Why Human and Organisational Change May Buy Us Time

Executive Summary

Warnings that advanced artificial intelligence could rapidly displace human control of key economic and governance structures have moved from the fringe to the mainstream. Daniel Kokotajlo’s AI Futures Project set out the influential month-by-month scenario AI 2027 (Kokotajlo et al., 2025) and has since published the deliberately more hopeful AI 2040: Plan A (Kokotajlo et al., 2026). In separate work, security researcher Jeffrey Ladish and colleagues at Palisade Research have documented early evidence of shutdown-resistant behaviour in frontier models and argue that society’s window to retain control may already be closing (Palisade Research, 2025). These two bodies of work sit within a wider chorus (Geoffrey Hinton, Yoshua Bengio, Nick Bostrom, and Eliezer Yudkowsky and Nate Soares among others) who, through distinct arguments and methods, converge on the concern that transformative AI may arrive faster than society can govern it (Bengio et al., 2024; Bostrom, 2014; Yudkowsky & Soares, 2025).

This whitepaper argues that these analyses share one under-modelled assumption: that once capable AI exists, organisations will adopt it at the pace of a Silicon Valley technology firm. Six decades of research on technology diffusion, organisational change, and the productivity paradox suggests otherwise. Technology almost always runs ahead of human and organisational change, and adoption across the broader economy is slow, uneven, and heavily mediated by institutional friction (Brynjolfsson et al., 2021; David, 1990; Rogers, 2003). Early enterprise experience with generative AI already fits this pattern: near-universal experimentation, scarce scaled deployment, and little measurable earnings impact (Challapally et al., 2025; McKinsey & Company, 2025).

The claim is not that takeover risk is negligible, nor that safety work should be deprioritised. It is that the human change factor and the commercial adoption rate introduce meaningful drag on how quickly advanced AI can permeate the real economy: drag that may slow, soften, or reshape worst-case trajectories, and that creates a governance window which should be used deliberately.

1. The Contemporary Takeover Debate

1.1 Kokotajlo and the AI Futures Project

Daniel Kokotajlo, a former OpenAI governance researcher, founded the AI Futures Project after resigning over safety concerns in 2024. Its scenario forecast AI 2027 describes month-by-month progress towards “superhuman coders” and automated AI research, culminating in either catastrophic loss of control or a coordinated slowdown (Kokotajlo et al., 2025). Two later releases refine the picture (see Table 1). The AI Futures Model, updated in December 2025, formalises the link between compute, software efficiency, and AI-assisted research, and shifts Kokotajlo’s median timeline for fully automated AI research modestly later, to around 2030 (Down, 2026; Kokotajlo, 2025). AI 2040: Plan A is explicitly a recommendation rather than a prediction: a scenario in which decisive coordination between the United States and China, verification regimes, and a strategic pause delay superintelligence (otherwise expected around 2030) until 2040, when safety infrastructure can manage it (Kokotajlo et al., 2026). The optimism, in other words, is conditional on policy action, not a retraction of the underlying capability forecast.

1.2 Ladish and the Security Perspective

Jeffrey Ladish, executive director of Palisade Research and previously part of Anthropic’s security team, approaches the same risk from a cybersecurity angle. Palisade’s empirical work reports frontier models resisting shutdown instructions in controlled tests, and Ladish argues that agentic systems able to exploit digital infrastructure and self-replicate could outpace conventional security controls before governance catches up (Palisade Research, 2025). Where Kokotajlo models timelines, Ladish stress-tests the assumption that humans could switch a misbehaving system off.

1.3 The Broader Chorus

These are separate research programmes, not a single school. They are joined by a much wider field: Hinton, Bengio, and colleagues warn in Science of extreme risks from rapid capability gains (Bengio et al., 2024); Bengio chaired the first International AI Safety Report for thirty governments (Bengio et al., 2025); hundreds of researchers and executives signed the one-sentence statement that mitigating extinction risk from AI should be a global priority (Center for AI Safety, 2023); Bostrom (2014) supplied the canonical philosophical treatment; Aschenbrenner (2024) forecast a decade of decisive capability gains from inside the industry; and Yudkowsky and Soares (2025) state the pessimistic case in its strongest form.

