Quam angusta porta, et arcta via est, quae ducit ad vitam: et pauci sunt qui inveniunt eam!
Bibli Sacra—Matthaeus 7:14
The avalanche has already started. It is too late for the pebbles to vote.
Ambassador Kosh—Babylon 5
Not so fast, Ambassador. I have an idea.
TLDR: The real problem isn’t job loss, though that is a problem, the real problem is pathway collapse, because entry levels get hit the hardest, and, we might fix that by leveraging the benefits of an AI augmented workforce, if we just make a few small changes to education. Let’s not fight the market, let’s steer it, and change everything.
Introduction
By 1455 though it’s almost certain no one could read it, the writing was on the proverbial wall; what Europe’s still relatively few literate[1] persons could read was the Gutenberg Bible, the first major book to be produced by the printing press—and the crack in the dam that would lead to the democratization of knowledge, the Protestant reformation, and set human civilization onto a fundamentally different path—Gutenberg probably thought that his invention was mostly a more efficient way to produce and distribute printed materials. Life just moved more slowly back then, it’d be about another 400 years or so before widespread literacy really took hold.
Unlike the Europe of 1455, in 2026, Geoffrey Hinton, Yoshua Bengio, and the rest of us certainly can read the writing on the wall, where the world’s newest disruptive technology, the widespread adoption of artificial intelligence will almost certainly be as disruptive; the process has already begun. As these and others warn, our window to steer the avalanche we’ve unleashed might be exceptionally narrow, but at the very least most of us agree; we do have a narrow path out of the worst, most pressing, and already unfolding AI dystopia scenario, that of gradual disempowerment.[2] This work will be making the case that our best shot is to steer the avalanche.
Gradual Disempowerment & the Avalanche Already Underway
As of this writing, 28 March 2026, Layoffs.fyi is showing that 40,482 tech employees had been laid off since the start of 2026.[3] If the trend holds, the back of the napkin math, assuming that this constitutes roughly the first calendar quarter of 2026, we might surmise that we can expect perhaps 170,000 total tech layoffs for 2026, which actually compares favorably to say, 2023 which saw a whopping 264,320 employees lose their jobs—that’s about 100,000 less! But the aggregate numbers don’t tell the whole story, Stanford’s economics lab found that employment for young, entry level workers was being hardest hit.[4] What this points to is a broader trend in labour, the elimination of entry-level roles; “21% of companies have already frozen entry-level hiring due to AI.”[5]
This points to a very troubling trend; the elimination of the lowest-rung of the economic ladder, and the pathway for most people to achieve prosperity, is most vulnerable, as has always historically been the case (citation needed here).
This is the crux and thrust of gradual disempowerment; that these trends, incentivized by existing market forces slowly but surely erode human bargaining power in favor of AI automation and replacement. A world without work could become the promised utopia, (although, I think more likely a dystopia) but we ought to remember that utopia literally means “no place,” and no place is nowhere human workers want to end up because, there’s some evidence to suggest that total retirement is a predictor of mortality.[6]There’s a certain intuitive sense in which purposeless humans aren’t exactly happy and healthy humans. And while humans aren’t mice, if there’s a chance we could end up like Universe 25, we really ought to try. [7][8]
Permanent pathway collapse: the real threat
Since entry-level roles are always the greatest casualty of disruptive technology, we need to reframe the discussion by focusing on those roles in particular, and what entry roles provide. We’re trying to solve for job losses, and while job elimination and job losses are important, What we should be trying to solve is more specific: pathway collapse. By pathway collapse, I mean the process by which some persons, potentially a great many are locked out of prosperity by an inability to secure an entry-level position, because in most circumstances such positions, internships, apprenticeships or equivalents constitute the first, and lowest run on the career ladder. Your first job, gets you your second job, gets you your third job, and so on.
Entry‑level roles serve two functions: they deliver economic value and they gradually transform novices into experts. The pyramid replacement model shows how market incentives gradually transform individual firms away from using humans for anything but the most senior-level jobs.[9] This has the effect of concentrating power and money solely at the upper-levels, and this is the most likely path to gradual disempowerment for a simple reason; it’s the path of least-resistance.
This is not a deliberate assault on workers. It is the outcome of relentless cost‑optimisation. Models get better and cheaper; firms rationally adopt them to cut costs; the training ground quietly erodes. Because the harm accumulates gradually, it is easily missed. The ultimate outcome is one that looks more like the third estate needing to unite against the first two. And earlier citations clearly show; the avalanche has already started, but it might not be too late for the pebbles to vote. But they’d best do so quickly, loudly, and decisively.
But what if there a way to hold pathways open, and stave off gradual disempowerment, at least for a while? Even a short delay could prove decisive, and it is the purpose of this work to demonstrate how AI augmentation could become the narrow road out of the seeming inevitability of gradual disempowerment.
Misaligned Incentives
In 1943, Abraham Maslow proposed a Theory of Human Motivation, and thought he didn’t craft the diagram himself, the pyramid known as “Maslow’s Hierarchy of needs” is in 2026 a fairly ubiquitous lens for understanding human behavior.[10] The industrial revolution started by solving for the physiological, and generally followed the the structure of the pyramid upward. The issue is, the pyramid structure is intentional; it’s very difficult to say, pursue a romantic relationship, the third rung of Maslow’s pyramid if you don’t have shelter, which rests at the bottom rung of the pyramid. Eliminate the bottom rung, and the whole pyramid collapses, as people’s immediate physical needs for food, water and shelter will be re-prioritized away from say, scientific endeavor, or even literacy, and the trend there is alarming to say the least.[11] Because market incentives are such that profitability may favor automation and AI replacement, we might suppose that from a civilizational or national perspective, market incentives, along with those that favor a strong and successful middle-class have gradually been diverging, and so what the discourse is anticipating with artificial intelligence is the completion of a pattern that began with the printing press, but which may hold its final completion in AI replacement.
Critics (and I am one!) will however rightly point out that previous points of inflection and industrialization didn’t eliminate work, quite the contrary, what actually happened historically was a shift in labor distribution.[12] AI we’re told is different, and it might be, because with the advent of humanoid robots, eventual replacement of the workforce at scale is a real possibility, and one which is manifesting, at least on the margins https://fortune.com/2025/08/29/nvidia-ceo-jensen-huang-4-day-work-week-productivity/.
But this brings us back to market incentives—a tried and tested way to improve profitability and productivity is to embrace new technologies quickly, because market competition forces that trend; whoever makes the tech work fastest wins. And with AI, this ends in gradual disempowerment. However, creating or otherwise adjusting incentives isn’t a deterministic process, and here lies the wedge that may forestall pathway collapse. But first, we need to consider the alternative proposals, and why they’re inadequate to stopping pathway collapse.
