again, this is a very significant design choice that can be played around with in alternative models.
This has been chosen in this instance to reflect the gap between the $150,000 benchmark salary and the income that may be available to an analyst if they were to pursue some alternative career that may be more resistant to AI-related displacement in the near-to-medium term.
An alternative that could have been posited in this piece would be to place the maximum insurable amount closer to ~90% of this salary would mean the analyst (without sponsorship) has to pay a premium of $23,260 p.a. (15.5% of total salary) when calculating based on a conservative insurance load of 1.25 (i.e. insurer expects to pay $1.25 for every $1 of claims that arise from legitimate displacement, due to leaks arising from soft claims, moral hazard or adverse selection issues).
For the purposes of modelling a startup structurer/originator model, it is deemed unlikely that the analyst and/or sponsor would find paying this percentage of annual income in premiums economically reasonable.
Other moral hazard controls that can be introduced to the dual-staged approach above may include:
>> Employer-certified reason codes for filing and processing claims;
>> Payroll-verified re-entry into policies;
>> Single claim per policy limit;
In order to hedge the insurer’s exposure, 90% of quotas will be shared with reinsurers, with the potential to introduce a displacement-rate index (share of starting cohort still employed in field) as a second line of defence.
SAMPLE MODEL HERE FOR YOUR OWN EXPLORATION OF SCENARIOS.
To run some napkin math on what this would look like for the policyholder at the per policy scale, observe some of these rough calculations below based on 4 scenarios for the AI displacement risk: overstated (+5% Δ in analysts displaced), mild (+10% Δ), moderate (+25% Δ), and severe (+45% Δ).
And for the structurers (service businesses responsible for designing, selling and managing claims) and originators (holder of financial exposure) of these policies:
This then provides the following illustrative outcomes for the structurer:
Based on the assumptions used above, returns trickle through stakeholders as follows:
Even a cursory review of the above sample data will show that much of the wedge in the issuance of early livelihood insurance products under the proposed model would settle with the structurers of these instruments.
A structurer ceding 90% of claims to alternative financial sponsors or investors earns $13mm whilst remaining solvent even if claims exceed double the expected rate under the napkin math example above.
In such severe cases, the solvency of the structurer remains dependent on investors paying. In this worst case, claims reach $300.9mm, requiring investor payouts of $185.2mm as a shortfall against premiums. This requires the cession of risk to be fully collateralised, in a manner similar to catastrophe bonds. The reason for this cession being entirely necessary in this case is because of the correlation of the risks in the proposed design, in that claims are highly likely to arrive together.
One interesting point to observe from these financial examples is the role and expected outcomes of the investor. Purchasing the risk being ceded by the structurers of these instruments is an expression of the investors’ conviction that the AI community may be overselling the risks associated with their products – in the event that the doom / moderate-doom scenarios are oversold, these investors would stand to gain an aggregate $65.7mm across a localised 20,000 policy book (though there is also significant downside if AI wipes jobs across the board to the tune of $185.2mm per the model cited above).
Keep in mind as well that all of the data above does not account for many of the exciting possibilities of third-party sponsorship or support that would smooth the go-to-market process for livelihood insurance or similar products. Given the civilisation-scale risk that mass displacement would cause, it does not seem unfeasible that governments might move to provide subsidies (e.g. explicit consumer discounts on premiums) on these products, absorb losses above a certain threshold or act as a market maker on these policies, all of which would serve to change the economics of the deal for consumers, structurers, originators and investors.
Secondarily, private firms could be incentivised to offer cover as a payroll-deducted benefit or explicit perk for employers, again improving accessibility and justifiability of the price tag from an individual consumer’s perspective.
In sum, these products serve significant value not just for the insured (who now have a backstop in the event of their career pathways being closed off or at least temporarily derailed), but also for the universe of market participants who want to express a view on the risk probabilities of AI tools replacing the value of human labour against what the labs are telling us.
