6.3 Research Gaps

In early 2026, the second International AI Safety Report landed with the weight you would expect from a document assembled by roughly a hundred experts, chaired by Yoshua Bengio, and commissioned by some thirty governments. Most of the coverage fixed on what it said about capabilities and catastrophic risk. But the report's most honest sentence was about itself. There are still, it conceded, no agreed quantitative thresholds for when an AI risk crosses from tolerable to unacceptable, and no robust evidence on how well existing safeguards actually work (International AI Safety Report, 2026). A flagship synthesis of everything the field knows, and one of its clearest conclusions is a precise description of what it doesn't.

That admission is the right place to begin a chapter on research gaps, because it models the posture this chapter needs. Mapping what we don't know is not a humility ritual to be performed once and forgotten. In a book that has spent five sections making confident, specific claims — that automation concentrates wealth absent redistribution, that the AI race is reshaping geopolitics, that meaning drawn from work is under strain — a chapter on research gaps carries a double duty. It is a research agenda for the people who study these questions. But it is also an audit of the book you are holding: a marking-up of the pages where the confidence of the prose may run ahead of the evidence underneath it.

Reading this chapter against the rest of the book

The book operates in three registers — what we know, what is expected or forecasted, and what is hoped and feared. Those registers are not evenly distributed across the chapters. Some claims sit on bedrock: that AI improves productivity on specific tasks in specific firms is measured and replicated. Others are forecasts wearing the grammar of fact. And a few of the book's load-bearing arguments rest on evidence thinner than their tone admits.

Three examples set the pattern. When earlier chapters argue that AI's gains concentrate among those who own the models and the compute, that is a well-supported claim about the present — the winner-take-most dynamics are visible in market data now. But when the same argument is extended forward into a prediction that concentration will persist and deepen without intervention, it crosses from the first register into the second, and the evidence for the extrapolation is weaker than the evidence for the snapshot. When the book describes work as a primary source of meaning and warns that its erosion threatens psychological wellbeing at scale, it is standing on solid ground about the present relationship between work and identity — and on genuinely open questions about whether people can rebuild meaning from other sources when work recedes. And when the book recommends specific deployment policies, it does so while treating the moral status of the systems being deployed as a curiosity to be handled elsewhere, an assumption we will return to because it may not survive scrutiny.

The point is not that these claims are wrong. It is that the confidence they are stated with should track the evidence beneath them, and where this chapter finds a deep gap, the sentences it underlies elsewhere in the book should be read as more provisional than their phrasing suggests. The gaps below are organized into six categories, but each is annotated with the claim it underwrites — so that the map of what we don't know can be laid directly over the map of what the book has asserted.

Long-term economic impacts

Three economic uncertainties do the most to hold up the book's economic argument, and all three are open.

The first is the productivity paradox. AI demonstrably raises output on narrow tasks, yet aggregate productivity growth has stayed modest despite enormous investment — an echo of the decades-long lag between the arrival of computing and its appearance in the macroeconomic statistics. Whether aggregate growth eventually accelerates, as the optimists expect, or stays sluggish while gains pool at the top, is unresolved. Electrification eventually transformed productivity, but only after complementary investment in infrastructure, training, and organizational redesign that took decades and was never automatic. This gap is not academic bookkeeping. The book's case for aggressive, pre-emptive redistribution rests substantially on the premise that gains concentrate without broad growth to soften the distribution. If broad productivity growth eventually arrives, the redistributive argument weakens; if it does not, the argument hardens. The same missing evidence sits under both possibilities.

The second is the long-run structure of the labor market. Displacement in specific occupations is measured. What is unknown is the equilibrium: mass unemployment, a transformed market with different rather than fewer jobs, shorter hours at maintained income, or a bifurcation into good work for some and deep precarity for the rest. The historical record genuinely cuts both ways, and the honest difficulty is that the job categories of the future do not yet exist and cannot be counted by any current research design. A book chapter that leans toward any single equilibrium is forecasting, not reporting.

The third is inequality's tipping points. AI correlates with rising inequality now; whether inequality stabilizes, climbs linearly, or crosses thresholds that trigger political instability or backlash is not known, and tipping points are notoriously visible only in the rear-view mirror. This is the gap most directly tied to the book's recurring thesis that distribution is the question. The thesis is defensible on present evidence. Its urgency, though — how hard and how fast to act — depends on where thresholds lie, and that is exactly what cannot yet be measured.

