Two words come up in nearly every conversation about training data, and they are the defense the industry reaches for: fair use. What I do not understand is how the balance inside them got skipped. This one phrase, a combination of two simple English words, defines when and how one can use a type of data. Extracting a passage of text from a web page and using it for training is usually justified with urgency: the race to AGI. I don't find that persuasive on its own. A.I. is a means to an end, the same end computing has been optimizing towards since DARPA connected the world, and the knowledge it carries should be accessible. But somewhere along the way, what it took to protect the creative commons got left behind. Without that balance, the idea of an "internet" is not possible.
When a person puts in the time to stitch words together into a document that gives a reader knowledge, the reader extracts that knowledge and can then contribute to society more effectively, remixing it into a return of some kind, whether a new UX feature in their application or an addendum to their own blog post. The idea of open source does not just apply to code. It applies to how the world shares intelligence and evolves together. "Decentralizing intelligence centralizes growth for the whole."
When that knowledge is taken without consent to train siloed, compressed machines of thought that can't credit the people who produced these paths of learning, the essence, the balance, the idea of the internet starts to erode. My objection is not to machines reading the public web — search does, and so does the detector I build. It is to generative systems that absorb people's work and hand it back with no consent, credit, or compensation. I believe "dead internet theory" is reversible, and the reversal starts with building solutions: courts are still adapting, and there isn't yet a working alternative to point to when the usual arguments for collecting training data without consent come up.
I don't believe "AGI" is the goal that matters most. Humans evolve, and A.I. will mostly prove that humans scale: a model will always be racing to gather more knowledge, from every source it can reach, just to keep pace with how people adapt to the systems around them. That race never ends, and I will always bet on the human to outmaneuver a machine. So the intelligence problem is being framed incorrectly. Data should be clearly owned by the individual, private and personal, and leased on request into the systems that aggregate it.
Large language models, at their core, are aggregates of sources within a corpus of data. This aggregate can be cited back to its sources, and that truth should be exposed and transparent rather than jumbled together in an S3 bucket, waiting to be pulled for another epoch.
That calls for a framework: the architecture and the proper storage units to do exactly that. Progress at the frontier only keeps its credibility if the roads are paved along the way. Otherwise, the credibility of the very invention billions are poured into falters at the expense of the people who helped build it in the first place. The creatives, the intellectuals, the commons.