Method: Creative Work in AI Training Data

1. The question

What share of the text that frontier AI models are trained on is journalism, books, academic writing, reference, forum posts, blog writing, code, or film and TV scripts and subtitles? What is each class worth to the AI developers who train on it? And what do known licensing deals pay for it?

Scope, for every figure. Text used to pre-train AI models, with no music, images or video. The frontier models (GPT, Claude and Gemini) as a group, not any one lab. Whole creative classes, not any single rights holder's catalogue.

Framing. No frontier lab discloses what its training data contains. Every frontier figure here is an estimate built from public evidence, with its range shown. None is a claim about any lab's actual data.

Terms. A token is a short chunk of text, about three quarters of a word (OpenAI: "1 token is approximately 4 characters or 0.75 words for English text"). A range is the spread of results across our assumptions and filing choices. It is not a statistical confidence interval.

The method has five steps, numbered as on the results page. Step 0 defines the classes. Steps 1 and 2 give each class its share of pre-training text. Steps 3 and 4 turn shares into money and compare them with known licensing payments.

2. Step 0: define the classes

There are eight creative classes: blogs and web writing, journalism, academic and scientific writing, code, forums and social media, encyclopaedic reference, books, and film and TV scripts and subtitles.

Film and TV scripts and subtitles covers scripts and screenplays for film, TV and radio, subtitles and captions of film and TV, and transcripts of film, TV and radio programmes. Reviews, synopses, fan commentary and news about films stay in their other classes. Transcripts of news programmes, podcasts and online video are recorded separately and kept out of the core class. They appear only at the top of the class's range.

Two further lines are not creative classes. Other web text is text that cannot be confidently put in a class, such as product pages, listings and boilerplate. AI-generated text has no human author, so it gets a share of the training text and no value. AI-generated scripts count as AI-generated text.

Each class is anchored on two sides. On the legal and economic side, it follows the categories of copyright works and the UK DCMS and WIPO classifications of the creative industries. Film and TV scripts follow dramatic works in copyright law (UK Copyright, Designs and Patents Act 1988, s.3; US 17 U.S.C. §102(a)(3)) and the DCMS sub-sector "Film, TV, radio and photography". On the corpus side, each class follows the WebOrganizer format taxonomy (Wettig et al., 2025). Forums, social media and blogs extend the industrial classifications on purpose. They are copyright-protected literary works written outside firms.

3. Step 1: sort open training data into classes

The starting point is AI2's Dolma 3 pool, a training collection published in full, 9.31 trillion tokens by its dataset card.

Counted. We counted the named sources in Dolma 3 in full: academic PDFs and arXiv papers, Wikipedia and Wikibooks, code (StackEdu) and maths (FineMath). The dataset cards for the Dolma 3 pool and the Dolma 3 training mix list no script, subtitle or transcript source.

Estimated. The web crawl is 8.14 trillion tokens, 87.7% of the pool. We did not classify the crawl ourselves. We estimated it from WebOrganizer's published token shares by format, mapped onto our classes:

Film and TV scripts and subtitles in the crawl come out of Audio Transcript (other web text) and Creative Writing (blogs). Section 4 gives the amounts.

Checked against a sample, and labelled by AI. We drew 998 documents at random from Dolma 3 and counted their tokens exactly. Claude Opus classified every document. For 499 of them, a second Claude Opus pass checked each label. For the other 499, Claude Opus at a higher effort setting re-labelled every document it was unsure of. No person labelled the sample. Here is how the published breakdown compared with the sample, as shares of the whole pool:

We could not settle the three disagreements. Every range on the results page therefore covers both views, together with the Nonfiction Writing filing choice, which moves journalism between 7.9% and 16.7% of the pool.

Result, shares of the pool (tokens and documents).

4. Step 1, continued: counting film and TV scripts and subtitles

Scripts and subtitles are too rare for the 998-document sample to measure. None of the 998 is a script, subtitle file or programme transcript. Six are close: four podcast or online-video transcripts and two news-programme transcripts.

A larger scan. We split the crawl into two groups by Dolma 3's own topic folders. Group A is the entertainment and literature topics (11.6% of crawl tokens). Group B is the other 22 topics. We drew 120 parts (shards) of each group at random, with chance proportional to size, and read every document in them. That came to 9,235,335 documents and 8.86 billion tokens.

Three format checks. A program flagged documents that look like:

• a script: two or more scene headings (lines starting INT. or EXT.), or one scene heading or a FADE IN: or CUT TO: line with five or more character names above dialogue;
• a subtitle file: three or more timing lines (00:01:02,345 --> 00:01:04,567), or a WEBVTT header;
• a transcript: six or more lines starting with a speaker's name, with at least two speakers repeating, plus a programme word (transcript, episode, broadcast, radio, BBC, NPR, season) or a production cue such as (LAUGHTER) or [MUSIC]; or three or more production cues on their own.

AI confirmation. Claude Opus, with the same instructions as the 998, classified every flagged script and subtitle document (81), a random 200 flagged transcripts from each group (400 of 51,760), and a random 150 from each group of documents that were not flagged but mention a screenplay, teleplay, subtitles, a transcript or an episode (300). No person checked these labels.

