The book left a rare-books dealer with a hidden pulse. On a screen, a reporter watched it move through the logistics bloodstream of America until the signal came to rest at VGT3, an Amazon facility in Las Vegas, then went dark. Inside, printed books arrive in bulk. Their bindings are cut away so the loose pages can be scanned faster. The physical volumes are destroyed. The team's logo shows a dinosaur clutching a book in its claws, teeth bared. There is no bonfire, smoke or jeering crowd, only a blade, a scanner and an efficient workflow. Are we watching the last human harvest, or the beginning of a fight over who controls the data that AI's hungry ghosts demand?
The Amazon episode caused outrage. It felt like a lost scene from Ray Bradbury's Fahrenheit 451 rewritten by William Gibson. Yet the problem is more complicated. Book burning destroys the object to suppress the ideas it carries. Destructive scanning extracts the patterns and discards the vessel. The language may survive inside a machine while the binding, edition, ownership history and route back to that copy disappear. A library preserves a source so future readers can return to it. A model turns the source into a capability that can answer, imitate, summarize and generate. The book becomes part of what the machine can do.
AI companies are searching for material that is clearly human, hard to find and missing from the public internet. Research has shown that repeatedly training models on model-generated material without enough real data can cause them to lose some of the range and variety found in the original material. While carefully designed synthetic data can still be useful, older texts with clear human origins have acquired new value. A useful analogy is low-background steel, historically used for some sensitive radiation measurements because it was made before atmospheric nuclear testing added faint radioactive traces to later steel. Older books carry a clearer signal from before machine-generated content became common. The analogy has limits, but it marks an important threshold. We are moving from a world in which most culture was made by people into one where human and machine output will be harder to separate.
The appetite for raw data is already moving beyond our cultural record. Shift, an offshoot of German AI lab MicroAGI, offers free apartment cleaning in New York in exchange for permission to record the work. Cleaners wear head-mounted cameras as they wash dishes, mop floors and fold laundry, producing first-person footage for AI and robotics training. Figure's new Index platform takes the same idea to global scale. Figure says more than 44,000 weekly contributors across 108 countries have uploaded over 16 million videos. Figure calls them "Creators", which seems a little odd. Whatever the label, the company has paid them $15 million and receives 30 minutes of new footage every second.
Robotics needs a different kind of data from the text models that power customer-service chat. The internet contains countless pictures of clean kitchens. It does not contain a full record of how hands, tools and objects interact in millions of real homes. A robot needs to see what people do when a drawer sticks, a glass begins to slip or an unfamiliar appliance behaves differently from anything it has seen before. Each contributor to Index brings another room, another set of objects and another way of solving the same problem. Figure is trying to turn that variety into machines that can handle ordinary life.
Index also hints at a different bargain. Contributors know they are creating training data and receive payment for accepted work. That is a better starting point than invisible scraping. Even so, the lasting stake remains with Figure. It chooses the tasks, filters the footage, owns the data pipeline and controls the robot capability that results. Once the system has seen enough examples of folding a shirt, the value of the next example may fall sharply. People are paid for the demonstration. Figure keeps what the demonstration teaches.
The same one-sided exchange is reaching professional work. Mercor recruits doctors, lawyers, engineers, consultants and other specialists to review model outputs, write strong examples and explain why an answer fails. This captures judgment that rarely appears in a textbook: which detail changes a diagnosis, which exception breaks a rule and which recommendation would survive contact with a client. An expert can be paid for an hour of evaluation. The capability shaped by that hour can be used across thousands of future tasks. People who spent years building their expertise are being hired to make parts of it easier to reproduce.
Even a failed company now leaves behind data worth buying. Google placed the winning $10 million bid at a bankruptcy auction for Spirit Airlines' internal business data, with the sale still subject to court approval. The proposed deal includes employee emails, Teams messages, spreadsheets, calendars and operating records. Spirit's flight attendants have objected, warning that links across the datasets may still expose information about individuals or small groups. The aircraft can be sold and the routes abandoned, while the record of how employees communicated and solved problems remains valuable. A dead company can leave behind a cognitive estate.
These stories point to the same transfer of power. The obvious story is a mass ingestion of human knowledge and a coming bonfire of human jobs. The harder question is who owns the capability after the transfer. Human knowledge begins in scattered places: a writer's mind, an engineer's hands, a professional's judgment or the daily routines of a company. Once that knowledge becomes machine capability, it is easier to copy, combine and deploy. The person or institution at the source often receives a one-time payment. The buyer gains an asset that can keep producing value. Data is becoming the channel through which useful human capability moves into systems owned by someone else.
There is a fin de siècle feeling to this moment. We may be living through the closing years of a world in which most culture and work were produced without generative systems in the loop. More books will be written with AI assistance. More professionals will make decisions alongside agents. Homes and factories will contain more sensors and robots. The record of culture and work created before generative AI became part of everyday production is finite. There will never be more of it. That creates a rush to capture it. Seen from this angle, the great ingestion looks like a final harvest of human-made experience.
Useful AI, however, cannot live on old archives alone. Systems that perform real work need fresh examples, corrections and outcomes as the world changes. There may eventually be enough footage to teach a robot one narrow household task. There will never be a final dataset for treating disease, managing a supply chain, running a laboratory or leading a company. Rules change, equipment fails in new ways and unfamiliar situations appear. A book can be scanned once. A hospital can keep connecting decisions to outcomes. A factory can keep revealing failures. The most valuable resource may become ongoing access to places where new evidence is produced.
Anthropic's new Model Hardware Standard shows how far this could go. It allows AI agents to operate programmable equipment such as microscopes, robotic arms and lasers, monitor the results and change their approach. In one project with quantum-computing company QuEra, an agent tested ways to recover a laser that had lost its precise frequency lock. The controller it developed restored the lock in 99.3 percent of blind trials without human intervention. The significance is simple: the agent could act on the world, observe the result and use that evidence to produce a better procedure.
Once AI can generate evidence through action, access to the real world becomes a source of power. AI companies will need homes, factories, laboratories, hospitals and professional communities. They will need permission to observe, test, measure and return for the next case. Figure's Index, for all its unequal terms, shows the beginning of this shift. The company is paying because the physical data it needs cannot simply be found online.
This dependence gives people a chance to regain some leverage. Workers could negotiate rights over recordings made during their jobs. Professionals and their associations could pool current case knowledge and license access on shared terms. Companies could provide limited, metered access to operational feedback instead of selling complete archives. Laboratories could let agents use equipment while retaining the experimental data. Communities could set conditions for the use of their language, history and cultural knowledge. Payment could continue for as long as access continues, or rise when the data produces measurable value.
None of this will happen on its own. Platforms can gather thousands of contributors and weaken the bargaining power of any one person. People under financial pressure may sign away permanent rights for a small payment. Once a model has learned a capability, withdrawing the original data may have little effect. The window for setting better terms may be brief.
The same hunger that drives extraction also reveals a dependency. The most useful future data will often come from trusted sources that can maintain quality, explain where it came from and connect actions to outcomes. AI companies need people and institutions that can keep producing relevant experience. That need gives us a chance to demand a better role than raw material or temporary tutor.
The tracker went dark at VGT3. It could follow the book into the warehouse, but not the capability extracted from it. That is the weakness in today's bargain. Once human experience enters a model, its origin becomes hard to see, its contribution hard to price and its creators lose leverage. The deeper danger is that we hand over the record, the pipeline and the power in one transaction. The choice is ours: surrender our knowledge and experience one upload at a time, or set the terms every time AI comes back for more.