Trying to fight the avalanche by lighting a match
Several proposed solutions are offered in the public discourse, and they’re all problematic because of a fundamental flaw; they’re trying to fight market incentives instead of working with existing incentives. This section will discuss these suggestions generally before discussing the limits of specific proposals.
Preserving Jobs Directly
If we acknowledge that we need to worry about job loss, and we do, while I’m focusing on pathway collapse, that shouldn’t exclude the vast projected displacement of white collar workers. Various measures are therefore proposed to ensure that people can stay at their jobs, despite any level of AI replacement. These proposals generally take the form of forcing a human to be in the loop. We don’t even need to consider numbers to know why this won’t work; this solution is
Uneven, and limited, even in professions where protection exists,
intentionally introduces inefficiency in a system designed to eliminate it,
will not scale at all,
very temporary at best.
The unevenness looks like this; the bar association is not about to allow ChatGPT Esq. to practice law; presenting legal arguments will remain the domain of lawyer. Law firms who make lawyers have every incentive to take on fewer, and fewer articling students and interns. Since a ChatGPT-like legal assistant can research and summarize complex legal faster and better, it’s more profitable to outsource that task, heretofore often subject of human labor. This is a direct illustration of pathway collapse in the first, and in the second, points out that the profession of law is already protected by a professional association.
This does nothing for professions which enjoy no such protection, such as, software developers, insurance adjusters, finance professionals and stock traders etc. Which leads directly to the second objection, the free market capitalist system seeks to reduce inefficiency, not introduce it, and therefore trying to force a human in the loop through regulation will be opposed by lobbying. And even if that succeeds, if the AI is much faster, and it is, then one human can supervise an increasingly larger team of AI agents. The AI can be trained to surface only those situations which require the humans attention, which leads directly to the fourth objection, this is a temporary at best solution, because of a fundamental flaw; trying to battle market pressure instead of working with it.
A Specific Proposal: Transition into Human Services
A Brookings Institute op-ed proposed what’s more or less a two-pronged approached, which can be boiled down to 1) creating an artificial floor for humans in the loop in human-services jobs like education, healthcare, and counselling, and 2) retraining or otherwise funneling displaced workers into these new vacancies to meet the artificial demand.[13] If we treat this proposal as a type, it fails for five reasons.
1) Imposing an artificial floors for human staffing is a top-down solution which injects friction into a system that profits by minimizing friction. In government-controlled industry like education, this may attract little in the way of pure market-pushback, save that politics will generally favor reducing taxes.
2) Friedler et. al. acknowledge that there’s a bottleneck, and friction in retraining displaced workers. Their solution is more national-level programs, which is top-down, imposed market forcing, rather than leveraging existing trends in the market.
3) It’s doubtful that the macroeconomics will align sufficiently; there are likely to be far more displaced software engineers than new vacancies created, which points back to the previous point.
4) This proposal requires heavy intervention in the market, a market which has long resisted and been stifled by this kind of heavy-handed meddling.
5) Finally, this admission on the part of the authors, “But other institutions will likely have to be built from scratch,” buttresses my previous two points, but additionally, it risks creating a vicious cycle of government liability and dependency. Universe 25: the “Human Utopia” awaits, but in the interim, the demand for government revenue to pay for these “make work” projects spirals upward, imposing a greater burden on the market, which incentivizes the market towards more automation.
Universal Basic Income (UBI)
A second popular suggestion is to push for a top-down, government basic minimal guaranteed standard of living, usually in the form of a species of universal basic income. [14] The proposal is more or less, that the government will tax and re-distribute to ensure that all citizens have access to a minimal standard of living, whether employed or not. Sometimes this is means-tested, sometimes it’s truly universal, meaning everyone who qualifies gets it automatically, employed or not. This is the most seductive option, because it sounds like it could work, until you think about the friction, who pays, and how much? And how does this actually distort the market pricing signals needed to make economic decisions?
It’s not clear that UBI won’t cause an inflationary spiral; after all, why wouldn’t landlords move to squeeze as much as possible out of their UBI-supported tenants, leaving them with little more than a subsistence existence? This would force a phony inflationary pressure on the regular payouts that the beneficiaries of UBI get, which could end up looking much like hyper-inflation. In this manner, estimating the rate needed for UBI, even with the best AI models ends up becoming a precarious exercise, and now we’re into the realm of hoping that hypothetical cause of the problem, widespread AI, needs to become ASI just to manage the whole thing.
But for the sake of argument, let’s presume that the economics of UBI are sound. UBI still has to contend with an error that slew more people in the 20th century than any other; the idea that you can reshape human nature itself. Even under the best-case scenario, UBI might strip individuals of meaningful activity, and one predictor of premature death is full-retirement. In other words, when AI is doing literally everything, what’s the incentive for humans to do anything at all? And the human creature which doesn’t have a gainfully employed purpose is a creature unhappy and this typically predicts morbidity. It is highly optimistic, and perhaps unrealistic to presume that if humans have all their needs and desires met by AI, that they will be happy finding things to do. This suggests that the government might engage in tokenistic, or make-work programs for those of us who feel the desire to remain productive.
Finally, there are the actual trials, which suggest that UBI actually hurts human beings.[15]
A hard pause on AI development
The proverbial nuclear option, but one which aims at converging popular sentiment, with AI doomerism, mandating, banning or otherwise halting AI development.[16][17] The response highlights an uncomfortable, and oft-repeated tension, that (alignment problem notwithstanding, something else I, and of course many others are working on.[18]) if the United States, or the West doesn’t develop advanced AI capabilities, then China will, and the western world can’t afford to fall behind. Any economy therefore that doesn’t leverage AI, to accelerate and develop its economy is making itself less competitive.
One approach calls for an international treaty to pause development of frontier AI systems, until we can solve the alignment problem.[19] There are several problems with this suggestion, for one, it likens a pause on AI development to nuclear weapons, and proposes an international agency similar to the International Atomic Energy Agency (IAEA). The issue with that is, AI development is not the same as a state-actor trying to build a nuclear weapon; if a motivated actor had the resources, it’s possible to disguise frontier AI development say, as a crypto-currency mining operation, and thus scrutiny would be difficult. This also calls into question motivated state actors who’ve resisted international pressure before, and built nuclear weapons anyway, such as North Korea. The establishment of the IAEA happened in 1957, about 8 years after the USSR had successfully tested an atomic weapon. And finally, the efficacy and deadliness of weapons of mass destruction is an established fact. The risks of AI though real, are mostly speculative, and the argument that if there’s any risk, we mustn’t take any chance would seem to risk stifling development. If an international treaty creates a disincentive for capital, where the market has clearly spoken that it anticipates AI adoption as a good return on investment, then creating a disincentive to research risks an outflow of capital, and stifling AI development prematurely. LLMs are still relatively new, highly disruptive, and not well understood.