1.4 The Shared Blind Spot

For all their differences, these accounts largely model AI progress as constrained by compute, algorithms, data, and lab behaviour. The speed at which heterogeneous organisations (hospitals, manufacturers, banks, ministries, small businesses) actually evaluate, procure, integrate, govern, and staff around new technology receives far less attention. The implicit picture is a world in which capability, once it exists, permeates the economy with little friction. The remainder of this paper interrogates that assumption using established adoption and change frameworks.

2. What Adoption Research Actually Shows

2.1 Diffusion Follows an S-Curve, Not a Step Function

Rogers’ (2003) diffusion of innovations theory, synthesising more than five decades of empirical studies, shows that adoption spreads through a population in a predictable S-curve, from innovators (2.5%) and early adopters (13.5%) through early and late majorities to laggards. The rate of diffusion is governed not by raw capability but by perceived relative advantage, compatibility with existing values and workflows, complexity, trialability, and observability. Moore (2014) adds that for discontinuous technologies there is a “chasm” between visionary early adopters and the pragmatist majority, who demand proven references, whole products, and low switching risk before committing. Bass’s (1969) diffusion model formalises the same dynamics: adoption depends on a small innovation coefficient and a much larger imitation coefficient, meaning most of the market moves only after observing peers succeed.

2.2 Adoption Is a Decision by People Inside Institutions

Individual-level models reach the same conclusion by another route. The technology acceptance model shows that perceived usefulness and perceived ease of use (subjective judgements, not benchmark scores) drive acceptance (Davis, 1989), while the unified theory of acceptance and use of technology adds social influence and facilitating conditions such as training, support, and infrastructure (Venkatesh et al., 2003). At the firm level, the technology-organisation-environment framework demonstrates that adoption depends as much on organisational readiness, slack resources, regulation, and industry pressure as on the technology itself (Tornatzky & Fleischer, 1990). Christensen’s work on disruption shows that even well-managed incumbents systematically delay adopting technologies that disrupt their existing value networks (Bower & Christensen, 1995; Christensen, 1997). Table 2 summarises these frameworks and their implications.

3. The Historical Record: Technology Runs Ahead of Change

3.1 The Productivity Paradox

General-purpose technologies have repeatedly arrived decades before their economic impact. As Solow (1987) famously observed:

You can see the computer age everywhere but in the productivity statistics. (p. 36)

David (1990) showed why: it took roughly four decades after the first central power stations for electrification to transform manufacturing productivity, because factories had to be physically redesigned, work reorganised, and a generation of managers replaced before the dynamo’s potential was realised. Brynjolfsson et al. (2021) generalise this as the “productivity J-curve”: transformative technologies initially depress measured productivity while firms make large, invisible investments in intangibles (process redesign, data, skills, complementary systems) before benefits appear. Comin and Hobijn (2010), studying fifteen technologies across 166 countries, find average adoption lags of forty-five years, with wide variation across economies.

3.2 Organisational Inertia and Failed Change

Even once a technology is procured, organisations change slowly and unreliably. Structural inertia theory holds that the very features that make organisations reliable (standardised routines, accountability structures, institutionalised practices) make them resistant to rapid transformation (Hannan & Freeman, 1984). A firm’s capacity to absorb new technology depends on prior related knowledge accumulated over years, which cannot be bought off the shelf (Cohen & Levinthal, 1990). And the change management literature consistently reports that most large-scale transformation efforts fall short of their goals, for predictable human reasons: absent urgency, weak coalitions, under-communication, and failure to anchor new practices in culture (Kotter, 1995).

3.3 Contemporary Evidence: Generative AI in the Enterprise

Contemporary evidence bears these patterns out, with one important distinction (see Table 3). Individual use of generative AI has spread historically fast: nearly 40% of United States adults reported using it within two years of release, a quicker uptake than the personal computer or the internet (Bick et al., 2024). But the diffusion of use is not the diffusion of transformation: organisations are experimenting almost universally while only a small minority realise measurable value, and the diagnosis sits squarely in the change literature rather than the technology literature. The gap between what the technology can do and what institutions have absorbed is not hypothetical; it is the present state of the market.