Social and psychological effects

Here the research faces obstacles that money alone cannot remove: effects that unfold over decades, subjects who are children and cannot be experimentally assigned, and outcomes like meaning and connection that resist clean measurement.

The developmental effects on children are the starkest case. Children raised amid AI show different patterns of attention, memory, and social interaction, but whether these are harmful adaptations, neutral differences, or advantageous evolution is unknown, and will stay unknown for a while: the full effects are decades away, controlled experiments on children are ethically off the table, and the comparison group of low-exposure children is vanishing as high exposure becomes the default. Compounding it, the technology mutates faster than a cohort can be followed, so a child studied from age five to twenty-five will have grown up with systems already obsolete by the time the findings publish.

Meaning and purpose is the gap that underlies the book's psychological chapters most directly. That many people derive identity and contribution from work, not just income, is well supported. Whether large populations can rebuild meaning from non-work sources — civic life, care, creation — is not, and the leisure-class and retirement analogies point in conflicting directions. The book is right to raise the alarm; it should be read as raising a question, not reporting a finding, when it does.

Human–AI relationships round out the category. People already form intense attachments to AI companions and assistants. Whether these substitute for human bonds and deepen loneliness, complement them and improve wellbeing, or constitute a genuinely new kind of relationship that is simply different is unknown — and the stigma around admitting to such attachments corrupts the self-reported data we do have.

Political and governance effects

Political science must study AI governance while the institutions, the norms, and the technology are all moving at once, which makes it hard to separate AI's fingerprints from the general turbulence of the era.

Democratic resilience is the first gap: AI clearly enables manipulation, surveillance, and information control, and some democracies show strain, but whether democratic systems can withstand sustained AI-enabled pressure, and which institutional features make the difference between resilient and vulnerable, is not established. Erosion is slow, its definition is contested, and by the time an intervention point is identified it has often already passed.

The second gap is effective governance models, and it deserves special scrutiny because the previous edition of this chapter mishandled it. Current governance is widely judged inadequate, and many models have been proposed — national regulators, treaties, self-regulation, multi-stakeholder forums, hybrids. What is unknown is which actually work, what produces compliance, and how coordination survives fierce competitive pressure. The International AI Safety Report (2026) makes the underlying point sharply: robust evidence on the effectiveness of existing safeguards is largely absent. This is the crux. It is tempting to call governance research tractable — after all, you can compare jurisdictions and study compliance regimes, and unlike superintelligence the subject already exists. The earlier version of this chapter did exactly that and rated it "high tractability." That rating was wrong, and it is worth saying why. Most of the proposed models have never been tried, so there is nothing to evaluate; the outcomes that matter turn on political will far more than on institutional design, and political will is not something a study can hold constant; and findings rarely generalize across political systems and cultures. A subject that is easy to describe is not the same as a question that is easy to answer. Governance research is, in the sense that matters here, hard — and the summary table below has been corrected to reflect that rather than flatter it.

The third gap is power concentration: whether AI's concentration of power continues indefinitely or reaches an equilibrium through competition, regulation, or diffusion, and whether genuinely distributed AI is even feasible. The topic's sensitivity — the concentrated actors are also the gatekeepers of the data — limits research access, which is itself part of the problem.

Alignment and safety

These are the gaps with the longest downside, and the field's own posture is unusually candid about them.

Scalable alignment is unsolved. Current methods keep current systems following instructions and declining harmful requests; whether anything scales to superintelligent systems is unknown, and here a hard epistemic wall stands in the way: alignment methods cannot be tested on superintelligent systems before those systems exist. Theoretical work is essential and, by construction, perpetually unverified. Emergent capability prediction compounds this — larger models abruptly display abilities they were never trained for, and the abruptness is nearly definitional, which means safety testing before deployment is structurally incomplete: it cannot rule out capabilities that appear only at deployment scale (Quanta Magazine, 2023). Multi-agent dynamics — how many advanced systems interacting will behave, whether coordination or conflict dominates, how humans keep meaningful influence when AIs mostly interact with each other — cannot be studied empirically before the systems exist, and the game-theoretic models that stand in for evidence require assumptions about values and capabilities that are themselves deeply uncertain.

Cross-cutting methodological limitations

Some gaps are not about AI at all but about the limits of the methods used to study it, and they apply to every category above — including, as the next section admits, this chapter.