Livelihood Insurance, in this case, charges both sides of the AI displacement argument a premium for the privilege of having this argument.
Then again, there remain plenty of question marks over whether such markets are viable – not just from a design perspective, but also from the threat of it being an expensive experiment for market participants to try and introduce in the presence of competing models or rough equivalents.
Governments already offer unemployment insurance for labour displacement in many of the major jurisdictions where high-income knowledge workers are concentrated. One of the limitations to these schemes is that the insurance cap is far below the benchmark wage index we have been working off above, and as such behaviourally may not make sense for the early career Investment Analyst.
The beauty of these schemes though, is that unlike with private insurers, it saves a lot of the hassle derived from charging premiums, operating claims analysis procedures or marketing to potential buyers (benefits are in many cases paid straight out of payroll taxes rather than premiums). The counterargument to state unemployment insurance being an adequate replacement is that across civilisation-wide risk events (such as AI-related displacement) it is extremely expensive and also does nothing for the threats that this poses to long-term career paths, instead acting as a short-term stabiliser.
As in any insurable market, the other major objection would have to do with moral hazard. The policy design proposed above relies a lot on operating procedures that may have to be built from the ground up or repurposed with twists in order to ensure claims are legitimate at scale (one such procedure being employer-certified involuntary termination related to a specific dismissal code). Imagine the case where an analyst is nearing the end of the eligible 2-year policy period – this timing aligns with approximately the point where many junior investment bankers begin to look for alternative work after a couple of years of tough toil.
By this point, with a livelihood insurance policy in place, they have every incentive to try and trigger their own dismissal and reap a payout. This risk may be difficult to replace in full, but can only be fenced and priced with a reasonable premium loading.
Closing Thoughts
In the event that firms begin to trust AI to handle a large portion of grunt work, the early career analyst introduced at the beginning finds themselves faced with a nearly intractable problem. Their route into the asset-owning class and the potential escape routes this provides them from the ‘permanent underclass’ is predicated on future career earnings that they cannot guarantee.
Hypergambling a portion of short-term wages or starting a high-risk startup are potential alternatives that some might espouse. However for a short-term, medium-probability risk like what this analyst faces, models such as the livelihood insurance idea proposed by Shiller and its cousins promise to provide adequate stopgap protections for adapting to an AI-led economy without necessitating significant expense to the taxpayer.
In the most hopeful scenario, such vehicles may offer such highly talented people the necessary bridge from the service-first economy to the unforeseeable career paths or alternative income-generating pathways in a world where the human mind is no longer the primary engine of innovation.
Livelihood insurance also may serve an important function outside of assuaging the employment anxiety issues of the knowledge working class it is designed to protect. Instruments that rely on finding a correct price for the risk of significant, easily-imagined AI scenarios occurring require a very correct analysis of the probability of these scenarios in order to price this risk – by providing these prices, we can move further away from the kind of anxiety-inducing doom scenarios and closer towards a better signal for how AI is impacting labour. If these sources do point towards a ‘permanent’ underclass being built at rapid pace, at least there is a means of compensation to get people back on the ladder.
Further Questions for Exploration
What kind of non-policy-style securities could be issued to help employers / employees hedge employees against the risk of their disemployment? (HIP-4, prediction markets)
What other risks to livelihood / identity could be effectively hedged against using event contracts?
What are the barriers to governments acting as a market maker on index-linked products such as these livelihood insurance markets?
How much of a risk is dominant sports gambling volumes to the viability of prediction markets platforms like Kalshi or Polymarket in becoming dominant players in event insurance markets?
Does the existence of markets for displacement change the probabilities / outcomes from displacement itself?
Would the endorsement of these products by AI labs / private employers hinder or improve upon the perception of these parties?
Key References and Further Reading
The New Financial Order: Risk in the 21st Century (Shiller, 2004)
Macro Markets and Financial Security (Athanasoulis, Shiller and van Wincoop)
Toward Livelihood Insurance (Abramowicz)