AI Safety · Open Testing · Civic AI · 2026-08-26

AI Is Speed, Life Waits

Safety does not only grow where technology is invented; it also grows among those who live with its consequences.

Artificial intelligence is often associated with speed: faster generation, faster deployment, faster iteration. But many important things in life draw their value from waiting. Trust takes time, craft takes maturing, and culture takes generations to pass on. When what we are building will affect the next generation, we usually prefer to go slower—because those not yet born, not yet present, should also have a place in today's decisions.

Slowness is not automatically a virtue. When justice is delayed, waiting becomes harm. But when speed pushes risk onto others and onto the future, going slower can be proof of responsibility. It reminds us that what we seek now is not simply to finish faster, but to avoid making others bear consequences they never chose.

When Paper Became Cheap, Knowledge Began to Flow

I often think about paper.

In the era when paper was expensive, knowledge could only be copied, preserved, and interpreted by a few. People also used stone inscriptions, bamboo slips, parchment, and oral tradition to carry memory across time. As papermaking and printing lowered costs, what changed was not only that "more people could read," but that more people could each keep a copy, write annotations, compare versions, and even point out errors in authoritative editions.

In the Czech writer Bohumil Hrabal's novel Too Loud a Solitude (Příliš hlučná samota), the hero spends half a lifetime working beside the hydraulic press of a wastepaper station, compacting books into bales while rescuing, again and again, the thoughts inside them from destruction. He praises how paper and books, though fragile and thin, carry human thought and wisdom that are immensely tough — impossible to crush or destroy. However light the paper, thought cannot be crushed; yet when words no longer come to rest in a book you can confront, what remains is only noise.

When paper was expensive, knowledge was not a flow but a thing. A single text consumed materials, time, and days or even months of a literate person's life. Whoever could copy, hide, or recite held the power of interpretation. Earlier carriers pushed this to the extreme: oracle bones, bronze inscriptions, and stone steles could barely be moved; writing was bound to ritual and kingship. Bamboo slips made writing possible, yet a single book could still weigh "five carts." Silk was light but cost like clothing. Europe long used parchment, and one manuscript could require the skins of many animals; copying scripture in monasteries was silent labor. What could not be written or carried was entrusted to oral transmission — and oral tradition was monopoly too: whoever held it in memory did the teaching.

Paper slowly dissolved the "heavy slips and costly silk," but books still had to be copied by hand. Ye Mengde wrote that before the Tang, "all books were manuscript copies; the method of printing did not yet exist, and people prized the owning of books." Copying was a profession; a library was status. A single book might exist in only a handful of copies in the whole country, and an error, once made, stayed made. Printing changed more than "more people could read." Woodblock printing existed under the Tang; in the Five Dynasties, Feng Dao petitioned for the official cutting of printing blocks for the Nine Classics, and the Song truly scaled it up; Shen Kuo recorded in the Dream Pool Essays Bi Sheng's movable type of the Qingli era. Once printing spread, book prices fell by more than half within a generation. The same book could now lie open on many desks at once: you could annotate, lay two editions side by side and collate them, point out errors in official or commercial printings. Textual criticism was born this way. Authority was no longer the single scroll.

What truly turned the flow into living water was the digital age and the internet. Copying no longer consumed paper, plates, or people, and publishing no longer needed a bookshop. That anyone could keep a copy, write notes, and point out errors was the promise printing had made. But the shared single volume disappeared. On the timeline, everyone faces a different waterfall. You are no longer sitting across from a book; you are pushed along by endless pushes. Words no longer stop so you can copy, annotate, and compare. Beside the paper press there were at least books worth rescuing; the timeline does not even leave the book. Loneliness has become loud — not because there are no books, but because the book is no longer a place you can confront.

This shift is worth carrying into AI: can we treat intelligent agents as knowledge that is public, portable, and open to inspection?

