AI Might Not Be Taking Your Job Anytime Soon

But It’s Making It a Whole Lot Harder to Find a New One

Generative artificial intelligence is introducing a new wave of economic challenges. In the long run, there are legitimate concerns about large-scale automation, model misalignment, and even existential risk. Yet these dramatic scenarios are difficult to evaluate given the limited evidence so far. More immediate issues, however, are already visible. Many observers worry that the rapid spread of AI threatens entry-level employment, and the rising rate of recent graduate unemployment lends credibility to this view. This article explores some of the near-term consequences of AI adoption, focusing on how shifts in hiring and productivity interact with the business cycle. In particular, I examine how information signalling failures, and search-and-match inefficiencies may explain why finding a job feels unusually difficult in the age of AI.

AI and Productivity

First and foremost, the question on everyone’s mind: is AI a bubble? I am not a hardened AI sceptic. Recent advances are far from meaningless, further, I believe AI has genuine potential to transform productivity. The issue lies in where that transformation is happening. The problem with the development trajectory of AI is that too much of the consumer-facing output is dull generative slop or actively deleterious to the social fabric, while the economically transformative use case is just enhancing technical and administrative productivity on the back end.

What this means is that the visible faces of AI are valued in excess of their economic contribution, and in some cases even a net negative. I am referring to the kind of general intelligence tools and chatbots that can complete tasks like writing, image and video generation at around a median level, albeit with wild inconsistency and moderate human effort in prompting. The hype that these features outperform elite students and professionals is isolated to a few tests, and industry surveys show uptake has been fairly slow and not had much bearing on profits.

In contrast, specialised AI tools are an entirely different matter. Their ability to process vast amounts of unstructured data, assist with transcription, coding, editing, data cleaning, and workflow automation, the so-called “boring stuff,” is where the real productivity gains lie. Further to this point, in an academic context I’ve experienced first-hand how LLMs have the capacity to fundamentally transform paper revision and replication, labour-intensive tasks with immense (if overlooked) value. Yet most of the excitement seems centred around replacing high school level essays, formulaic structure and basic content.

Turning to the replacement of higher-skilled labour, it is worth recalling that fears of white-collar displacement are nothing new. The advent of computers once provoked similar anxieties about the future of administrative and clerical work. Many such roles have indeed been automated by successive waves of information technology, but LLMs may not represent a radical departure from these historical trends. Administrative burdens grow with economic complexity, and even as automation expands, demand for human workers persists due to complementarities between people and machines, more so than because of simple bureaucratic inertia or rent-seeking. What we are witnessing, I would argue, is not labour replacement but labour enhancement: technology amplifying human capability rather than rendering it obsolete. This is a marked distinction from claims of structural unemployment, productivity is increasing through labour-enhancing and not labour-replacing technologies.

The threat of technological unemployment resurfaces with every major wave of innovation. From artificial intelligence and computers back to the mechanised loom, each era has inspired fears of jobless growth. Yet, historically, the “compensation effects” of new technologies, through productivity gains, lower costs, and the opening of new sectors, have tended to offset the job losses caused by labour-saving or even labour-replacing tools. The real hardship typically arises not from aggregate job destruction, but from its uneven distribution: the concentration of losses in certain industries or regions, while the benefits are spread more broadly and diffusely across the economy.

This raises an important question: does the current wave of AI-driven innovation actually show signs of technological unemployment in the data? So far, the evidence suggests otherwise. Overall unemployment remains relatively low. Even as major firms announce layoffs, these appear to reflect broader economic stagnation rather than an AI-specific shock. Labour turnover is slowing, while GDP growth is increasingly concentrated among a small number of dominant companies, a sign less of mass displacement than of structural imbalance within the economy.

If I had to forecast the near-term economic impact of AI on productivity, I would argue that its labour-enhancing effects will be most pronounced among workers with high levels of human capital. The comparison to the dot-com bubble may feel overused, but it remains instructive. The firms that endured and ultimately prospered after the crash were those providing essential IT infrastructure, tools that complemented rather than replaced existing commercial activity. In many ways, the internet’s true economic impact resembled that of the fax machine: transformative, but largely through its role in enabling other forms of productivity rather than an independent source of growth.

But if this is the case, why are we seeing a drop off in entry-level hiring? Shouldn’t productivity enhancing technologies increase the demand for labour, and particularly skilled labour in industries that utilise information technologies? This is where I want to shift the scope away from the big picture to more applied micro. The economic effects of the aforementioned generative products seem fairly limited in terms of growth, but we need to recognise the frictions they produce in education and hiring, two areas where domains where information asymmetry and signalling play an outsized role.

AI and the Breakdown of Information Signalling

The rise of generative AI has not yet triggered a widespread collapse in employment via labour-replacing productivity gains, at least not that we can tell. Instead, the more immediate disruption lies in how AI has weakened signalling mechanisms in environments with asymmetric information.