Every proposed solution so far shares the same flaw: they attempt to resist the direction of the market. But markets are not persuaded—they are redirected. If there is a viable path forward, it must align with the incentives already driving adoption
Working with the Market
In all the despair, is there hope? Yes, but the path is so very narrow. [20]
The answer is hinted at, and often buried under the dire prognostications of AGI doomers, but findings have shown that human-AI augmentation, not replacement, not human job preservation, but humans using AI as cognitive scaffolding outperforms all human, and all AI teams. While it wasn’t on every task, a significant meta-study of AI augmentation found that AI augmentation was a significant enhancement in the efficacy of creation tasks, just not decision tasks.[21] Early market signals are absolutely both definitive, and daunting.[22][23] The market has identified and is pivoting towards AI native skills. Yet, this recognition has caused an intriguing downstream effect: the market is now struggling to identify, recruit, onboard, and retain AI talent in particular, but not AI talent alone.[24]
But the biggest incentive the market has is the same as it’s always had, up to an estimated 5.5 trillion dollars, that’s trillion with a T.[25] But that, is contingent upon leveraging the new tool, and in that regard, it’s estimated that up to 90% of industries will face critical skill shortage soon. Read that with the reality that recruiting AI talent is now a serious preoccupation, and a path out of pathway collapse emerges, narrow though it is.
The Cognitive Unlock, and a very personal anecdote
Is there a solution therefore, that doesn’t try and fight the market? Maybe. This is what I call, the narrow path, and it comes from experience.
On April 2nd, 2025, I knew nothing about AI. I don’t have a computer science, nor mathematics, nor engineering background. By around April 20th, I installed VS Code for the first time. By August, I was contributing regularly to my GitHub (though, most repos are private). By November 2025, I was probing L1 depthwise activations in order to try to answer whether AI models had a personality. The preprint of my findings were made public on TechRxiv in January, 2026.[26] Since then, my preprint paper has attracted 130 downloads, and would have been updated with a much sharper V2, but for the fact that I was waiting on reviews from ICML, and the double-blind review process led me to delay updating my work with more definitive findings.
How? Because artificial intelligence can, and has, at least in my case served as a cognitive unlock. With AI, I am able to do things I couldn’t have done before. And there’s absolutely no reason that I could, or should be alone in that.
So if I can do it, and I shouldn’t be alone, then how do we use my example to adopt a similar narrow path without fighting market incentives, but seeking to work with them instead?
It’s not a guarantee by any means, but there is a way, and it starts with education. But how? After all, isn’t education suffering from AI just as much as students outsource their thinking?[27]
Educational incentives reconstructed
What is the pith and substance of a university-level or college-level education, but to create and train sharp thinkers, whose purpose is to do the cognitive labor necessary to advance society, especially in the STEM fields? So how do we evaluate therefore the quality of that thinking process? Heretofore, our model was to present assignments, and course work, grade the outputs of that coursework, and thereby use a letter, or numerical grade for a student’s ability to work with and through that cognitive labor. But, just as soon as the paradigm was introduced, students, whose incentives were to obtain the best possible letter grade, at any cost, found ways to work the system for efficiency. Mostly incentives did align, but that mismatch gave birth to plagiarism, cheating on tests, and copying another’s homework. Where the system optimized for deliverables, mirroring the market, students also optimized to reduce the friction of producing those deliverables as much as possible. And with LLMs why wouldn’t students decide that their best move is to outsource all deliverable production to LLMs, and they are![28]
Education alone cannot stop pathway collapse, but schools don’t need to restrict their incentive structure to deliverables alone. They can enhance their incentive structure by leveraging the tool that’s causing all the headaches. If students are incentivized to think with AI, to use AI to develop, enhance, sharpen and improve their thinking itself, then those inputs as made to AI can be graded continually, and we can move to evaluating the strength of human cognition itself. Of all things, AI is uniquely positioned to evaluate the structure of a student’s thoughts, without having to rely on the proxies of tests and deliverables.
And therefore, if we change the structure of evaluation itself, if we reward thinking directly in addition to the legacy indicators of proxies, like tests and assignments, we can begin to change the outcome from the bottom-up, instead of trying to impose it from the top down, and, this critically does not require trying to battle against market forces.
And this, actually has another benefit for education itself, because if quality critical thinking can be evaluated and rewarded, thus incentivizing its development, plagiarism, cheating and gaming the system are correspondingly de-incentivized. Students no longer want or feel the need to cheat, not because it’s punished, but because the investment in it is no longer the optimal path to educational success.
Use the AI for improv
So how to restructure education to create AI enhanced thinkers? By making it standard for students to provide an LLM trace along with their essays and other assignments. A model can be used to scan for, and highlight how the student engages with the AI, to create a parallel score to be read in conjunction with the student’s absolute score on deliverables. It’s not unheard of to experiment with grading students according to more than deliverables alone, so why not create a positive incentive where students are encouraged to make use of the tools designed and intended to optimize their learning anyway?[29]
What the model surfaces as high, or low quality thinking could be flagged for human in the loop review. Students could be awarded for;
the quality of their Socratic dialogue,
how they steered the model,
catching and correcting hallucinations, and
overall engagement with the material.
As a general rule, or effective proxy educators could be looking for the “yes, and” rule, meaning, situations where a student builds on a model’s outputs is likely to indicate high engagement with the material. In this way, by shifting the incentives towards engagement, and not making deliverables the only metric, education would be able to more directly focus and zero-in on rewarding the cognitive process, and no longer solely be bound to estimating cognition based on proxies.
We’d also help curb the incentive towards cheating, because cheating becomes a higher friction strategy. While it would still be possible to game the system by having AI talk to AI, it’s easily detectable, and, concealing it takes more work than using the system as intended. In order to try and defeat possible detection, the student would need to prompt a “student AI proxy” to speak in their manner, and then feed those outputs to their personal AI tutor. The submitted trace would be flagged as heavy AI use, but more, the cost of investment in this is actually higher than just prompting the AI tutor, because it involves the extra step of copying and pasting AI outputs back and forth.And worse, doing this repeatedly almost guarantees a low mark on the deliverables, as the student’s voice and effort disappear below AI slop. In this way, cheating is structurally disincentivized, and no longer just punished retroactively.
By shifting the educational rubric to reward the Socratic trace, to markers of student cognitive engagement, and other indicators that the student is maximizing the resource that is AI, we aren’t fighting the market’s drive for efficiency; we are providing the high-level human steering that the market is already desperate to find.
