Why AI still needs data made by people
AI teams are running short of good human-made data, and the open web is closing its doors. Licensed company data is one of the few new sources left.
Picture made for ABURA Bureau.
Today's AI models were trained largely on public text written by people. Epoch AI estimates the stock of good public text at about 300 trillion tokens, and that AI training could use it all up between 2026 and 2032.
At the same time, that public data is closing off. Researchers tracking AI training sources found that within one year, 45% of a widely used training set had become restricted by websites' terms of use.
Could AI simply make its own data? Researchers writing in Nature showed that models trained on other models' output lose the rare cases over time, and the 2026 Stanford AI Index finds AI-made data is still not replacing real data in training. A further study shows the answer is to keep real, human-made data in the mix.
That leaves data that was never public: the code, drawings, records and research inside companies, made by people doing real work. It is the kind of data AI teams cannot scrape and cannot fake.
Licensing it, with the owner's agreement and with personal data taken out first, is the honest way for that data to reach AI. It gives AI teams what they need, and gives the people who made it a say and a share.
