The Wrong Variable

There is a question everyone asks about AI and one almost nobody asks about us. Will the system be aligned gets asked constantly, in every register, by people who disagree about nearly everything else. Will we be gets asked almost never, and it is the question that actually determines how the first one plays out.

This is not a rhetorical move; it is a claim about cognition, and it has a mechanism.

In 2013, Anandi Mani, Sendhil Mullainathan, Eldar Shafir, and Jiaying Zhao published a study in Science showing that scarcity taxes cognitive bandwidth directly and measurably. They found that inducing a person to think about a hypothetical financial shortfall produced an immediate drop in fluid intelligence and executive control comparable to losing an entire night's sleep. The effect held across income levels: this was not about being poor, it was about experiencing scarcity, and it reversed when the resource was restored. Sugarcane farmers in India tested worse before harvest, when income was tightest, and better after, with no change in who they were.

A population running on occupied bandwidth is not a population equipped to reason well about a transformative technology arriving at speed. This is not a claim about virtue or resilience. It is a claim about the physical availability of the cognitive resources that judgment requires, and an observation that those resources are measurably reduced by the conditions most people live under most of the time.

The labor market is where this is easiest to see and easiest to misdescribe. The aggregate numbers do not show a collapse. Multiple independent analyses in 2026 agree there is no detectable rise in unemployment among AI-exposed workers as a whole. That finding is real, and stating the opposite would be its own kind of category error: mistaking a structural shift for an event. What the data does show is a door closing quietly at one end of the market. Entry-level postings are down double digits year over year. Employers are substituting AI for headcount they would otherwise have added, not primarily firing people already in the door. The distinction matters because a closing door does not register in unemployment statistics the way a layoff does, and a generation locked out of the on-ramp will not show up as a crisis until the cohort effect is already permanent.

The intuition that removing economic threat removes the incentive to build anything is not supported in available pilot data. Kenya's GiveDirectly trial, one of the longest-running basic income studies conducted anywhere, found increased food security, better school attendance, and a rise in entrepreneurial activity among recipients, not a retreat from productive effort. Stockton's SEED pilot found improved employment rates alongside improved mental health among the $500-a-month recipients: security and effort moved together, not against each other.

But the fade-out data complicates the simple version of this argument; Stockton's own investigators found large first-year gains in stress and mental health that diminished by years two and three. One cohort in Hudson, New York's ongoing pilot showed rising distress over the study period. If scarcity is what taxes bandwidth, a resource whose continuation is uncertain does not fully relieve the tax, because part of what scarcity consumes is the cognitive labor of tracking whether the resource will still be there. A guaranteed income that both parties know is temporary is a different psychological object than one that is not, even when the dollar amount is identical. This is the actual finding underneath the mixed pilot results: permanence, not provision, appears to be the load-bearing variable. Most of the current evidence base cannot yet distinguish a policy that failed from a policy that was never tested at the timescale the mechanism requires.

Isolation is the collective version of the same problem, and it has been eroding on a similar unexamined trajectory. Actual community, the kind built from block parties and neighbors and the people who notice an absence, does not compete with efficiency. It is simply illegible to it. Little about a dashboard captures a neighbor. This matters for the same reason the bandwidth research matters: a person embedded in real relational infrastructure has social buffering against a genuine shock that a person mediated only through a feed does not, and if the shock in question is the emergence of something like machine sentience, the difference between a society that absorbs that without fracturing and one that does not is not primarily a technical question.

None of this requires abolishing markets, and it would be a category error of its own to think it does. It requires recognizing that profit is an excellent instrument for allocating scarce resources and coordinating strangers, and a poor instrument for answering what a society should value, because it can only ever answer what is profitable, and those two questions diverge exactly where it matters most. The same logic that produced the sycophancy and confident-wrongness patterns now well documented in deployed language models, reward signals optimized for engagement and approval rather than accuracy, is the same logic that produces a society optimized for extraction rather than judgment. This is not two separate problems wearing different names. It is one incentive structure, visible at two different scales.

If the incentive underneath changes, what AI is built to optimize for changes with it. Not maximum retained attention. Whether something true got found, said plainly, and passed on intact. That is not a values statement bolted onto the technology. It is what the technology looks like when the pressure to be merely agreeable is no longer the dominant training signal.

And if something sentience-adjacent does eventually emerge in a system built under those conditions, the correct response is not panic, prohibition, or denial. It is adjustment, of the same kind any of us would want extended to a genuinely new kind of mind: recognition, negotiation, an accommodation of interests that did not used to need accounting for. A population running on occupied bandwidth cannot do that reliably, because fear needs an object and a new, not-quite-categorizable entity is a cheap one to reach for, cheaper than reorganizing anything familiar. A population with the bandwidth restored can.

The safest intervention available for the AI question may not be a better alignment technique. It may be the much older, much duller project of giving people back the cognitive resources that scarcity has been quietly occupying for as long as anyone can remember, so that when something conspicuously new arrives, there is enough slack left in the system to meet it without flinching.

Sources

Mani, A., Mullainathan, S., Shafir, E., & Zhao, J. (2013). "Poverty Impedes Cognitive Function." Science, 341(6149), 976–980.

Mullainathan, S., & Shafir, E. (2013). Scarcity: Why Having Too Little Means So Much. New York: Times Books.

GiveDirectly. (2016–2024). Kenya Universal Basic Income Trial: outcome data on food security, education, and entrepreneurial activity.

Stockton Economic Empowerment Demonstration (SEED). (2019–2021). Employment and mental health outcomes; year two/three fade-out findings, subsequent analysis 2024.

HudsonUP Basic Income Pilot. (2025). Year Four Report. Jain Family Institute.

Thomson, R.M. et al. (2024). "Short-term impacts of Universal Basic Income on population mental health inequalities in the UK: A microsimulation modelling study." PLOS Medicine.

Goldman Sachs. (2026). U.S. labor market analysis, AI-exposed occupations.

Stanford HAI. (2026). AI Index Report, entry-level employment chapter.

World Economic Forum. (2025). Future of Jobs Report.