Do we need to overhaul the educational system? Hardly! Teachers regularly engage in professional development and upskilling throughout their careers, and grading rubrics have been changed, enhanced, and standards are updated. If we’re serious about AI augmentation, and we should be, pilots can be launched over a summer of teacher professional development on how to use the tools, and how to help students learn to use them. And teachers could spend less time grading papers and assignments by leveraging AI tools, something they’re already doing.[30] This isn’t a reinvention of how teachers do their job, it’s a small nudge. Maybe the pebble can steer the avalanche after all.
How do you fight one avalanche? With another one
So, if the pebbles can vote, and they should, then who starts? What’s sometimes called post-secondary education; colleges and universities.
Unlike primary and secondary education, which don’t suffer market pressure, universities must compete for students, and students want to choose quality. In a world where AI is rapidly reshaping the nature of work, the institution that most effectively prepares students to operate in an AI-augmented environment gains a decisive advantage.
A university that produces graduates who can:
effectively steer AI systems,
identify and correct model errors, and
demonstrate high-quality cognitive engagement
will outperform one that continues to evaluate students purely on deliverables.
The university, or university cohort (like the Ivy League) who adopts this first will differentiate themselves from the competition, secure a first-mover advantage in preparing students for the transition into an AI-native workforce. And so begins the counter-avalanche.
What begins as a pilot in a single department can rapidly propagate across institutions—not through mandate, but through competition.
And as universities shift, secondary education will face downstream pressure to prepare students for this new standard. All that’s needed now, is for one brave pebble to start the counter-avalanche.
Wu C, Odden MC, Fisher GG, et al Association of retirement age with mortality: a population-based longitudinal study among older adults in the USA J Epidemiol Community Health 2016;70:917-923. https://jech.bmj.com/content/70/9/917
Universe 25 refers to the collapse of Calhoun’s “Mouse Utopia” experiments. In summary, John Calhoun created a “mouse utopia” environment, where mice would have no natural predators, and everything a mouse could possibly want—plenty of living space, food etc. The last living mouse died about 5 years after the beginning of the experiment, with the last successful birth taking place about 2 years into the experiment. A full explanation of the pathologies evidenced in the Universe 25 experiment are beyond the scope of this work, and humans are not mice, and thus, projecting from the Universe 25 experiment into human affairs is necessarily imperfect, however, looking for patterns is not without merit. Consider looksmaxxing, where some practitioners have admitted to using abusing methamphetamine solely for the creation of a more desirable appearance. This sounds a lot to this author like Callhoun’s documentation of “the Beautiful Ones.”
Calhoun JB. Death Squared: The Explosive Growth and Demise of a Mouse Population. Proceedings of the Royal Society of Medicine. 1973;66(1P2):80-88. doi:10.1177/00359157730661P202
OECD (2024), Do Adults Have the Skills They Need to Thrive in a Changing World?: Survey of Adult Skills 2023, OECD Skills Studies, OECD Publishing, Paris, https://doi.org/10.1787/b263dc5d-en.
In this paper, Bélisle-Pipon both points out that UBI is desired by tech oligarchs, and constitutes a form of stagnation. This points back towards the mouse utopia; a world where humans have nothing to do, but trend towards hedonism, or indolence, and not a desirable.
Bélisle-Pipon J-C (2025) AI, universal basic income, and power: symbolic violence in the tech elite’s narrative. Front. Artif. Intell. 8:1488457. doi: 10.3389/frai.2025.1488457
“In our broad sample of recent experiments, the vast majority (about 85%) of the effect sizes were for decision-making tasks in which participants chose among a predefined set of options. But in these cases we found that the average effect size for human–AI synergy was significantly negative. In contrast, only about 10% of the effect sizes researchers studied were for creation tasks—those that involved open-ended responses. And in these cases we found that the average effect size for human–AI synergy was positive and significantly greater than that for decision tasks. This result suggests that studying human–AI synergy for creation tasks—many of which can be done with generative AI—could be an especially fruitful area for research.”
Organizations that integrate AI into their processes intentionally are better positioned to operate faster, make stronger decisions, and achieve sustainable growth. The true advantage of AI automation and augmentation lies not in efficiency alone, but in the strategic benefits that extend beyond day-to-day operations. https://online.hbs.edu/blog/post/business-process-automation
Adrian Hau. Emergent Depthwise Activation Structure in Transformer Language Models The Hau Curve and Its Early-Training Attractor Pattern. TechRxiv. January 12, 2026. DOI:10.36227/techrxiv.176823049.96584847/v1
Vieriu, Aniella Mihaela, and Gabriel Petrea. 2025. “The Impact of Artificial Intelligence (AI) on Students’ Academic Development” Education Sciences 15, no. 3: 343. https://doi.org/10.3390/educsci15030343
Paustian T and Slinger B (2024) Students are using large language models and AI detectors can often detect their use. Front. Educ. 9:1374889. doi: 10.3389/feduc.2024.1374889
Arcta via est—the Narrow path out of gradual disempowerment
Bibli Sacra—Matthaeus 7:14
Ambassador Kosh—Babylon 5
Not so fast, Ambassador. I have an idea.
TLDR: The real problem isn’t job loss, though that is a problem, the real problem is pathway collapse, because entry levels get hit the hardest, and, we might fix that by leveraging the benefits of an AI augmented workforce, if we just make a few small changes to education. Let’s not fight the market, let’s steer it, and change everything.
Introduction
By 1455 though it’s almost certain no one could read it, the writing was on the proverbial wall; what Europe’s still relatively few literate[1] persons could read was the Gutenberg Bible, the first major book to be produced by the printing press—and the crack in the dam that would lead to the democratization of knowledge, the Protestant reformation, and set human civilization onto a fundamentally different path—Gutenberg probably thought that his invention was mostly a more efficient way to produce and distribute printed materials. Life just moved more slowly back then, it’d be about another 400 years or so before widespread literacy really took hold.
Unlike the Europe of 1455, in 2026, Geoffrey Hinton, Yoshua Bengio, and the rest of us certainly can read the writing on the wall, where the world’s newest disruptive technology, the widespread adoption of artificial intelligence will almost certainly be as disruptive; the process has already begun. As these and others warn, our window to steer the avalanche we’ve unleashed might be exceptionally narrow, but at the very least most of us agree; we do have a narrow path out of the worst, most pressing, and already unfolding AI dystopia scenario, that of gradual disempowerment.[2] This work will be making the case that our best shot is to steer the avalanche.