In models representing markets and contracts characterised by hidden information, one party (the agent) possesses private knowledge about themselves that the other party (the principal) cannot directly observe. This asymmetry can produce inefficiencies such as adverse selection, in which uninformed parties make systematically poor decisions. Signalling provides a way for the informed party to credibly communicate their private information through observable actions.

In the context of the labour market, the classic model Spence (1973) frames education as a signal of ability, a concept later refined by Cho and Kreps (1987). Education serves as a costly signal: high-ability individuals can more easily bear the cost of acquiring education, while low-ability individuals find it prohibitively expensive to imitate. This difference in cost leads to different agents choosing different levels of education allowing employers to infer productivity based on education or other effort-intensive indicators (see Mas-Colell, Whinston, and Green, 1995, Ch. 13 for more on signalling).

Essentially, under certain conditions education can act as a signal of a worker’s ability: the higher ability agents find it less costly to obtain education, so acquiring a degree credibly signals higher productivity. The challenge in signalling models lies in ensuring the equilibrium is separating (different types choose different signals) and incentive compatible (each type prefers their signal to imitating another). These models are central in mechanism design and contract theory, and they rely heavily on Perfect Bayesian Equilibrium to characterise consistent beliefs and credible signalling strategies.

In education, AI significantly reduces the effort cost for low-performing students to imitate their higher-performing peers. An outward shift in their indifference curve reflects a lower cost of achieving the same or even greater observable outcomes. High-ability students, by contrast, experience a smaller reduction in relative effort, as they already operate near the efficient frontier of performance. The result is a pooling equilibrium, in which observable measures of achievement, grades, essays, or portfolios, no longer reliably distinguish between ability types. In effect, AI induces a convergence around the lowest-cost path to acceptable performance, eroding the informational content of traditional academic signals.

Figure 1 illustrates this dynamic through a hidden-information signalling model in the labour market, visually demonstrating how AI augmentation alters the equilibrium between wages and education. Under asymmetric information, high-ability individuals can credibly signal their type by investing in education, which is relatively less costly for them than for low-ability individuals. This cost differential sustains a separating equilibrium, where employers correctly infer worker productivity from education levels. The blue curve represents the net benefit for high-ability individuals, while the red curve depicts the steeper indifference curve for low-ability individuals, who face higher education costs and are thus deterred from imitation.

The two black lines represent the employer’s wage schedule, illustrating how wages are determined based on the employer’s beliefs about the worker’s productivity and representing the payoffs of hiring the respective agent-types. Specifically, these lines show the expected value of output conditional on a worker’s type, high or low ability, minus the wage paid. In equilibrium, employers offer wage levels that reflect the perceived productivity associated with different education levels, as seen in the tangency points between agent indifference and employer schedule..

When AI tools enter the picture, illustrated by the dashed lines, they flatten the effective cost curve for low-ability individuals. Generative technologies reduce the effort required to complete education-related tasks such as writing, coding, and exam preparation. This compression of costs narrows or even eliminates the cost gap between high- and low-ability types, undermining the incentive compatibility constraint that previously maintained separation. The wage of low type workers does increase along the low-type schedule, but it also allows imitation at the point where the indifference curve intersects with the high-type schedule. Consequently, both types may converge at the same wage-education level p, producing a pooling equilibrium in which education ceases to function as a reliable signal.

Figure 1: The diagram illustrates a signalling model with asymmetric information, showing how AI augmentation reduces the cost of education for low-ability workers, thereby shifting the labour market from a separating to a pooling equilibrium in the wage–education space.

This shift carries substantial implications for labour market efficiency, particularly in entry-level hiring, where education has traditionally served as the primary screening mechanism. Figure 1 illustrates how AI-driven reductions in signalling costs for low-ability individuals erode equilibrium separation, destabilising the informational role of education in the matching process. When employers can no longer infer ability from credentials, the labour market loses one of its most reliable filters, increasing both search costs and the likelihood of mismatched hires.

At an application level, generative AI tools such as ChatGPT, Copilot, and similar plug-ins have dramatically lowered the cost of producing high-quality submissions, application materials, essays, CVs, portfolios, and code samples. As a result, less-qualified or lower-quality candidates can now closely replicate the application signals of their higher-quality counterparts. This compression of signalling costs flattens the differential that sustained a separating equilibrium, driving convergence toward a pooling equilibrium.

Traditional hiring signals, writing quality, responsiveness, or standardised assessments, lose their discriminatory power. In economic terms, the incentive compatibility constraint that made signalling credible collapses, undermining employers’ ability to distinguish genuine productivity from AI-augmented imitation. In the context of the model, hiring the wrong type at a higher wage lowers firm payoffs, sometimes below not hiring at all. This uncertainty produces hesitancy, firms might put off new hires until they can be certain that they will be credibly hiring the correct type of applicant.

