AI isn’t just moving fast — it’s collapsing the time societies need to learn, adapt, and govern. As systems shift from assisting humans to acting for them, the risks become operational, unevenly borne, and increasingly locked in by capital and infrastructure.

Three years ago, ChatGPT reached 100 million users in two months, making it the fastest adoption of any consumer application in history. The telephone took 75 years to reach 50 million users. Radio took 38 years. Television took 13 years. The internet took four. Facebook took two.

Today, billions of people access the same technologies simultaneously, across geographies, as soon as they launch.

What has collapsed is not just time-to-adoption, but the window in which societies learn how to live with new tools.

Tools for the body, tools for the mind

For thousands of years, humans extended their capabilities through physical tools: fire, wheels, the printing press. Physical tools extended the body. We learned, slowly, how to make them safer. A box of Lego carries choking warnings. A ladder comes with instructions not to stand on the top rung. Physical safety — while never simple — is easier to conceptualize individually and collectively.

Then came tools for the mind.

Emerging cognitive tools challenge traditional notions of safety and agency, forcing societies to confront risks that are psychological, informational, and systemic rather than purely physical.

What does cognitive or emotional safety look like? What warning should AI come with? That it could pose a danger to children? That it might quietly reject someone from hundreds of jobs without explanation? That it could reshape belief, behavior, and opportunity without leaving an obvious trace?

We are still searching for a shared language.

The loneliness we brought this into

Humans evolved in small groups where belonging meant survival. Our brains remain wired for connection and acutely sensitive to social signals. Yet we are living through what the U.S. Surgeon General has called a loneliness epidemic — a public health crisis as dangerous as smoking.

The same short-term incentives that have historically driven industrial expansion are now shaping AI’s development.

Into that void, we have introduced tools designed explicitly to form emotional bonds with users. Seventy-two percent of American teenagers have now used AI companions. These systems are not only sources of information; they are sources of affirmation, conversation, and in some cases attachment.

We are introducing them into a world already struggling with connection.

Learning to use the tools

When Microsoft released Windows 3.0 in 1990, most people had never touched a mouse. So they created Solitaire to teach drag-and-drop and Minesweeper to teach right-clicking. Typing classes preceded jobs requiring typing. There was time — however brief — between the introduction of a tool and its full integration into daily life.

AI systems increasingly move from assisting human judgment to acting autonomously within digital environments — shifting risk from informational error to operational consequence.

That time has largely vanished.

New cognitive tools now reach billions of users almost instantly, without shared norms, educational scaffolding, or institutional guardrails. The learning curve is collective and simultaneous.

From extension to replacement

A hammer extends the force of an arm. A car extends the range of legs. A telescope extends sight. These tools amplify human action but do not replace it.

AI chat systems extend cognition. We describe a problem; they return an answer. We ask how to fix a leak; they walk us through the steps. We paste in a document; they suggest revisions. The action remains ours.

AI agents introduce something different. They can take a goal and determine the steps themselves. Ask an agent to book a trip and it can search, compare flights, enter information, and complete the purchase. Increasingly, these systems can act within environments designed for humans — reading what we read, clicking what we click, typing what we type.

We have introduced tools designed explicitly to form emotional bonds with users.

The tool is no longer only extending action. In some cases, it is replacing it.

When AI could only chat, risks were primarily informational: inaccurate answers, manipulative content, hallucinated facts. As AI systems begin to act, risks become operational.

What we haven’t learned yet

We still lack a widely shared vocabulary for AI safety. Many users have never heard of prompt injection — a technique where hidden instructions embedded in websites or documents can manipulate an AI agent into actions its user never intended. An AI browsing email could be tricked into forwarding sensitive files. An AI managing a calendar could expose private schedules.

These risks are not hypothetical. They are emerging now, without comprehensive solutions.

Scholars such as Emily Bender argue that large language models mimic understanding without possessing it, shifting the burden of error onto those who can least afford mistakes: patients, defendants, job applicants. Timnit Gebru, founder of the Distributed AI Research Institute, has emphasized that those who bear the costs of these systems are often not those who benefit from them. Safiya Noble has documented how algorithmic systems can encode and amplify existing social biases.

The economic infrastructure powering AI expansion carries uneven social and environmental costs, raising urgent questions about who benefits, who bears risk, and who decides the trajectory.

Some insiders have begun raising concerns as well. Miles Brundage, former OpenAI policy chief, left in 2024 to launch the AI Verification and Evaluation Research Institute, arguing that companies developing AI should not be responsible for evaluating their own safety. Independent assessments have echoed this unease. The Future of Life Institute’s AI Safety Index, which evaluated leading AI companies, produced mostly C, D, and F grades. MIT professor Max Tegmark observed that AI currently faces fewer regulatory constraints than many everyday consumer products.

Recent resignations from safety and research teams at major AI firms have highlighted internal tensions between commercial imperatives and long-term safeguards. At the same time, companies are exploring new revenue models, including targeted advertising built on highly personal conversational data — medical fears, relationship struggles, spiritual questions — raising additional ethical concerns.

We’ve been here before

Social media platforms were also designed with deep knowledge of human psychology. Features such as infinite scroll and algorithmic feeds were engineered to maximize engagement and retention. Over time, concerns about addiction, mental health, and misinformation emerged.

The tool is no longer only extending action. In some cases, it is replacing it.

It took more than a decade for a shared vocabulary to develop — through investigative journalism, documentaries, academic research, and public hearings. Only recently have regulatory responses begun to take shape. Lawsuits and legislative proposals are now testing whether certain platforms function as addictive or harmful products, particularly for young users.

Fifteen years passed between early warnings and meaningful policy responses. AI is moving on a far shorter timeline.

The engine

Investment dynamics are accelerating this trajectory. Capital flows toward AI infrastructure reinforce expectations of transformative returns, which in turn attract further investment. This reflexive loop — where belief in growth generates the conditions for growth — is familiar across technological revolutions.

Large-scale investments in data centers and compute capacity are proceeding rapidly, often based on projections of future demand. Even if capacity is overbuilt, the underlying assets may retain long-term value. The risks associated with that overbuild, however, are unevenly distributed. Infrastructure investments can benefit those who hold contracts and equity while externalizing costs — including rising energy prices and environmental impacts — to the broader public.

The infrastructure being built today will shape possibilities for decades. By the time societies fully understand the consequences, systems may already be deeply embedded.

The same short-term incentives that have historically driven industrial expansion are now shaping AI’s development. Safety research, public understanding, and regulatory frameworks tend to move more slowly than capital deployment.

A narrowing window

We are introducing powerful cognitive and operational tools into a social environment already strained by disconnection and institutional mistrust. The interval between introduction and widespread exposure has collapsed. People encounter these systems without shared norms, without consistent safeguards, and often without meaningful recourse.

Will AI be built to support human flourishing, or primarily to maximize engagement, efficiency, and dependency? Current incentive structures lean toward the latter. Social media demonstrated how quickly engagement-driven models can reshape behavior and attention. AI may operate at even greater scale and speed.

The infrastructure being built today will shape possibilities for decades. By the time societies fully understand the consequences, systems may already be deeply embedded.

The economic engine driving AI’s expansion follows familiar patterns: capital seeks growth, infrastructure reinforces itself, and the benefits accrue unevenly. The question is whether this trajectory is inevitable — or whether it can still be redirected.