Gradual Disempowerment & the Avalanche Already Underway
As of this writing, 28 March 2026, Layoffs.fyi is showing that 40,482 tech employees had been laid off since the start of 2026.[3] If the trend holds, the back of the napkin math, assuming that this constitutes roughly the first calendar quarter of 2026, we might surmise that we can expect perhaps 170,000 total tech layoffs for 2026, which actually compares favorably to say, 2023 which saw a whopping 264,320 employees lose their jobs—that’s about 100,000 less! But the aggregate numbers don’t tell the whole story, Stanford’s economics lab found that employment for young, entry level workers was being hardest hit.[4] What this points to is a broader trend in labour, the elimination of entry-level roles; “21% of companies have already frozen entry-level hiring due to AI.”[5]
This points to a very troubling trend; the elimination of the lowest-rung of the economic ladder, and the pathway for most people to achieve prosperity, is most vulnerable, as has always historically been the case (citation needed here).
This is the crux and thrust of gradual disempowerment; that these trends, incentivized by existing market forces slowly but surely erode human bargaining power in favor of AI automation and replacement. A world without work could become the promised utopia, (although, I think more likely a dystopia) but we ought to remember that utopia literally means “no place,” and no place is nowhere human workers want to end up because, there’s some evidence to suggest that total retirement is a predictor of mortality.[6]There’s a certain intuitive sense in which purposeless humans aren’t exactly happy and healthy humans. And while humans aren’t mice, if there’s a chance we could end up like Universe 25, we really ought to try. [7][8]
Permanent pathway collapse: the real threat
Since entry-level roles are always the greatest casualty of disruptive technology, we need to reframe the discussion by focusing on those roles in particular, and what entry roles provide. We’re trying to solve for job losses, and while job elimination and job losses are important, What we should be trying to solve is more specific: pathway collapse. By pathway collapse, I mean the process by which some persons, potentially a great many are locked out of prosperity by an inability to secure an entry-level position, because in most circumstances such positions, internships, apprenticeships or equivalents constitute the first, and lowest run on the career ladder. Your first job, gets you your second job, gets you your third job, and so on.
Entry‑level roles serve two functions: they deliver economic value and they gradually transform novices into experts. The pyramid replacement model shows how market incentives gradually transform individual firms away from using humans for anything but the most senior-level jobs.[9] This has the effect of concentrating power and money solely at the upper-levels, and this is the most likely path to gradual disempowerment for a simple reason; it’s the path of least-resistance.
This is not a deliberate assault on workers. It is the outcome of relentless cost‑optimisation. Models get better and cheaper; firms rationally adopt them to cut costs; the training ground quietly erodes. Because the harm accumulates gradually, it is easily missed. The ultimate outcome is one that looks more like the third estate needing to unite against the first two. And earlier citations clearly show; the avalanche has already started, but it might not be too late for the pebbles to vote. But they’d best do so quickly, loudly, and decisively.
But what if there a way to hold pathways open, and stave off gradual disempowerment, at least for a while? Even a short delay could prove decisive, and it is the purpose of this work to demonstrate how AI augmentation could become the narrow road out of the seeming inevitability of gradual disempowerment.
Misaligned Incentives
In 1943, Abraham Maslow proposed a Theory of Human Motivation, and thought he didn’t craft the diagram himself, the pyramid known as “Maslow’s Hierarchy of needs” is in 2026 a fairly ubiquitous lens for understanding human behavior.[10] The industrial revolution started by solving for the physiological, and generally followed the the structure of the pyramid upward. The issue is, the pyramid structure is intentional; it’s very difficult to say, pursue a romantic relationship, the third rung of Maslow’s pyramid if you don’t have shelter, which rests at the bottom rung of the pyramid. Eliminate the bottom rung, and the whole pyramid collapses, as people’s immediate physical needs for food, water and shelter will be re-prioritized away from say, scientific endeavor, or even literacy, and the trend there is alarming to say the least.[11] Because market incentives are such that profitability may favor automation and AI replacement, we might suppose that from a civilizational or national perspective, market incentives, along with those that favor a strong and successful middle-class have gradually been diverging, and so what the discourse is anticipating with artificial intelligence is the completion of a pattern that began with the printing press, but which may hold its final completion in AI replacement.
Critics (and I am one!) will however rightly point out that previous points of inflection and industrialization didn’t eliminate work, quite the contrary, what actually happened historically was a shift in labor distribution.[12] AI we’re told is different, and it might be, because with the advent of humanoid robots, eventual replacement of the workforce at scale is a real possibility, and one which is manifesting, at least on the margins https://fortune.com/2025/08/29/nvidia-ceo-jensen-huang-4-day-work-week-productivity/.
But this brings us back to market incentives—a tried and tested way to improve profitability and productivity is to embrace new technologies quickly, because market competition forces that trend; whoever makes the tech work fastest wins. And with AI, this ends in gradual disempowerment. However, creating or otherwise adjusting incentives isn’t a deterministic process, and here lies the wedge that may forestall pathway collapse. But first, we need to consider the alternative proposals, and why they’re inadequate to stopping pathway collapse.
Trying to fight the avalanche by lighting a match
Several proposed solutions are offered in the public discourse, and they’re all problematic because of a fundamental flaw; they’re trying to fight market incentives instead of working with existing incentives. This section will discuss these suggestions generally before discussing the limits of specific proposals.
Preserving Jobs Directly
If we acknowledge that we need to worry about job loss, and we do, while I’m focusing on pathway collapse, that shouldn’t exclude the vast projected displacement of white collar workers. Various measures are therefore proposed to ensure that people can stay at their jobs, despite any level of AI replacement. These proposals generally take the form of forcing a human to be in the loop. We don’t even need to consider numbers to know why this won’t work; this solution is
Uneven, and limited, even in professions where protection exists,
intentionally introduces inefficiency in a system designed to eliminate it,
will not scale at all,
very temporary at best.
The unevenness looks like this; the bar association is not about to allow ChatGPT Esq. to practice law; presenting legal arguments will remain the domain of lawyer. Law firms who make lawyers have every incentive to take on fewer, and fewer articling students and interns. Since a ChatGPT-like legal assistant can research and summarize complex legal faster and better, it’s more profitable to outsource that task, heretofore often subject of human labor. This is a direct illustration of pathway collapse in the first, and in the second, points out that the profession of law is already protected by a professional association.
This does nothing for professions which enjoy no such protection, such as, software developers, insurance adjusters, finance professionals and stock traders etc. Which leads directly to the second objection, the free market capitalist system seeks to reduce inefficiency, not introduce it, and therefore trying to force a human in the loop through regulation will be opposed by lobbying. And even if that succeeds, if the AI is much faster, and it is, then one human can supervise an increasingly larger team of AI agents. The AI can be trained to surface only those situations which require the humans attention, which leads directly to the fourth objection, this is a temporary at best solution, because of a fundamental flaw; trying to battle market pressure instead of working with it.