What would these dynamics look like in the data? Figure 2 shows the U.S. nonfarm hiring rate from 2005 to 2025. While the series displays the expected cyclical declines during major downturn, the more recent trend is striking. Since 2022, hiring has steadily fallen despite the absence of an official recession.

By 2025, the hiring rate has approached its lowest non-recessionary level in two decades. This pattern points to rising matching frictions in the labour market, not from a shortage of applicants, but potentially from a breakdown in the informational quality of hiring signals. As firms struggle to differentiate between candidates of varying ability or authenticity, many may choose to delay or reduce hiring altogether, reflecting growing uncertainty rather than declining demand for labour.

Figure 2: While the hiring rate exhibits the expected cyclical declines during recessions, 2008–2009 and again during the 2020 COVID-19 shock, what’s striking is the persistent decline in hiring since 2022.

The decline in hiring rates is consistent with the theoretical breakdown of signalling equilibria described earlier. When AI reduces the cost of imitation, observable credentials lose their power to separate high- and low-ability candidates. Employers, facing greater uncertainty about true productivity, respond rationally by tightening screening standards, lengthening recruitment processes, or postponing hiring altogether. In effect, the collapse of credible signals raises the transaction costs of matching, leading to a slowdown in hiring even when aggregate demand for labour remains strong. The data in Figure 2 thus capture not a cyclical downturn, but a structural friction in the market for new entrants, an informational drag on employment driven by the erosion of reliable indicators of ability.

Search and Match Frictions

Taking a step back from the pooling equilibria, what does this mean for the labour market more broadly? To understand recent developments in market inefficiency, it is useful to look the Beveridge curve (1944), which characterises the inverse relationship between job vacancies and unemployment. Traditionally, this curve reflects the efficiency of the matching process: how well the labour market pairs unemployed workers with available jobs. Movements along the curve reflect cyclical dynamics (higher unemployment during recessions, more vacancies during an expansion), while shifts of the curve itself, particularly outward, are interpreted as evidence of deteriorating matching efficiency.

This framework helps explain why unemployment can persist even when there are many job openings and is widely used to study labour market dynamics and policy impacts. Because it takes time and effort for firms to find suitable workers and for people to find jobs, the model includes frictions that prevent the labour market from clearing immediately. It uses a matching function to describe how efficiently job seekers and vacancies come together and often includes wage bargaining between firms and workers.

As discussed above, recent advances in application technology, particularly the widespread adoption of AI tools in recruitment and job applications, have introduced new and unanticipated frictions into the matching process. If you’ve applied for a job recently, you may have encountered disclaimers such as:

“This application will be assessed with the help of AI; all final hiring decisions will be made by a human.”

“Please agree to our AI terms of use or disclose any use of generative AI in your application materials.”

The increasing use of generative AI has dramatically lowered the cost of producing tailored resumes and cover letters, and on the other side of the table, employers are using AI to sift through the growing stacks of applications. Even lower quality applicants can now generate high-quality application materials at scale, flooding firms with large volumes of superficially well-matched candidates. This contributes to signal degradation: traditional markers of suitability, bespoke cover letters and polished CVs, no longer reliably distinguish motivated or qualified applicants from those who have simply optimised prompts.

It’s been remarked on before that the state of the job market is AI recruiters posting AI job listings and sorting through AI resumes and AI cover letters. While this is hyperbole, it does reflect a problem of volume. The median-level artificial intelligence with an eye for detail and regurgitating the language of the job listing is really all you need to produce a optimised resume and cover letter for a job application, drastically reducing the marginal cost of applying. That is to say nothing of agentic AI, eliminating the cost of applying altogether.

Consequently, job listings now often receive hundreds or even thousands of applications, many of which are nearly indistinguishable on the surface. Further, If you’ve applied for a job recently, you have almost certainly encountered:

“We’ve received an unusually high volume of applications for this role.”

While this may not be that new or uncommon of a phrase, a nice way of letting down applicants, it can take months for some firms to sift through the pile of applications. In the case where a company gives a rough number of the applications to positions ratio, they often exceed 1000:1, with the common wisdom that a recent grad will need to submit at least 100-200 applications to secure a position.

In the context of AI-enabled job search and recruitment, the Beveridge curve offers a natural lens through which to interpret the growing disconnect between high vacancy rates and slowly rising unemployment levels. The widespread adoption of generative AI has drastically lowered the cost of job applications, leading to a surge in application volume. On the firm side, AI-based applicant tracking systems now filter resumes and cover letters at scale, often relying on keyword heuristics and scoring algorithms. While these tools aim to improve efficiency, their overuse may degrade the signal-to-noise ratio in applications, particularly when applicants themselves use AI to generate optimised but formulaic materials.