An intelligent agent is very much the next sheet of paper: light, thin, taking up almost no room, yet able to speak a thought all over again — as the demos of services like Claude already show. But right now it resembles the timeline more than the book. You cannot easily keep "that single volume," lay two answers side by side and compare them, or point to errors in an authoritative edition, because sentences shift with the phrasing of a question and the sources may no longer be there. After printing, humanity took another step: from having books to having public libraries — the same volume placed in a public space, where anyone could borrow it and argue with it. What today's agent services lack is precisely that step: public annotation, and the legitimacy that lets annotations stand. If an agent only sends and will not stop to be copied, annotated, and confronted, it is not the new book — it is only a louder loneliness.

This does not mean all data, memory, and models should be made public without boundaries. Personal privacy, a community's cultural data, and systems with dangerous capabilities all need appropriate protection. The point of publicity is not "expose everything," but that verification rules, sources, and necessary evidence no longer rest in the hands of a few institutions.

When paper was expensive, knowledge could only be kept by the few; when paper became cheap, knowledge could be examined by the many. We hope AI moves toward the latter: letting more people understand, replay, question, and correct—only then does safety have somewhere to grow.

The Story of Radium: Discovery and Safety Grow in Different Places

A story of “radium” research.

In 1898, Marie Curie and Pierre Curie studied pitchblende, reporting polonium and then radium. These discoveries opened the field of radiochemistry and quickly carried radioactive substances into medical and industrial imagination. What people saw then was a "miracle material" that glowed and seemed to promise new cures; the dangers of long-term exposure were barely understood.

In the early twentieth century, young women workers in the United States were employed to paint luminous watch dials with radium-based paint. To keep their brush tips fine, they were taught to point the bristles with their lips. Ingested radium settled into bone, and the damage often appeared only years later: loose teeth, jaw necrosis, anemia, fractures, and cancers revealed another face of the once-miraculous material.

Here an important reversal is needed: those who discovered radium were not the ones who built safety rules for the dial painters. What made the harm visible to society were the women who lived with the consequences, the physicians and pathologists who recorded the lesions, and the lawyers, journalists, and reformers who helped demand accountability. Scientific discovery arises in one field; safety grows in another—even in injured bodies.

So the answers to AI safety will not necessarily come only from researchers in Silicon Valley, corporate leaders, or models atop the leaderboards. They may also come from caregivers, teachers, workers, artists, speakers of endangered languages, people with disabilities, and local communities—those who live daily with the consequences of these systems. They may not use the vocabulary of safety research, but they are the first to know where it hurts, what is unreasonable, and which risks have not yet been named by experts.

From Personal Choice to Public Risk

Suppose a phone is used three hours a day; its owner may not mind a tiny probability of battery explosion. But if everyone in society uses similar devices ten hours a day, explosion is no longer only the buyer's choice. Devices might harm bystanders on trains, in classrooms, hospitals, or homes. Risk crosses the boundary of product and user, and becomes a public problem.

AI is the same. When systems are used occasionally by a few, errors seem individually bearable; when they enter education, care work, employment, credit, journalism, and public administration, model bias, hallucination, and safety flaws travel along institutions, affecting people who never used—and never consented to—the system.

This is why safety cannot be understood merely as "user beware," nor left to the self-certification of makers.

The Importance of Open Testing

Open testing does not mean handing over every secret and dangerous capability without limits. It means letting people from different fields check safety claims under clear boundaries and shared rules: which version was tested? What settings were used? Do complete records exist? Can outsiders ask new questions? Can those affected appeal, correct, or exit?

Because the place that truly hurts is often not where the technology is invented, but where life is lived.

When testing is held only by a few internal experts, many problems are named only after accidents happen. When testing tools, rules, and necessary evidence can be used by more people, experience from another field can illuminate risks in advance. Openness is not the opposite of safety; when designed well, it is how safety leaves the laboratory and enters society.

AI is speed; life waits. Safety stands between them, willing to leave time—for checking, for correction, and for saying no—on behalf of those not yet present.