The Tether
AI has been built by learning from people, and it can't keep improving by learning only from itself. Models trained on the output of earlier models instead of human data lose the rare and subtle parts of what people know, and within a few generations their output degrades. Later research found the fix, which is to keep real human data in the mix, and newer work suggests that even a small amount of fresh outside data can make the difference. That dependence gives us a tether, because every person is a source of something AI can't make for itself: new data that comes from a real human judgment about what matters and why. This paper lays out the theory, where it holds and where it breaks. It holds best for the things only people can judge, like values, intent and meaning, and it's weakest where machines can check their own work or where human data can be faked. We close with what it would take to hold the tether in place (proof that data came from a real person, fair terms for the people who provide it, and independent checks on the systems that learn from it) and with what the tether means for organizations that deploy AI. Part of the series How Do We Trust AI.
Authors
- A TUCKER
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24231132
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00