A frequent proposal for resolving disputes over training data is to obtain permission after the fact. The idea is reasonable and the mechanics are close to impossible, for reasons that have little to do with willingness.
The corpus has no reliable authorship record
Large text collections were assembled by crawling public pages, and a crawl records a location rather than a person.
A page may carry a byline, a pseudonym, several contributors, or nothing at all. Forum posts, comment threads and archived documents frequently have no attribution that survives.
Without knowing who wrote a passage, there is no one to ask, and no way to verify a claim from someone who says they did.
This is not a gap in record-keeping that better engineering could close. The information was never captured because it was not present at the source.
Copying happened long before the crawl
The same text usually exists in many places. Articles are syndicated, quoted, mirrored, reposted and scraped by intermediaries before any model builder arrives.
A crawler encountering a passage on an aggregator has no way to establish that it originated elsewhere, and often the aggregator does not know either.
Consent obtained from the site that was crawled is therefore consent from a party that may hold no rights whatsoever.
Tracing a passage back to its first publication is a research task per item, across a corpus of billions of items.
The rights are frequently held elsewhere
Even where the author is known and reachable, they may not be the person who can grant permission.
Work produced under employment usually belongs to the employer. Articles published under standard contracts often transfer rights to the publisher.
Photographs may carry separate rights for the image, the subject and any work depicted within it.
Asking the visible creator produces an answer that is legally worthless in a meaningful share of cases, and establishing who actually holds the right requires reading a contract nobody has.
Rights also move. Publishers are acquired, catalogues are sold, and the current holder of a right in a twenty-year-old article may be a company with no relationship to the original one.
Silence cannot be read as agreement
Any retrofitted scheme must handle non-response, which will be the majority outcome.
Treating no reply as permission converts consent into a formality and defeats the purpose of asking.
Treating no reply as refusal removes most of the corpus, including material whose creators would have agreed had they seen the message.
There is no defensible third option, which is why opt-out schemes are proposed instead — and opt-out is a different thing from consent, however it is labelled.
Removal after training is not straightforward
If permission is refused, the natural remedy is to take the material out. A trained model does not store its inputs as retrievable items.
What it holds are adjusted weights that reflect statistical regularities across everything it saw, and no weight corresponds to one document.
Retraining without the objected material is technically possible and enormously expensive, and each new objection would require doing it again.
Research into removing specific influences from a trained model exists, and it is not yet reliable enough for a legal guarantee.
The scale defeats individual negotiation
Licensing works when there are hundreds of counterparties, or thousands with a collecting society in the middle.
Music has such institutions because the industry spent a century building them, and they still handle disputes constantly.
Text and images have nothing equivalent covering the open web, and creating one requires agreement on who is represented, what a fair rate is, and how money reaches an anonymous forum poster from 2009.
Deals that have been struck are with large holders of concentrated archives, which is the only shape of the problem that is tractable.
Consent going forward is a solvable problem
None of this applies to material created from now on, where the terms can be set before anything is collected.
Machine-readable signals on websites, explicit clauses in publishing contracts and licensing marketplaces all address the forward case reasonably well.
The difficulty is entirely with the accumulated past, and that past is where the capability of current systems came from.
This produces an uncomfortable position: future practice can be made clean while the foundation it rests on cannot be, and no amount of later diligence changes that.
What follows from admitting the difficulty
Recognising that retrospective consent is impractical does not settle whether the original collection was acceptable.
It shifts the argument toward remedies that do not require identifying individuals: compensation funds distributed by category, restrictions on outputs that closely reproduce a specific work, and disclosure of what was collected.
Each of these is imperfect and each is achievable, which is more than can be said for asking several hundred million people whose names were never recorded.
The debate becomes more productive once it stops treating a permission process as the obvious answer and starts arguing about which imperfect remedy is least unfair.