A Specific Proposal: Transition into Human Services
A Brookings Institute op-ed proposed what’s more or less a two-pronged approached, which can be boiled down to 1) creating an artificial floor for humans in the loop in human-services jobs like education, healthcare, and counselling, and 2) retraining or otherwise funneling displaced workers into these new vacancies to meet the artificial demand.[13] If we treat this proposal as a type, it fails for five reasons.
1) Imposing an artificial floors for human staffing is a top-down solution which injects friction into a system that profits by minimizing friction. In government-controlled industry like education, this may attract little in the way of pure market-pushback, save that politics will generally favor reducing taxes.
2) Friedler et. al. acknowledge that there’s a bottleneck, and friction in retraining displaced workers. Their solution is more national-level programs, which is top-down, imposed market forcing, rather than leveraging existing trends in the market.
3) It’s doubtful that the macroeconomics will align sufficiently; there are likely to be far more displaced software engineers than new vacancies created, which points back to the previous point.
4) This proposal requires heavy intervention in the market, a market which has long resisted and been stifled by this kind of heavy-handed meddling.
5) Finally, this admission on the part of the authors, “But other institutions will likely have to be built from scratch,” buttresses my previous two points, but additionally, it risks creating a vicious cycle of government liability and dependency. Universe 25: the “Human Utopia” awaits, but in the interim, the demand for government revenue to pay for these “make work” projects spirals upward, imposing a greater burden on the market, which incentivizes the market towards more automation.
Universal Basic Income (UBI)
A second popular suggestion is to push for a top-down, government basic minimal guaranteed standard of living, usually in the form of a species of universal basic income. [14] The proposal is more or less, that the government will tax and re-distribute to ensure that all citizens have access to a minimal standard of living, whether employed or not. Sometimes this is means-tested, sometimes it’s truly universal, meaning everyone who qualifies gets it automatically, employed or not. This is the most seductive option, because it sounds like it could work, until you think about the friction, who pays, and how much? And how does this actually distort the market pricing signals needed to make economic decisions?
It’s not clear that UBI won’t cause an inflationary spiral; after all, why wouldn’t landlords move to squeeze as much as possible out of their UBI-supported tenants, leaving them with little more than a subsistence existence? This would force a phony inflationary pressure on the regular payouts that the beneficiaries of UBI get, which could end up looking much like hyper-inflation. In this manner, estimating the rate needed for UBI, even with the best AI models ends up becoming a precarious exercise, and now we’re into the realm of hoping that hypothetical cause of the problem, widespread AI, needs to become ASI just to manage the whole thing.
But for the sake of argument, let’s presume that the economics of UBI are sound. UBI still has to contend with an error that slew more people in the 20th century than any other; the idea that you can reshape human nature itself. Even under the best-case scenario, UBI might strip individuals of meaningful activity, and one predictor of premature death is full-retirement. In other words, when AI is doing literally everything, what’s the incentive for humans to do anything at all? And the human creature which doesn’t have a gainfully employed purpose is a creature unhappy and this typically predicts morbidity. It is highly optimistic, and perhaps unrealistic to presume that if humans have all their needs and desires met by AI, that they will be happy finding things to do. This suggests that the government might engage in tokenistic, or make-work programs for those of us who feel the desire to remain productive.
Finally, there are the actual trials, which suggest that UBI actually hurts human beings.[15]
A hard pause on AI development
The proverbial nuclear option, but one which aims at converging popular sentiment, with AI doomerism, mandating, banning or otherwise halting AI development.[16][17] The response highlights an uncomfortable, and oft-repeated tension, that (alignment problem notwithstanding, something else I, and of course many others are working on.[18]) if the United States, or the West doesn’t develop advanced AI capabilities, then China will, and the western world can’t afford to fall behind. Any economy therefore that doesn’t leverage AI, to accelerate and develop its economy is making itself less competitive.
One approach calls for an international treaty to pause development of frontier AI systems, until we can solve the alignment problem.[19] There are several problems with this suggestion, for one, it likens a pause on AI development to nuclear weapons, and proposes an international agency similar to the International Atomic Energy Agency (IAEA). The issue with that is, AI development is not the same as a state-actor trying to build a nuclear weapon; if a motivated actor had the resources, it’s possible to disguise frontier AI development say, as a crypto-currency mining operation, and thus scrutiny would be difficult. This also calls into question motivated state actors who’ve resisted international pressure before, and built nuclear weapons anyway, such as North Korea. The establishment of the IAEA happened in 1957, about 8 years after the USSR had successfully tested an atomic weapon. And finally, the efficacy and deadliness of weapons of mass destruction is an established fact. The risks of AI though real, are mostly speculative, and the argument that if there’s any risk, we mustn’t take any chance would seem to risk stifling development. If an international treaty creates a disincentive for capital, where the market has clearly spoken that it anticipates AI adoption as a good return on investment, then creating a disincentive to research risks an outflow of capital, and stifling AI development prematurely. LLMs are still relatively new, highly disruptive, and not well understood.
Every proposed solution so far shares the same flaw: they attempt to resist the direction of the market. But markets are not persuaded—they are redirected. If there is a viable path forward, it must align with the incentives already driving adoption
Working with the Market
In all the despair, is there hope? Yes, but the path is so very narrow. [20]
The answer is hinted at, and often buried under the dire prognostications of AGI doomers, but findings have shown that human-AI augmentation, not replacement, not human job preservation, but humans using AI as cognitive scaffolding outperforms all human, and all AI teams. While it wasn’t on every task, a significant meta-study of AI augmentation found that AI augmentation was a significant enhancement in the efficacy of creation tasks, just not decision tasks.[21] Early market signals are absolutely both definitive, and daunting.[22][23] The market has identified and is pivoting towards AI native skills. Yet, this recognition has caused an intriguing downstream effect: the market is now struggling to identify, recruit, onboard, and retain AI talent in particular, but not AI talent alone.[24]
But the biggest incentive the market has is the same as it’s always had, up to an estimated 5.5 trillion dollars, that’s trillion with a T.[25] But that, is contingent upon leveraging the new tool, and in that regard, it’s estimated that up to 90% of industries will face critical skill shortage soon. Read that with the reality that recruiting AI talent is now a serious preoccupation, and a path out of pathway collapse emerges, narrow though it is.
The Cognitive Unlock, and a very personal anecdote
Is there a solution therefore, that doesn’t try and fight the market? Maybe. This is what I call, the narrow path, and it comes from experience.