Two ways to count. We fixed a rule before counting: a check that is right less than 30% of the time counts only the documents the AI confirmed. Under that rule, the class is 0.0011% of crawl tokens (range 0.0005% to 0.0019%). That figure is a floor, because it ignores the flagged transcripts the AI never saw. Applying every check's confirmation rate to everything it flagged gives 0.061% (range 0.018% to 0.12%). We report both and carry the span from 0.0005% to 0.12%. Adding news-programme and podcast transcripts lifts the top to 0.63%.

Where the confirmed documents came from. Of the confirmed text, 62% sat in WebOrganizer's Audio Transcript format (filed in other web text) and 38% in Creative Writing (filed in blogs).

5. Step 2: estimate the frontier pre-training mix

The frontier mix is inferred from the open collection in two layers plus an overlay. A simulation of 20,000 draws, with fixed seeds, carries the uncertainty through.

Layer 1: how tokens split across sources. Each draw samples a share for each source from a range, then rescales them to sum to 100. The ranges bracket what partly open labs disclose: Llama 1, The Pile, RedPajama, Falcon, Dolma 1.6 and 1.7, OLMo 2 and the Dolma 3 training mix. AI2 publishes both the Dolma 3 pool and the mix it trained on. That makes one move from pool to mix directly measurable: code rises 4.6 times, maths 6.4 times, academic 1.4 times, and web falls to 0.87 times.

The subtitle range runs from none, as in every disclosed mix but one, to The Pile's 1.55% for OpenSubtitles (Gao et al., 2020, Table 1). The Pile's YouTube Subtitles (0.60%) is online video, so it is left out.

Layer 2: what is inside the web block. No lab says. The web interior comes from step 1. Frontier quality filtering is applied as a survival weight per class, sampled over a range. Blogs are fixed at 1.0. Other web text survives at 0.25 to 0.55, journalism at 0.90 to 1.40, forums at 0.70 to 1.00, and scripts and subtitles at 0.50 to 1.00. Encyclopaedic and academic survive at 1.20 to 1.70. The range for scripts and subtitles is our judgement. Dialogue and subtitle text is short and informal, which quality filters tend to mark down, and no source measures it.

Overlay. Books enter as the curated books slice. Licensed news is added to journalism.

Three settings. The simulation runs with the web interior from the published breakdown, with Nonfiction Writing filed as blogs, and from our 998 sample. Each runs with scripts and subtitles at the core and the wide definition. A class's range runs from the lowest 2.5th percentile to the highest 97.5th percentile across the six runs. The central figure is from the first setting with the core definition.

Result. Share of all pre-training text, including AI-generated text at 5% to 15%:

The eight creative classes are 79% of pre-training text.

6. Step 3: value each class

The value step uses the top-down chain published in our article What the Reddit–Google Deal Suggests Creative Work Is Worth to AI. Here it runs across every class, counted in tokens.

Value a year is the revenue pool x training data's share of model value x the class's share of pre-training text.

• Revenue pool. OpenAI's run rate of about USD 50 billion (Bloomberg, 8 October 2026), plus Anthropic's third-quarter revenue of USD 16.8 billion, USD 67.2 billion a year (Reuters, 14 October 2026), plus Gemini's attributable revenue, sampled at USD 8 to 13 billion as in the Reddit–Google article. The median pool is USD 127.7 billion a year. Run rates annualise a recent month or quarter, so they show current scale. They are not forecasts.
• Data share. Nobody has measured it. We use 10%, the anchor in the Reddit–Google article, and also show 2% and 30%. 2% is close to BMI's radio licence rate of 2.2% of gross revenue. 30% is the top of the article's range. The sources behind 10% (Ho et al., 2024; Erdil and Besiroglu, 2022) measure what drives model improvement, not what share of revenue data earns.
• Class share. The share of pre-training text from step 2. AI-generated text takes its share of the text and gets no value.

Value here is a royalty-style share of revenue, not of profit.

Value a year at a 10% data share:

The results page rounds each value to the nearest USD 0.5 billion. Other web text is human-written but not creative work, so it is not in the USD 10.1 billion. The total is the sum of the class figures.

The eight classes together are worth about USD 2.0 billion, 10.1 billion or 30.4 billion a year at 2%, 10% or 30%. That is 1.6%, 7.9% or 23.8% of the revenue pool.

Unit value is the value of a class ÷ its training tokens. Training tokens are the class share x one frontier training run (14.8 to 36 trillion tokens, from the DeepSeek-V3, Llama 3, Kimi K2, Llama 4 and Qwen3 reports) x three developers. The chain splits value in proportion to tokens, so the unit value is the same for every class: about USD 167 per million tokens a year at a 10% data share (119 to 277), and one fifth or three times that at 2% or 30%.