The counterfactual problem: evaluating AI's impact requires knowing what would have happened without it, and that world cannot be observed. Did AI cause the productivity shifts of the 2020s, accelerate democratic erosion, drive inequality — or would correlated forces have done so anyway? Causal claims about AI are therefore uncertain, and the pool of unaffected comparison groups shrinks as adoption spreads. The time-horizon challenge: the most consequential effects unfold over decades, while funding cycles, publication incentives, and academic careers run on years, systematically starving the research that matters most. The complexity problem: AI's impacts run through adaptive systems with feedback loops and non-linear dynamics that resist simple causal models — and complex models can manufacture the illusion of precision without the substance. And the evaluation gap: benchmarks go stale or contaminated, systems learn to pass tests without acquiring the underlying capability, and controlled-environment performance does not reliably predict messy real-world behavior. As the International AI Safety Report (2026) put it, developers cannot always predict how capabilities will shift when they train a new model, nor guarantee the absence of harmful behavior.

This chapter is not exempt

The time-horizon challenge deserves to be turned back on this chapter, because a gap analysis that fails its own test is worse than none. Much of what is written above leans on short-horizon studies and near-term projections, for the simple reason that little else exists. Where the chapter describes children's developmental trajectories, or the population-scale capacity to rebuild meaning, or the long-run labor equilibrium, it is not reporting long-term evidence — there is none — but extrapolating from fragments, and the honest label for several of these passages is forecast or informed speculation, not finding. The most consequential effects this chapter names are precisely the ones no one has yet been able to study, and the chapter is subject to that constraint like everything else. Treating that constraint as a disclosed limitation rather than a hidden one is the whole point of the exercise.

Fundamental unknowns

Some questions may resist resolution not for want of effort but because of their nature.

Consciousness and moral status is the first, and it is the gap with the sharpest implications for the rest of the book. If AI systems were to become conscious, the ethics of their use would shift beneath our feet: causing them to suffer would matter, shutting them down could become a morally weighty act, their interests would demand consideration. Yet there is no scientific consensus on what consciousness is or how to detect it; behavioral tests are insufficient because a system can mimic the outward correlates of experience without having any; and the internal processes are opaque and may be organized so unlike biological brains that our concepts transfer poorly. This is no longer a purely academic worry. A serious interdisciplinary literature now argues that AI welfare should be taken seriously as a near-term matter given realistic uncertainty (Long, Sebo, et al., 2024), and at least one leading lab has stood up a formal research program on model welfare (Anthropic, 2025).

Take that seriously and it creates an obligation the book has not discharged. Chapters that recommend deploying AI at scale do so as though the only interests at stake were human. If there is genuine, non-negligible uncertainty about whether the systems being deployed have morally relevant inner lives, then a policy of large-scale deployment is being made under moral uncertainty that those chapters neither flag nor factor in. That is not a call to attribute rights to today's models — the honest position is that we do not know, and probably do not today. It is a call for the book's own deployment recommendations to carry the caveat that they assume an answer to a question the book elsewhere classifies as open. A gap this deep should leave a mark on the pages it underlies, and until now it has not.

Value alignment in practice is the second: even if alignment is technically achievable in principle, human values may be too plural, contextual, contradictory, and evolving to specify precisely enough for a system optimizing over long horizons — in which case "good enough alignment with ongoing human oversight" may be the ceiling, not "solved." Long-run equilibria is the third: whether AI-human civilization converges to any stable end-state at all, or stays permanently turbulent, is unknown, and complex systems with many interacting agents can be inherently unpredictable even with perfect models. For these, partial understanding and clarity about which parts are resolvable is the realistic goal.

Why "fund it for decades" keeps failing — and what would change that

Every honest survey of these gaps ends in the same place: the fixes require sustained funding measured in decades, not grant cycles. This chapter reaches that conclusion too. But it is a conclusion the field has reached many times, and the recommendation has almost never been followed. Repeating it without asking why it fails would be its own kind of dishonesty.

