Where new human data comes from when the public supply runs out

The headline version of this problem is that the internet is running out of text. The operational version is more specific and more useful. The stock of public human writing has a measured size, training runs are consuming it on a projected schedule, and the substitutes on offer work in some places and fail in others. If your roadmap assumes another decade of scraping, the question worth answering now is not whether the supply tightens but which mechanisms produce genuinely new human data once it does, and what each one costs you in access, exclusivity, and control over quality.

Key takeaways

• Epoch AI estimates the stock of public human-generated text at roughly 300 trillion tokens, and projects that training sets reach parity with it between 2026 and 2032.
• Substituting generated data does not close the shortfall. Recursive training on model output produces irreversible defects, and the tails of the original distribution disappear first.
• Retaining a fraction of real data slowed collapse in the published experiments, which makes fresh human data a continuing requirement rather than a one-time purchase.
• Three mechanisms produce net-new human data: licensing what was never public, instrumenting proprietary operations, and eliciting data against a specification.
• The three differ most in exclusivity and in how much quality control you can exercise, and that is what should drive the choice.

The ceiling has a measured size

Work from Epoch AI put the stock of publicly available human-generated text at roughly 300 trillion tokens and projected that models will be trained on datasets of comparable size somewhere between 2026 and 2032. Overtraining, which has become standard practice because it produces cheaper inference, pulls that window earlier. Text is the modality the analysis treats as most likely to bind first, because it is the primary input to frontier language models and the easiest to exhaust.

Treat those dates as a projection rather than a deadline, since they depend on assumptions about compute growth and training practice that continue to move. The structural point survives the uncertainty. The public supply is finite, it is measurable, and it is not replenished at the rate models consume it. Every plan that depends on scraped text has an expiry that is now inside a normal planning horizon.

Why substituting generated data is a partial answer

The obvious response is to generate what you can no longer collect, and it works in narrower conditions than the enthusiasm suggests. The same Epoch analysis notes that synthetic data has been shown to reliably improve capability mainly in domains with verifiable answers, such as mathematics and code, where a generator can be checked against a ground truth that already exists.

Outside those conditions, recursive training carries a documented failure mode. Work published in Nature found that indiscriminately training on recursively generated content causes irreversible defects, with the tails of the original content distribution disappearing first. The mechanism matters for planning: what degrades is not average-case performance, which can look stable, but the rare and ambiguous cases that were thinly represented to begin with. Those are the cases production traffic tends to surface.

The same work found that retaining a fraction of real data alongside generated data slowed the collapse. That result is the one with budget consequences, because it reframes human data as an ongoing input rather than a corpus you buy once. HumanSignal's published position on this is that generated data is useful and that generated data still needs a verification layer built from human judgement, and it is worth being precise about where relying on generated data stops being safe for a given task. Evaluation inherits the same constraint, which shows up as measurable differences between benchmarks built on generated inputs and those built on real-world ones.

Three mechanisms that produce net-new human data

Once scraping is no longer the default, the supply routes reduce to three. They are not interchangeable, and most teams end up using more than one.

Licensed corpora that were never public

Archives, publishers, professional bodies, and commercial operators hold text and multimodal records that were never crawlable. Licensing moves data across a legal boundary rather than creating it, which makes it the quickest of the three to deliver volume and the weakest on exclusivity, since the same archive can license to your competitors on similar terms. It also inherits whatever quality and consistency the original holder maintained, and archives assembled for other purposes rarely carry the annotation your task needs.

Proprietary data your operation already generates

Support transcripts, technician notes, inspection records, sensor logs, and internal review decisions accumulate as a byproduct of running a business. This route is the strongest on exclusivity, because nobody else can obtain it, and the most consistently underused, because the data exists in forms nobody designed for training. Realizing it usually means instrumentation and annotation rather than acquisition, and it is worth being clear that what the term human data covers includes these operational records and not only deliberate labeling work.

Elicited data, collected against a specification

The third route creates data that did not previously exist, by designing a collection protocol and paying people to produce against it. This covers expert annotation, preference judgements, on-site capture, and the multimodal recordings that were never on the open web at all. It is the slowest and the most expensive per unit, and the one route where you specify the distribution in advance rather than accepting whatever the source happened to contain. When a model fails on a specific class of input, this is the mechanism that addresses it directly, because collection designed against a specified requirement can target the class rather than hoping coverage improves.

How to choose between the three

Rank the three on the dimension your situation is genuinely constrained by. If you need volume quickly and can tolerate shared access, licensing wins. If your differentiation argument depends on data nobody else can get, the proprietary route is the one that delivers it, and the work is internal rather than procurement. If you have a specific, identified weakness in model behavior, elicited collection is the mechanism that lets you aim at it.

Quality control follows the same ordering in reverse. Licensed data arrives as-is, proprietary data is as good as your instrumentation, and elicited data lets you set the standard before anything is produced. That is also where the operational burden sits, which is why choosing who collects it for you becomes a live decision rather than a procurement detail.

Time to first usable batch runs the other way again. A licensing agreement can deliver a corpus in weeks once the legal work clears. Proprietary data takes as long as the instrumentation it depends on, which is often a quarter or more because it means changing how work is captured at the source. Elicited collection sits between the two on schedule and above both on cost, though it is the one route where a delay buys you something, since the specification improves while the protocol is being designed.

One consequence is worth stating plainly: these routes are complements rather than alternatives. Licensing gives you breadth cheaply, proprietary capture gives you a differentiation argument nobody can copy, and elicited collection gives you the ability to fix a named failure. A team that treats the choice as exclusive usually ends up buying volume it cannot aim.

What this does to a data budget

Two changes follow. The first is that data spending stops being a project cost and becomes a recurring one, because the collapse research implies a continuing real-data fraction rather than a single corpus purchase. The second is that the unit economics invert. Scraped text was effectively free at the margin and unspecified in content. Elicited data is expensive per unit and specified in advance, so the return depends on how precisely you target it.

That inversion changes what a data plan should contain. Instead of a volume target, it needs a statement of which model behaviors are currently weak, which inputs produce them, and which of the three routes can supply those inputs. Evaluation becomes the input to procurement rather than a check that runs afterward, because the specification for the next collection comes out of the last set of failures. Teams that plan for this budget for a smaller volume of data they chose, rather than a larger volume of data they found, and they keep a standing line item for the real-data fraction instead of treating it as a one-time acquisition.

Scope a collection against the data you are missing

If your model is failing on inputs the open web never contained, the useful next step is specifying the distribution you need rather than sourcing more of what you already have. HumanSignal Services designs and runs collection across more than 5 modalities, 30 languages, 50 knowledge domains, and 75 countries, working to a specification your team sets. Book a scoping conversation to talk through what your collection would involve.