On April 2nd, 2025, I knew nothing about AI. I don’t have a computer science, nor mathematics, nor engineering background. By around April 20th, I installed VS Code for the first time. By August, I was contributing regularly to my GitHub (though, most repos are private). By November 2025, I was probing L1 depthwise activations in order to try to answer whether AI models had a personality. The preprint of my findings were made public on TechRxiv in January, 2026.[26] Since then, my preprint paper has attracted 130 downloads, and would have been updated with a much sharper V2, but for the fact that I was waiting on reviews from ICML, and the double-blind review process led me to delay updating my work with more definitive findings.
How? Because artificial intelligence can, and has, at least in my case served as a cognitive unlock. With AI, I am able to do things I couldn’t have done before. And there’s absolutely no reason that I could, or should be alone in that.
So if I can do it, and I shouldn’t be alone, then how do we use my example to adopt a similar narrow path without fighting market incentives, but seeking to work with them instead?
It’s not a guarantee by any means, but there is a way, and it starts with education. But how? After all, isn’t education suffering from AI just as much as students outsource their thinking?[27]
Educational incentives reconstructed
What is the pith and substance of a university-level or college-level education, but to create and train sharp thinkers, whose purpose is to do the cognitive labor necessary to advance society, especially in the STEM fields? So how do we evaluate therefore the quality of that thinking process? Heretofore, our model was to present assignments, and course work, grade the outputs of that coursework, and thereby use a letter, or numerical grade for a student’s ability to work with and through that cognitive labor. But, just as soon as the paradigm was introduced, students, whose incentives were to obtain the best possible letter grade, at any cost, found ways to work the system for efficiency. Mostly incentives did align, but that mismatch gave birth to plagiarism, cheating on tests, and copying another’s homework. Where the system optimized for deliverables, mirroring the market, students also optimized to reduce the friction of producing those deliverables as much as possible. And with LLMs why wouldn’t students decide that their best move is to outsource all deliverable production to LLMs, and they are![28]
Education alone cannot stop pathway collapse, but schools don’t need to restrict their incentive structure to deliverables alone. They can enhance their incentive structure by leveraging the tool that’s causing all the headaches. If students are incentivized to think with AI, to use AI to develop, enhance, sharpen and improve their thinking itself, then those inputs as made to AI can be graded continually, and we can move to evaluating the strength of human cognition itself. Of all things, AI is uniquely positioned to evaluate the structure of a student’s thoughts, without having to rely on the proxies of tests and deliverables.
And therefore, if we change the structure of evaluation itself, if we reward thinking directly in addition to the legacy indicators of proxies, like tests and assignments, we can begin to change the outcome from the bottom-up, instead of trying to impose it from the top down, and, this critically does not require trying to battle against market forces.
And this, actually has another benefit for education itself, because if quality critical thinking can be evaluated and rewarded, thus incentivizing its development, plagiarism, cheating and gaming the system are correspondingly de-incentivized. Students no longer want or feel the need to cheat, not because it’s punished, but because the investment in it is no longer the optimal path to educational success.
Use the AI for improv
So how to restructure education to create AI enhanced thinkers? By making it standard for students to provide an LLM trace along with their essays and other assignments. A model can be used to scan for, and highlight how the student engages with the AI, to create a parallel score to be read in conjunction with the student’s absolute score on deliverables. It’s not unheard of to experiment with grading students according to more than deliverables alone, so why not create a positive incentive where students are encouraged to make use of the tools designed and intended to optimize their learning anyway?[29]
What the model surfaces as high, or low quality thinking could be flagged for human in the loop review. Students could be awarded for;
the quality of their Socratic dialogue,
how they steered the model,
catching and correcting hallucinations, and
overall engagement with the material.
As a general rule, or effective proxy educators could be looking for the “yes, and” rule, meaning, situations where a student builds on a model’s outputs is likely to indicate high engagement with the material. In this way, by shifting the incentives towards engagement, and not making deliverables the only metric, education would be able to more directly focus and zero-in on rewarding the cognitive process, and no longer solely be bound to estimating cognition based on proxies.
We’d also help curb the incentive towards cheating, because cheating becomes a higher friction strategy. While it would still be possible to game the system by having AI talk to AI, it’s easily detectable, and, concealing it takes more work than using the system as intended. In order to try and defeat possible detection, the student would need to prompt a “student AI proxy” to speak in their manner, and then feed those outputs to their personal AI tutor. The submitted trace would be flagged as heavy AI use, but more, the cost of investment in this is actually higher than just prompting the AI tutor, because it involves the extra step of copying and pasting AI outputs back and forth. And worse, doing this repeatedly almost guarantees a low mark on the deliverables, as the student’s voice and effort disappear below AI slop. In this way, cheating is structurally disincentivized, and no longer just punished retroactively.
By shifting the educational rubric to reward the Socratic trace, to markers of student cognitive engagement, and other indicators that the student is maximizing the resource that is AI, we aren’t fighting the market’s drive for efficiency; we are providing the high-level human steering that the market is already desperate to find.
Do we need to overhaul the educational system? Hardly! Teachers regularly engage in professional development and upskilling throughout their careers, and grading rubrics have been changed, enhanced, and standards are updated. If we’re serious about AI augmentation, and we should be, pilots can be launched over a summer of teacher professional development on how to use the tools, and how to help students learn to use them. And teachers could spend less time grading papers and assignments by leveraging AI tools, something they’re already doing.[30] This isn’t a reinvention of how teachers do their job, it’s a small nudge. Maybe the pebble can steer the avalanche after all.
How do you fight one avalanche? With another one
So, if the pebbles can vote, and they should, then who starts? What’s sometimes called post-secondary education; colleges and universities.
Unlike primary and secondary education, which don’t suffer market pressure, universities must compete for students, and students want to choose quality. In a world where AI is rapidly reshaping the nature of work, the institution that most effectively prepares students to operate in an AI-augmented environment gains a decisive advantage.
A university that produces graduates who can:
effectively steer AI systems,
identify and correct model errors, and
demonstrate high-quality cognitive engagement
will outperform one that continues to evaluate students purely on deliverables.
The university, or university cohort (like the Ivy League) who adopts this first will differentiate themselves from the competition, secure a first-mover advantage in preparing students for the transition into an AI-native workforce. And so begins the counter-avalanche.
What begins as a pilot in a single department can rapidly propagate across institutions—not through mandate, but through competition.
And as universities shift, secondary education will face downstream pressure to prepare students for this new standard. All that’s needed now, is for one brave pebble to start the counter-avalanche.
