This research estimates the value of each class of creative work* that AI is trained on, including journalism, books, code, film and TV scripts and subtitles, and compares it with what AI developers pay for that work.
Creative work is worth an estimated $10bn a year to AI training. Known licensing deals cover about $0.5bn.
Value and payments by creative class
Blogs and web writing
Journalism
Academic and scientific writing
Code
Forums and social media
Encyclopaedic reference
Books
Film and TV scripts and subtitles
We assume training data is worth 10% of what AI developers earn (Reddit–Google article). Creative work is 79% of that data, so creative work is worth about 8% of revenues from AI.
OpenAI and Anthropic now report a combined revenue run-rate of about $117bn a year, an increase of more than 60% since June (Reuters, Bloomberg). Adding an estimate for Google's Gemini (Reddit–Google article) brings this to about $128bn a year.
That puts the value of creative work used to train AI at $10bn a year.
Known licensing deals that cover all creative classes total about $0.5bn a year of that value‡.
How we calculated the value of creative work to AI training
WIPO and UK DCMS creative industries
WebOrganizer format taxonomy
Define the classes
Eight creative classes, plus other web text and AI-generated text
Web crawl sorted by WebOrganizer format
Checked against 998 AI-labelled documents
9.2m pages scanned for scripts
Sort open training data into classes
Blogs and web writing are 29% of Dolma 3
Quality-filter weight per class
Licensed news added to journalism
20,000 simulated mixes
Estimate frontier pre-training mix
A share for each class, with a range. Blogs and web writing fall from 29% to 24%. Creative work is 79% in total.
10% comes from training data
Value each class
$10bn a year across eight classes
Compare with licensing payments
$0.44bn a year
In the absence of live data from AI developers, we use proxies: open training data, the training recipes from research and system cards that open labs publish, and the licensing deals made public.
Europe is changing access to training data. Since August 2025 the EU AI Act has required developers of general-purpose AI to publish a summary of their training content on a standard template (Pinsent Masons).
The summaries published so far give size bands and yes or no flags for each type of source, not shares by class. Few of the largest AI developers have published a summary so far (AI Accountability Lab).
The method note gives every step, range and source we used for this research.
A quarter of pre-training data is blogs and web writing
THE MAKE-UP OF AI TRAINING DATA BY CREATIVE CLASS
Creative work is 79% of pre-training text
Findings for each creative class
- News publishers and social media platforms receive almost all (96%) of AI developers' known licensing payments.
- Blogs and code account for over 30% of AI training data, are estimated to be worth $4.5bn a year, with very few known licensing deals and no disclosed payments.
- Film and TV writing has no licensing deals found, while subtitles of 138,000 films and TV episodes are in training sets used by Apple, Anthropic, Meta and Nvidia (Nieman Lab).
- Journalism is worth $2bn a year and receives $276m, through deals with large publishers such as News Corp.
- Forum posts are worth $1bn a year and receive $149m, with Reddit accounting for almost all of the known licensing payments.
- Encyclopaedic reference is worth $0.5bn a year and receives about $6m. Wikipedia's text is free to use under an open licence, so AI developers pay Wikimedia for faster access to wiki data, not for the right to use it in training (Wikimedia Enterprise).
- Books have no ongoing payment from these developers. One large payment is Anthropic's $1.5bn settlement for pirated copies, which is damages, not a licence (Authors Guild).
A market problem needs a market solution
Today AI developers license mainly from owners of large archives, such as News Corp and Reddit, who can offer a whole catalogue in one deal.
Most creators have no practical way to license their work to AI developers, and developers have no practical way to license from millions of individual creators.
Markets that pay creators at scale run on measurement and data that are transparent to both sides. For example, radio pays songwriters from play counts that stations and collecting societies both trust.
AI requires a similar method, an independent measure of what each class of creative work contributes to AI, which developers, publishers and collecting societies can use to value, attribute and pay creators for their contributions to AI.
The Creative Value Monitor measures the value that creative work contributes to AI. Its aim is that creators are paid for their contribution to AI, and that AI developers get clear prices, legal certainty and continued access to the creative work their systems depend on.
Developers, collecting societies, publishers and trade bodies who want to help build this can contact us at hello@creativevaluemonitor.ai.
Notes
*Creative work here means text used to pre-train AI models: writing, code and the dialogue of film and TV. It does not cover data used to refine models after pre-training, or work that AI systems look up when answering questions. Music, images and video are not covered.
‡AI developers also pay data companies such as Mercor (revenue over $2bn a year, TechCrunch) and Scale AI (revenue over $1bn in 2026, Forbes) for data written by paid experts, mostly used to refine and test models after pre-training, which this research does not count.
†Books and scripts are sold per title, so we show what the market charges per work rather than a yearly value. For books, HarperCollins licenses titles to Microsoft at $5,000 a title over three years. For scripts and subtitles, $515m is a one-off projection: the cost of licensing the 138,000 films and TV episodes in the OpenSubtitles collection once, script by script, to OpenAI, Anthropic and Google, at script option fees of $500 to $2,500 (Final Draft), weighted towards the low end of that range for an average of $1,243 a script. It is indicative only, because an option fee buys the right to purchase a script, not a training licence. Most produced scripts are works made for hire, so this money would go to studios first. The yearly totals count books at $0.4bn and scripts and subtitles at $0.09bn, their value by share of training text.
About the table. Worth is a share of AI developers' yearly revenue if training data accounts for 10% of it, split between classes by their share of training text. Books and scripts are the exception (see below). Yearly values are rounded to the nearest $0.5bn. Ranges for each class are in the method note.
Ongoing and one-off payments. Some classes are licensed every year, because developers need new content as it is published: journalism, forums and social media, and encyclopaedic reference. Others are usually licensed once rather than every year: books, scripts and academic writing. Blogs and code have no priced payment of either kind. Payments a year count ongoing payments only. One-off payments are shown beside them, marked “one-off”, and are not added to the yearly totals or drawn in the bars. One-off payments reported by a single secondary source are listed but not counted.
What payments include. Payments a year include our estimate for deals whose terms are not public, and deals with any AI developer, including Amazon and Microsoft. “Not disclosed” means deals exist but no terms are public. “None found” means we found no deal. Wikipedia's text is free to use under an open licence, but its value is still counted in the $10bn, because it is part of what AI is trained on.
Books and scripts. These are sold per title, so no yearly worth is shown for them. Scripts show a one-off projection instead. Books have no projection, and their per-title price is in the note (see †). Their bars are left blank because these figures are not yearly. Books show Anthropic's one-off $1.5bn settlement as a payment. The settlement is damages for pirated copies, not a licence. Academic writing is licensed once but priced as a catalogue, not per work, so its worth stays yearly.