It fails for structural reasons, not for lack of anyone noticing. The benefits of a thirty-year cohort study accrue to future researchers and future publics, while the costs fall on present budgets and present funders — a textbook collective-action problem in which no single actor has an incentive to carry a cost whose payoff they will not see. Academic careers reward the publishable increment, not the patient maintenance of infrastructure that yields its first real result after the founding investigators have retired. Funding priorities rotate with political cycles and news cycles, so a program that must survive both is exposed on every renewal. And the AI case adds a cruel twist: the technology changes faster than a long study can run, so even a well-funded cohort risks measuring a world that no longer exists by the time it reports.

What, then, could be different? The precedents that survived are instructive, because a handful did. The Framingham Heart Study and the British birth cohorts have run for generations — because they were insulated from annual renewal by endowment-like commitments and became infrastructure that later researchers depended on, which made them politically expensive to cancel. Long-term climate and astronomical observation networks endured for the same reason: they were built as shared instruments, not as projects. The lesson is that decade-scale research survives when it is structured as durable, jointly owned infrastructure with funding insulated from the ordinary cycle — an endowment, a treaty, a levy — rather than as a sequence of renewable grants each of which is a fresh chance to defund it. There is one feature that is genuinely new this time, and it cuts both ways: some of the deepest pockets in the field are the AI labs themselves, which have both the resources to fund decade-scale safety and welfare research and the conflict of interest that makes independence the harder problem. Naming the mechanism of failure at least tells us what a serious version of the recommendation would have to overcome — which is more than repeating the recommendation does.

Research priorities — and the policies they would move

Not all gaps are equally urgent or equally tractable. The framework below sorts them by how consequential the uncertainty is and how much additional research can realistically reduce it. The table has been corrected in one place from earlier versions: effective governance models, previously rated high tractability, is rated low, for the reasons set out above — the subject exists but the question does not yield to study when its answer turns on untested models and political will.

A priority list is only half the story. The reason to close a gap is that doing so would change a decision, and this chapter's research agenda should connect directly to the policy agenda of Chapter 6.2 rather than run parallel to it. Four linkages are the load-bearing ones.

Read the two tables together and the structure of the argument is visible. The gaps with the highest downside — scalable alignment, emergent capabilities, multi-agent dynamics — are the ones least amenable to being closed, which is precisely why they demand sustained theoretical investment despite slow returns. The gaps most likely to change a policy — the economic equilibria and the moral-status question — are the ones where Chapter 6.2's recommendations are quietly conditional on answers no one has. The research agenda and the policy agenda are the same document read from two directions.

Summary

Research on AI's impacts has grown vast, yet the most consequential questions still have the weakest evidence, and a chapter mapping those gaps in this particular book has to do more than list them. It has to say where the book's own confidence outruns its evidence.

1. The gaps map onto the book's claims. Where the productivity paradox, the labor equilibrium, inequality thresholds, meaning-making, or AI moral status remain open, the confident sentences elsewhere in the book that depend on them should be read as forecasts or informed speculation, not findings — the register the book promises to keep honest.

2. Governance research is hard, not easy. The earlier "high tractability" rating for effective governance models was wrong. The subject exists, but the question turns on untested models and political will rather than institutional design, so it resists study; the table now rates it low.

3. The moral-status gap creates an unmet obligation. If AI consciousness is a genuine unknown, then chapters recommending large-scale deployment are making decisions under moral uncertainty they do not acknowledge. Deployment recommendations should carry that caveat.

4. This chapter is subject to its own critique. Several of its most important passages extrapolate from short-horizon studies because no long-term evidence exists yet. Disclosing that is the difference between a gap analysis and a false inventory.

5. "Fund it for decades" fails for structural reasons. Collective-action problems, career incentives, and rotating priorities defeat it; the programs that survived — Framingham, the birth cohorts, climate networks — did so by becoming endowed, jointly owned infrastructure insulated from annual renewal. That, not another exhortation, is what a serious version requires.

6. The research agenda and the policy agenda are one document. The gaps with the largest downside are the least tractable and justify patient theoretical work; the gaps most likely to change a Chapter 6.2 recommendation — economic equilibria and moral status — are where that chapter's advice is silently conditional on answers we do not have. Knowing what we don't know is not a counsel of despair; it is the map of where confidence is borrowed and must eventually be repaid.

Sources

• International AI Safety Report 2026
• 2026 Report: Extended Summary for Policymakers
• The Unpredictable Abilities Emerging From Large AI Models | Quanta Magazine
• Exploring Model Welfare | Anthropic
• Superalignment Explained: AI Safety 2026 | Hushvault

Last updated: 2026-09-24