In this paper, Burnigh and Van Zanden use the numbers of printed materials as a proxy to estimate European literacy, it’s generally safe to assume pre-1500 European literacy would be less than 20%. Buringh, Eltjo, and Jan Luiten Van Zanden. “Charting the ‘Rise of the West’: Manuscripts and Printed Books in Europe, A Long-Term Perspective from the Sixth through Eighteenth Centuries.” The Journal of Economic History 69.2 (2009): 409–445. Web.
https://www.cambridge.org/core/journals/journal-of-economic-history/article/abs/charting-the-rise-of-the-west-manuscripts-and-printed-books-in-europe-a-longterm-perspective-from-the-sixth-through-eighteenth-centuries/0740F5F9030A706BB7E9FACCD5D975D4
Kulveit et. al. Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development. https://doi.org/10.48550/arXiv.2501.16946
https://layoffs.fyi/
Brynjolfsson et. al. 2025, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
https://www.resume.org/1-in-5-companies-have-stopped-hiring-entry-level-workers-because-of-ai/
Wu C, Odden MC, Fisher GG, et al Association of retirement age with mortality: a population-based longitudinal study among older adults in the USA J Epidemiol Community Health 2016;70:917-923. https://jech.bmj.com/content/70/9/917
Universe 25 refers to the collapse of Calhoun’s “Mouse Utopia” experiments. In summary, John Calhoun created a “mouse utopia” environment, where mice would have no natural predators, and everything a mouse could possibly want—plenty of living space, food etc. The last living mouse died about 5 years after the beginning of the experiment, with the last successful birth taking place about 2 years into the experiment. A full explanation of the pathologies evidenced in the Universe 25 experiment are beyond the scope of this work, and humans are not mice, and thus, projecting from the Universe 25 experiment into human affairs is necessarily imperfect, however, looking for patterns is not without merit. Consider looksmaxxing, where some practitioners have admitted to using abusing methamphetamine solely for the creation of a more desirable appearance. This sounds a lot to this author like Callhoun’s documentation of “the Beautiful Ones.”
Calhoun JB. Death Squared: The Explosive Growth and Demise of a Mouse Population. Proceedings of the Royal Society of Medicine. 1973;66(1P2):80-88. doi:10.1177/00359157730661P202
CALHOUN JB. Population density and social pathology. Sci Am. 1962 Feb;206:139-48. doi: 10.1038/scientificamerican0262-139. PMID: 13875732.
Drago L. and Laine R. The Intelligence Curse https://intelligence-curse.ai/
Maslow, A. H. (1943). A theory of human motivation. Psychological Review, 50(4), 370–396. https://doi.org/10.1037/h0054346.
OECD (2024), Do Adults Have the Skills They Need to Thrive in a Changing World?: Survey of Adult Skills 2023, OECD Skills Studies, OECD Publishing, Paris, https://doi.org/10.1787/b263dc5d-en.
Nunes, Ashley, Automation Doesn’t Just Create or Destroy Jobs — It Transforms Them. 2021 https://hbr.org/2021/11/automation-doesnt-just-create-or-destroy-jobs-it-transforms-them
Friedler et. al. A people-first vision for the future of work in the age of AI , Brookings Institute, March 2026 https://www.brookings.edu/articles/a-people-first-vision-for-the-future-of-work-in-the-age-of-ai/
In this paper, Bélisle-Pipon both points out that UBI is desired by tech oligarchs, and constitutes a form of stagnation. This points back towards the mouse utopia; a world where humans have nothing to do, but trend towards hedonism, or indolence, and not a desirable.
Bélisle-Pipon J-C (2025) AI, universal basic income, and power: symbolic violence in the tech elite’s narrative. Front. Artif. Intell. 8:1488457. doi: 10.3389/frai.2025.1488457
https://www.cato.org/commentary/universal-basic-income-not-answer-ai-comes-job
“Half of U.S. adults say the increased use of AI in daily life makes them feel more concerned than excited, according to a June 2025 survey. … Another 38% say they are equally concerned and excited.” https://www.pewresearch.org/short-reads/2026/03/12/key-findings-about-how-americans-view-artificial-intelligence/
https://futureoflife.org/open-letter/pause-giant-ai-experiments/
https://github.com/Founder-ArcaFutura/AI-Instrumental-Convergence-Audit
https://pauseai.info/proposal
Diaz, Kathryn, How AI is changing the nature of Entry Level Work. March 2026 https://www.weforum.org/stories/2026/03/how-ai-is-changing-the-nature-of-entry-level-work/
“In our broad sample of recent experiments, the vast majority (about 85%) of the effect sizes were for decision-making tasks in which participants chose among a predefined set of options. But in these cases we found that the average effect size for human–AI synergy was significantly negative. In contrast, only about 10% of the effect sizes researchers studied were for creation tasks—those that involved open-ended responses. And in these cases we found that the average effect size for human–AI synergy was positive and significantly greater than that for decision tasks. This result suggests that studying human–AI synergy for creation tasks—many of which can be done with generative AI—could be an especially fruitful area for research.”
Vaccaro, M., Almaatouq, A. & Malone, T. When combinations of humans and AI are useful: A systematic review and meta-analysis. Nat Hum Behav 8, 2293–2303 (2024). https://doi.org/10.1038/s41562-024-02024-1 https://www.nature.com/articles/s41562-024-02024-1
https://www.bcg.com/publications/2025/ai-shifts-it-budgets-to-growth-investments
Organizations that integrate AI into their processes intentionally are better positioned to operate faster, make stronger decisions, and achieve sustainable growth. The true advantage of AI automation and augmentation lies not in efficiency alone, but in the strategic benefits that extend beyond day-to-day operations. https://online.hbs.edu/blog/post/business-process-automation
https://www.bcg.com/publications/2023/how-to-attract-develop-retain-ai-talent
Work Rewired: Navigating the Human-AI Collaboration Wave https://www.idc.com/resource-center/blog/work-rewired-navigating-the-human-ai-collaboration-wave
Adrian Hau. Emergent Depthwise Activation Structure in Transformer Language Models The Hau Curve and Its Early-Training Attractor Pattern. TechRxiv. January 12, 2026.
DOI: 10.36227/techrxiv.176823049.96584847/v1
Vieriu, Aniella Mihaela, and Gabriel Petrea. 2025. “The Impact of Artificial Intelligence (AI) on Students’ Academic Development” Education Sciences 15, no. 3: 343. https://doi.org/10.3390/educsci15030343
Paustian T and Slinger B (2024) Students are using large language models and AI detectors can often detect their use. Front. Educ. 9:1374889. doi: 10.3389/feduc.2024.1374889
Omari H. Swinton, The effect of effort grading on learning, Economics of Education Review, Volume 29, Issue 6, 2010, Pages 1176-1182, ISSN 0272-7757, https://doi.org/10.1016/j.econedurev.2010.06.014. https://www.sciencedirect.com/science/article/pii/S0272775710000798
https://www.edweek.org/technology/more-teachers-are-using-ai-in-their-classrooms-heres-why/2026/01