Control Over Superintelligence Is Handed Over Before It Is Taken
The careful sources do not describe a machine that seizes power. They describe organizations that delegate decisions one at a time under competitive pressure, and delegation is an IS construct.
The popular image of losing control to a superintelligent machine is a takeover. The machine schemes, escapes and seizes power. The more sober sources I read describe something quieter, and for an IS reader something more familiar. Control is delegated, one decision at a time, by people and organizations with good reasons to delegate.
Bengio et al. (2024) put it in one sentence. "Finally, AI systems will not need to plot for influence if it is freely handed over." They also name the pressure behind it. "Companies, governments, and militaries might be forced to deploy AI systems widely and cut back on expensive human verification of AI decisions, or risk being outcompeted". The International AI Safety Report (2025) describes the same path. "In 'passive' loss of control scenarios, important decisions are delegated to AI systems, but the systems' decisions are too opaque, complex, or fast to allow for or incentivise meaningful oversight." It adds a second route, "Alternatively, people stop exercising oversight because they strongly trust the systems' decisions and are not required to exercise oversight". And it names the driver. "Competitive pressures can also incentivise companies or governments to delegate more than they would otherwise choose to, for instance if delegation allows them to stay ahead in a race with competitors." The report is also careful to say that "Expert opinion on the likelihood of loss of control varies greatly."
This is delegation, and IS has a framework for it. Baird and Maruping (2021) define delegation as "transferring rights and responsibilities for task execution and outcomes to another". They note that agentic artifacts "can now assume responsibility for tasks with ambiguous requirements and for seeking optimal outcomes under uncertainty". Ågerfalk (2020) makes a related point about our field. "Another implication for information systems is that these systems can perform social action with only indirect human intervention." Read next to the safety literature, these sentences describe the first steps of passive loss of control. Yet they were not written about superintelligence. They were written about ordinary information systems.
Morris et al. (2024) make the point that matters most to me, which is that capability and autonomy are separate choices. "Increasing capabilities unlock new interaction paradigms, but do not determine them." And "The choice of appropriate autonomy level need not be the maximum achievable given the capabilities of the underlying model." They even picture building something very capable and holding it back. "We may develop an AGI, but choose not to deploy it autonomously, or choose to deploy it with differentiated autonomy levels in distinct circumstances as dictated by contextual considerations." Their own summary is short. "Notably, we observed that AGI is not necessarily synonymous with autonomy."
Drexler (2019) argues that people will often keep the decisions that matter most. In strategic decisions, where the stakes are higher and there is less time pressure, he writes that "humans can exploit AI competence without ceding control". He goes further. "In light of these considerations, we can expect senior human decision makers to choose to retain their authority, and without necessarily sacrificing competitiveness." I would like to believe him. But his argument and the warning from Bengio and colleagues disagree on one point, which is competition. Drexler expects authority to be kept without a loss of competitiveness. Bengio and colleagues, and the Safety Report, expect competition to push organizations the other way. That disagreement is an empirical question about organizations, and it is the kind of question IS research knows how to study.
What would it take to study it? We already study who decides what in organizations. We can record which decisions are delegated to a system, who can override it, how often overrides happen, and whether checking steps are cut when budgets tighten. The phrase from Bengio and colleagues, "cut back on expensive human verification of AI decisions", names something we can observe. My guess is that if passive loss of control happens, it will show up first as a slow decline in that kind of checking, long before anything that looks like a takeover.
The note on delegation in human-AI collaboration found that lab studies nearly always keep the person as the final decision maker. Real organizations will not be so tidy. Under competitive pressure the arrangement drifts, and no single step looks like giving up control. That is why I think the IS view of delegation, a sequence of choices with rights and responsibilities attached to each, is more useful here than the image of a machine that takes power. It also connects to the older argument that AI agents create a principal-agent problem at machine speed. The faster the agent, the more tempting it is to stop checking.
None of this tells us how capable future systems will be, and the sources disagree about how likely loss of control is. What it tells me is where to look. My reading of these sources is that if control is lost, it is most likely to be lost through decisions that organizations make about delegation, and those decisions can be studied now.