This dynamic gives rise to a novel form of search-and-match friction. There is a circular dynamic where the filters are trained on AI applications, and the applications are updated to optimise for the filters. From the perspective of the Mortensen–Pissarides (1994) framework, it represents a decrease in the efficiency parameter of the matching function, the term that captures how well the unemployed and vacancies are matched. When job applications become indistinguishable due to AI homogenisation, and when firms are overwhelmed with inflated applicant pools, matching efficiency declines, even as vacancies remain relatively high.

In graphical terms, this results in an outward shift of the Beveridge curve: for a given level of vacancies, unemployment is higher than before. It reflects the fact that AI-induced pooling and filtering breakdowns distort the matching technology. This friction is not structural in the classical sense (e.g. skill mismatch), nor purely cyclical, but a function of informational noise and breakdown in the signalling mechanism, a kind of generative AI-driven market inefficiency.

Figure 3: illustrates an outward shift in the curve, indicating higher unemployment at every level of job vacancies. The shift reflects growing search-and-match frictions caused by AI: homogenised applications, weakened signals, and overloaded filtering systems. Together, these frictions reduce matching efficiency even as total vacancies remain high.

The matching function in the search-and-match framework, typically increasing in both unemployment and vacancies breaks down when the quality of matching signals is diluted. Firms struggle to extract relevant information from a saturated pool of indistinct applications, leading to delays in hiring or missed matches altogether.

Taking this theory to the data, the Beveridge curve for the general population shifted outward during the pandemic but has since corrected and returned to its 2010-2019 position. This has not been the case for recent graduates, as seen in Figure 4. In aggregate, for college graduates age 25-34, this appears as a persistent rightward shift of the Beveridge curve: higher vacancies for a given unemployment rate due to lower matching efficiency.

Figure 4: The relationship between job vacancies and unemployment among recent college graduates has shifted outward indicating a decline in matching efficiency, amidst higher job vacancies. The change suggests increased search-and-match frictions for young jobseekers, consistent with the breakdown of traditional signalling and screening mechanisms in the AI-augmented job market.

In this context, what does it take to actually get a job? Common wisdom suggests that applying through online portals increasingly resembles a roulette wheel, low odds, little feedback, and a high noise-to-signal ratio. Employers still rely on screening tests, but many of these tools like writing samples, logic games, even technical coding tasks, are now easily gamed with AI assistance. And if AI is becoming a legitimate productivity tool in the workplace, it’s not obvious why applicants should be penalised for using it in the hiring process.

For job seekers, this erosion of traditional, effort-based signals means that the labour market is reverting to a reputation and network-based hiring system, where referrals and trust substitute for noisy signals of observable effort. While this kind of informal screening may help employers reduce uncertainty, it also reduces the efficiency of job matching: capable applicants without network access may be overlooked, while noise from low-cost, AI-augmented applications clogs formal channels. In effect, the levelling effects of standardised, merit-based applications are being undone, leading to a dual-track labour market, one informal and relationship-driven, the other formal but increasingly unworkable. This breakdown in signalling, not AI-driven productivity per se, may be the more immediate driver behind the stagnation in hiring rates visible in the data.

The effects are particularly acute for young workers and recent graduates, whose limited professional experience leaves them most reliant on the very signals now losing value. For them, the labour market feels not only tighter, but less legible: the pathways that once led from education to employment have become uncertain, mediated by algorithms, and distorted by the same technologies that were supposed to expand opportunity.

In Conclusion

The anxiety surrounding artificial intelligence is not misplaced, but it may be misdirected. The true challenge AI poses to workers today is not mass structural unemployment, the breakdown of the informational systems that used to make the labour market legible. Jobs still exist, but firms are hesitant to hire, even as productivity continues to grow. What has changed is how individuals are seen, evaluated, and selected in that process.

In theoretical terms, the signalling collapse and search-and-match frictions outlined in this article represent a subtle but significant source of economic inefficiency. When employers can no longer distinguish between applicants, the cost of information rises. The result is hesitation, delay, and underemployment, especially among young workers who lack social capital. Ironically, a technology designed to improve efficiency has introduced new layers of friction by overwhelming the very systems it was meant to optimise.

Still, this outcome is not inevitable. Just as earlier technological revolutions prompted new institutions, credentialing systems, standardised testing, and professional licensing, the current wave of AI disruption will likely drive innovation in signalling and verification. Even as AI detection remains inconsistent at best, new digital credentials, task-based assessments, and rigorous, if transparent, AI disclosure norms may eventually restore some informational balance. But for now, the transition is messy, and its burden falls heaviest on those at the start of their careers.