Books priced per title. Books are sold per title, so the results page shows no yearly value for them. HarperCollins licenses titles to Microsoft at USD 5,000 a title over three years. Anthropic's settlement paid about USD 3,000 a book. Per token, those prices are 72 to 176 times the unit value. US copyright law helps explain the gap: statutory damages are set per registered work, from USD 750 (17 U.S.C. §504(c) and §412).

Film and TV scripts priced per script. As with books, we also price scripts per work. The count is about 138,000 works: The Atlantic found subtitles from about 53,000 films and 85,000 TV episodes in a training dataset (18 November 2024, read through The A.V. Club's report of the same day). The Pile's datasheet lists 446,612 OpenSubtitles files (Biderman, Bicheno and Gao, 2022), about 3.2 files per title, which fits that count. No training price for a script exists, so we use option fees for scripts by unsold writers as a guide, USD 500 to 2,500. An option buys exclusive production rights for a period, which is more than a training licence, so this range is more likely high than low. Few scripts command the top price, so we weight the range towards USD 500. Two spreads give similar averages: equal odds for each doubling of price (USD 1,243) and most common at USD 500, thinning to USD 2,500 (USD 1,167). We use USD 1,243. At that price, 138,000 works cost one developer about USD 172 million and three developers about USD 515 million (range USD 207 million to 1.04 billion), paid once, like the book deals. Most produced scripts are works made for hire, so this money would go to studios first.

The yearly totals count books and scripts by their share of training text, USD 0.40 billion and USD 0.09 billion.

7. Step 4: compare with licensing payments

Payments a year count recurring payments from known licensing deals with any AI developer: a per-year figure, a multi-year licence or a subscription. One-off payments and settlements for past use are listed separately and never annualised.

• Multi-year deals are divided by their term. "More than X" and "at least X" are read as X. "Up to X" is sampled from half of X to X.
• A deal that buys both training data and live display carries a training share of 25% to 100%.
• Deals with no public price are allowed for by scaling the priced sum by between 1 and the ratio of known deals to priced deals in the class.
• A one-off payment reported by a single secondary source is listed but not counted.

"Not disclosed" means deals exist but no terms are public. For books, HarperCollins' price per title is public but the number of titles is not. "None found" means we found no deal. Stack Overflow's deals are counted under forums. Wikipedia's text is free to use under an open licence, so payments to Wikimedia Enterprise buy faster access to the data, not permission to use it.

Known licensing payments are USD 0.44 billion a year, 4% of the USD 10.1 billion value at a 10% data share, and 0.3% of the revenue pool. The payments do not depend on the data share.

Film and TV scripts and subtitles. We found no recurring or one-off payment by an AI developer for script, subtitle or transcript text. What we found is context, read at source on 12 October 2026: Lionsgate's video-model partnership with Runway (September 2024, no price public); Google's approaches to the major studios over their intellectual property (Los Angeles Times, 31 August 2026, "no agreements have been reached yet"); the WGA's call for payment for training (Variety, 10 March 2026) and its 2026 contract term giving notice and a meeting over "remuneration, if any"; and 2024 reporting by Proof News and The Atlantic that YouTube and film and TV subtitles were used to train models without licences. The Atlantic article could not be read directly. Its content comes from The A.V. Club's summary.

Not counted. AI developers also pay data companies such as Mercor and Scale AI for new writing by paid experts. That data is mostly used to refine and test models after pre-training, so it is outside this research.

8. How the figures were checked

The checks and their pass marks were written down before any figure was computed. Every change was logged. The checks were run by AI tools. No person outside the project has reviewed the work.

9. What the method cannot yet say

1. It cannot see inside any lab. No frontier lab discloses its training mix. The frontier shares rest on judgement ranges for how much quality filtering strips each class, how far the frontier leans on curated sources, and how much licensed news and subtitle text is added. The share for film and TV scripts and subtitles comes almost entirely from the subtitle range.
2. The data share decides the value. Moving it from 2% to 30% multiplies every value by 15. The make-up of the training text changes the total far less.
3. It values every token alike. Markets do not. Books are priced per work, at 72 to 176 times the chain's unit value at a 10% data share. So the chain sizes the pool. It does not say how to divide it between a book, a script and a forum post.
4. The labels are AI labels. No person has checked a sample of them, so their error rate is unknown. Labelling the 998 a second time changed 11.5% of labels, so on those documents at least one of the two AI runs is wrong.
5. Payments are a floor with an estimate on top. Most deals are undisclosed. Journalism scales four priced deals up to 25. If the deals with a public price are larger than the rest, the journalism figure is too high.
6. Openly licensed text is not separated. Wikipedia is free to use under CC BY-SA. Wikimedia Enterprise sells a faster feed, not permission.
7. Academic writing includes maths datasets. Taking maths out lowers academic's share.
8. Revenue moves quickly. Every value scales with the revenue pool, which uses run rates from October 2026. The composition and payment figures do not depend on it.

10. Reproducing the figures

This note, the mapping table in section 3 and the sources below are enough to rebuild the composition and value figures. The count of scripts and subtitles needs the scan itself, which reads public Dolma 3 files. Our scripts, input files and document-level labels are not published.