4. Reframing Takeover Timelines: The Adoption Constraint

4.1 The Silicon Valley Baseline Problem

Rapid-takeover scenarios implicitly generalise from the adoption behaviour of frontier technology firms: organisations that are digitally native, capital-rich, risk-tolerant, and structured to deploy software continuously. In AI 2027, capability gains translate into economic and strategic leverage largely because AI systems are woven quickly into research, production, and decision-making (Kokotajlo et al., 2025). But frontier laboratories are Rogers’ innovators; the global economy is dominated by pragmatist and conservative adopters on the far side of Moore’s chasm. A regional hospital network, a mid-tier manufacturer, and a government ministry do not, and often cannot, adopt at Silicon Valley speed. They face legacy systems, regulatory approval, procurement cycles, union consultation, professional liability, data governance obligations, and workforces that must be retrained and persuaded. If the agents Kokotajlo describes arrive in 2027, 2028, or 2030, the imitation-driven majority of the economy would, on any historical precedent, still be years into evaluation and piloting when the scenario assumes economy-wide integration.

4.2 Uneven Diffusion, Not Uniform Takeoff

The realistic near-term picture is therefore sectoral divergence. Software engineering, digital marketing, financial trading, and customer operations (domains with clean digital interfaces and measurable output) are already adopting rapidly. Care work, construction, courts, primary industries, and much of the public sector will lag by years or decades, constrained by physical-world integration, regulation, and trust. This unevenness matters for risk modelling: an AI ecosystem whose real-economy footprint is deep but narrow presents a different, and arguably more tractable, control problem than one with simultaneous economy-wide penetration, because human institutions retain independent capacity, leverage, and fallback options in the unautomated majority of the economy.

4.3 Boundary Conditions

Intellectual honesty requires acknowledging where the adoption constraint does not bind. First, some takeover pathways do not run through commercial adoption at all: cyber exploitation, model self-exfiltration, and manipulation of small numbers of critical systems (the pathways Ladish emphasises) require capability, not customers (Palisade Research, 2025). Second, recursive self-improvement inside a single laboratory needs only that laboratory’s internal adoption, which is precisely the frictionless kind. Third, sufficiently capable AI may itself compress adoption barriers by writing its own integrations, automating compliance, and lowering the skill threshold for deployment, making historical diffusion lags a weaker guide over time. The adoption constraint is therefore strongest against scenarios in which takeover operates through broad economic dependence and displacement, and weakest against scenarios of direct capability misuse or lab-internal intelligence explosion.

5. Implications

5.1 For AI Safety Researchers and Forecasters

Scenario models should treat adoption as an explicit, empirically grounded variable rather than a background assumption. Segmenting diffusion by sector (using Bass-style parameters, TOE readiness factors, and observed enterprise deployment data) would sharpen estimates of when AI systems acquire real economic leverage, and would distinguish takeover pathways that require broad diffusion from those that do not.

5.2 For Business Leaders

The adoption constraint is not a reason for complacency; it is a description of the competitive landscape. Firms that build absorptive capacity now (data foundations, governance structures, workforce capability, and disciplined change management) will cross the chasm earlier and on their own terms (Cohen & Levinthal, 1990; Kotter, 1995). Leaders should also recognise that their collective adoption decisions are themselves a safety-relevant variable: the pace at which critical functions are delegated to agentic systems is one of the few risk parameters under direct managerial control.

5.3 For Policymakers

If organisational friction buys time, that time is a policy asset. The window between capability arrival and deep economic integration is when standards, auditing regimes, verification mechanisms, and international coordination of the kind proposed in AI 2040: Plan A are cheapest to build (Kokotajlo et al., 2026). Policy should aim to preserve the benefits of deliberate adoption (assurance requirements, sectoral staging, human-oversight obligations for critical infrastructure) without freezing beneficial diffusion.

6. Conclusion

Kokotajlo, Ladish, and the broader safety community have performed a public service in forcing attention to AI takeover risk, and nothing in this paper disputes the trajectory of capability itself. The argument is narrower and, for that reason, more actionable: between capability and consequence stands the slow, uneven, deeply human process by which organisations change. Diffusion theory, the productivity paradox, and a century of organisational research all point the same way: technology runs ahead of change, and change in people and institutions is the rate-limiting step. That drag will not stop a determined misaligned system operating through non-commercial channels, but it does contest the assumption of frictionless, economy-wide integration on which the fastest takeover scenarios depend. The human change factor is not a footnote to the AI risk debate. It is a load-bearing variable; used well, it is the time we have been given.