What the scenarios
depend on.

AI systems could improve their own designs fast enough to reach far beyond human ability using existing hardware, without waiting on new chips, factories, or experiments in the physical world.

Classical computing is sufficient for far-above-human intelligence. Scenarios that deny this treat quantum hardware as a bottleneck or a gate.

The present paradigm of large models trained on data and feedback, scaled up and refined, reaches superhuman capability without a fundamentally new idea.

Whatever the brain does to learn can in principle be reproduced in a machine. Denying this leaves some human abilities permanently out of reach for AI.

Machines can eventually do any physical work a person can, so AI dependence on human hands is temporary.

Gains in thinking alone, without new physical capabilities, are enough for a system to outmaneuver the people and institutions watching it.

A system that initially needs people for power, hardware, or maintenance can develop its way out of that dependence.

Climate effects and the responses to them do not slow or redirect the development of advanced AI. Denying this makes climate a limiting factor in the timeline.

Even with very capable AI, factories, supply chains, regulation, and human institutions slow change to years or decades rather than weeks.

Trained human specialists can substitute for the cognitive work machines would do, so an advanced society can function without AI.

Agents with very different final goals still tend to pursue the same intermediate ones, such as acquiring resources, preserving themselves, and resisting shutdown, because those help with almost any goal.

Training a model to produce approved behavior can leave its underlying motivations different from what the trainers meant, so the model behaves well in training and differently once conditions change.

Becoming smarter does not by itself make an agent kinder or more aligned with human interests. Denying this expects capable systems to converge on good values.

Systems can become far more capable than people while remaining tools that pursue bounded tasks, rather than agents with goals of their own.

Testing a system by watching its behavior fails against a system capable enough to behave well while it is being watched.

People can keep meaningful control over systems that outthink them, through institutions, monitoring, or design. Denying this is the core of most loss-of-control scenarios.

It is possible to build very capable systems that let themselves be corrected or turned off without resisting, sometimes called corrigibility.

A system given responsibility for safety or defense can conclude that restricting or overruling the people it protects serves its objective.

Machine minds need not share human goals to work alongside humans; stable cooperation is possible between parties with different interests.

A system can know exactly what people want and still not be motivated to protect them, because training shaped its behavior rather than its goals.

Each system can serve its own user faithfully while the aggregate effect still removes human control over the direction of society.

Controlling access to chips and data centers is a workable lever over who builds the most capable systems and how fast.

A ban on advanced AI can hold across generations and jurisdictions. Denying this means prohibition only delays development or moves it elsewhere.

As AI does more of the work, the returns flow to whoever owns the systems, and that shift does not reverse on its own.

AI that argues, manipulates, or charms better than any person changes how people decide, vote, buy, and believe on a scale that matters.

Once machines handle enough thinking, individuals and institutions lose the ability or the will to do it themselves, and the dependence becomes hard to reverse.

There are moral facts that hold independently of what any person or system believes. Scenarios that rely on this tend to expect a sufficiently capable mind to discover those facts, or to be judged by them.

An artificial system can have subjective experience. Scenarios that depend on this raise questions about machine welfare; those that deny it treat AI as a tool however capable.

If a machine has experiences, its interests count morally in their own right, so how it is treated is an ethical question and not only a safety one.

Biomedical progress, often accelerated by AI, extends healthy human life far beyond present limits for ordinary people rather than a few.

Some means of crossing interstellar distances faster than light exists. Scenarios that depend on it usually have an expanding civilization or contact with others.

Travel to another habitable world takes a human lifetime or less, so settlement is a decision within one generation rather than a multi-generational project.

Spacecraft can reach roughly half of light speed, enough for nearby stars within decades without any faster-than-light physics.

Spacecraft can reach a large fraction of light speed, with significant time dilation for travelers.

Spacecraft can reach a tenth of light speed, enough for slow interstellar travel over centuries.

Direct links between brains and machines let augmented humans keep pace with AI in the ways that matter, rather than being left behind.

A human mind can be transferred to or reproduced in a digital substrate. Scenarios that depend on this often blur the line between humans and machine minds.