Steve Jobs was right to call the computer a bicycle for the mind, but only within the limits of the metaphor. A bicycle lets the rider go farther with the same body. It improves the efficiency of effort. The computer did something similar for calculation, writing, design, memory, and search.
The computer was never only a bicycle. Word processors changed the sentence. Spreadsheets changed the organization chart. Search changed what could be remembered and what could be allowed to disappear. The tool did more than extend the individual mind. It altered the medium in which thought took place. Still, the person at the keyboard largely set the goal and judged the result.
AI breaks the analogy at the level of muscle. A bicycle strengthens propulsion without deciding intention. AI works on a different set of human capacities: attention, judgment, memory, imagination, and delegation. It does not simply help thought move faster. It pushes into the conditions under which thought forms.
Large-scale AI does not stay in that register. It proposes, ranks, completes, hallucinates, routes attention, and increasingly acts on behalf of the user. It shapes the destination, the route, and the terrain: the goals it offers, the paths it ranks first, and the information environment it builds around the task.
This is why AI agents matter. Saving time is the shallow version of the story. The deeper issue is delegation. Institutions are beginning to hand over cognitive acts once held inside professional judgment: searching, sorting, drafting, testing, recommending, escalating, and coordinating.
That shift is no longer speculative. Coding agents, security agents, finance agents, customer support agents, and workplace assistants are moving from demonstration to operating layer. The metaphor has already expanded from bicycle to exoskeleton to infrastructure. The question is not whether AI helps the rider move faster. The question is whether the road system is beginning to decide where movement happens.
The adjacent possible is the useful concept here: the set of moves a system can make next. Every tool changes what can be done. A bicycle expands the possible for one body in a given world. An exoskeleton expands what one body can bear, adding strength, memory, and endurance. Infrastructure changes the world that bodies move through. Roads, libraries, markets, databases, platforms, and protocols do not simply assist action. They organize possibility.
AI belongs in that category, with one difference that carries risk: it is infrastructure that learns inside the loop. It can widen the adjacent possible by generating new combinations, simulations, and shared context. It can also narrow it by standardizing language, compressing difference, and steering institutions toward the easiest pattern. When the business model rewards the extraction of attention, labor, and dependency, the adjacent possible is sorted toward what the system can monetize. Without deliberate boundaries, narrowing is the default.
The serious position is neither panic nor boosterism. AI should be treated as contested cognitive infrastructure, not as a neutral productivity tool. It quietly sets defaults for how thinking is distributed across people and machines.
There is AI worth building on these terms. Our work uses it in a narrower register: small language models and machine learning trained for pattern recognition in social and cultural data, kept auditable and under human framing. A system like that remains closer to tool or exoskeleton. It amplifies a judgment the analyst still makes. It does not stand in for that judgment.
The harder problem is the drift toward large, opaque, agentic systems that slide into the system-led column without anyone formally choosing to put them there. The boundary matters: what stays human-led, what becomes hybrid, and what moves into systems. That boundary will not hold because people admire it in a diagram. It has to be designed, audited, and revised as the systems change.
The boundary is the real question, and it does not hold still. Underneath it is a second question: which futures become easier to imagine, which become harder to reach, and who gets to configure the road system we all ride on. Without institutional work, that system settles toward what pays. Drawing it otherwise is not a technical fix. It is a decision about who does the thinking, with what, and on whose terms.
