Four community archives are licensed on the public record: Reddit, Stack Overflow, Automattic's WordPress.com and Tumblr, and Wikimedia Enterprise. None of their announcements states a price.
Forum archives are one text type within AI training data licensing. No party states a per-post or per-buyer price for a forum archive, and the main figure Reddit states is an aggregate contract value of USD 203.0 million from its S-1 of 2024-02-22, for counterparties it does not name. Not licensing is a valid outcome for an operator.
Platform statements were checked on 2026-11-09 unless stated otherwise; how we source, date and check every figure explains the method.
What a Forum Archive Holds Beyond the Text of the Posts
A forum archive holds the posts plus the structure the community added to them: who answered whom, which answer the community voted up, which tags it carries and how it was edited over time.
The table compares the elements that licensors and one dataset paper name, as of 2026-11-16, with the licensors' wording checked 2026-11-09 and the dataset paper from 2019.
| Element of the archive | What it records | Where a licensor or dataset names it | Source |
|---|---|---|---|
| Posts and comments | the text and the reply structure of a thread | "real-time, structured" content through the Reddit Data API (Google, 2024-02-22) | blog.google |
| Question-and-answer pairs | a question with its answers | "83M+ Q&A" from 150+ Stack Exchange sites, Stack Overflow states | stackoverflow.co |
| Tags | the topic labels the community assigned | named among licensed fields (Stack Overflow) | stackoverflow.co |
| Votes and karma | the community's quality judgment | votes named (Stack Overflow); WebText kept outbound Reddit links with at least 3 karma (OpenAI, 2019) | stackoverflow.co; cdn.openai.com |
| Revisions | how a post was edited over time | named among licensed fields (Stack Overflow) | stackoverflow.co |
| Community-written reference articles | collaboratively edited articles | more than 65 million articles in over 300 languages, Wikimedia states | enterprise.wikimedia.com |
Forum text is one of the data types AI companies buy, and it differs from a document set because the community's own judgments travel with it.
One dataset shows the idea in use: the 2019 GPT-2 paper describes WebText, built from outbound Reddit links with at least 3 karma. That is a historical selection rule for one dataset, not a statement about how any model is trained today.
Community Archives Licensed to AI Developers, by Holder and Delivery
Four community archives are licensed to AI developers on the public record, and the table lists each by holder, the AI developer named, the delivery and the party's own wording. It is as of 2026-11-16, every value is the party's own statement or filing checked 2026-11-09, and the rows run alphabetically by holder.
| Holder (legal party) | AI developer named | Archive and delivery | Use (party's words) | Condition stated | Value as stated | Source |
|---|---|---|---|---|---|---|
| Automattic Inc. (WordPress.com, Tumblr) | select AI companies, not named | public posts from hosted sites that have not opted out (2024-02-27) | not stated | AI companies updated about new opt-outs and asked to remove content from past sources | not disclosed | automattic.com |
| Reddit, Inc. | Google LLC (2024-02-22) | Reddit Data API, real-time structured content | "display, train on, and otherwise use" (Google) | none stated | not disclosed by the parties | blog.google |
| Reddit, Inc. | OpenAI (2024-05-16) | Reddit Data API, content for ChatGPT and new products | training not stated | none stated | not disclosed | redditinc.com |
| Reddit, Inc. | not named (January 2024 arrangements) | to "access, search, and analyze data on our platform" | not stated | terms of two to three years | USD 203.0 million aggregate contract value (S-1, 2024-02-22) | sec.gov |
| Stack Exchange, Inc. (Stack Overflow) | Google Cloud (2024-02-29) | technical knowledge surfaced in the Google Cloud console | not stated | none stated | not disclosed | stackoverflow.co |
| Stack Exchange, Inc. (Stack Overflow) | OpenAI (2024-05-06) | OverflowAPI; attribution in ChatGPT | collaborate to improve model performance (Stack Overflow) | attribution expected | not disclosed | stackoverflow.co |
| Wikimedia, LLC (Wikimedia Enterprise) | Amazon, Meta, Microsoft, Mistral AI, Perplexity (added 2026-01-15) | paid Enterprise API: on-demand, snapshot, realtime | not stated | no pricing published | not disclosed | enterprise.wikimedia.com |
"Not stated" in the use column is not a finding that training is forbidden, and no row states whether the license is exclusive.
The platform's own decision to license, its filings and its opt-out controls belong to the page on platforms and online communities licensing user content.
Continuing Access Versus a One-Off Copy: How Forum Data Is Delivered
Forum data is mostly licensed as access to a live feed, because a community keeps writing, so the buyer pays for continuity as well as for the archive.
Stack Overflow states continuous access for training and fine-tuning, with attribution expected and a free sample of 1,000 Q&A pairs, and Wikimedia Enterprise offers on-demand, snapshot and realtime modes.
Reddit's filings fit the access reading. Its 10-K for fiscal 2025 states that licensing revenue is recognized generally ratably over the license period, because the license is a right to access, and its S-1 gives terms of two to three years.
In its 10-Q for the quarter ended 2026-09-30, Reddit states that substantially all licensing contract value comes from two counterparties whose agreements are approaching the end of their current terms; no renewal or end is on the record as of 2026-11-09, and this page makes no forecast.
Reddit's values and renewal are covered under Reddit's AI data licensing deals with Google and OpenAI. Its shareholder letter of 2026-10-29 frames logged-in-only tests outside the U.S. as partly protecting access to its data corpus.
Who Holds the Rights in a Community Archive: Users, Operators and Open Licenses
Users write the posts and usually keep their copyright, and an operator can license the archive only as far as the rights its terms of service take from users, any open license on the posts, and its own platform terms allow. The operator's current terms decide.
Posts under an open license follow that license. Creative Commons states, checked 2026-11-09, that its licenses apply only where copyright permission is required, and that under a cautious reading content under a NoDerivatives license is not used as training data.
A hosting service's own terms can cut the other way. Discord's terms, effective 2025-09-29, prohibit in section 9 "selling, licensing, or otherwise commercializing content or data obtained from the services."
FTC staff wrote on 2024-02-13 that a retroactive or surreptitious change to adopt more permissive data practices may be unfair or deceptive (a staff blog post, not a rule). Reddit disclosed in its S-1/A an FTC letter of non-public inquiry of 2024-03-14 about licensing user content to train AI, with no outcome on the record.
Before an operator offers an archive, a rights audit before licensing an archive separates what users granted from what the operator may sublicense.
Personal Data in Public Threads and Private Messages
Public threads name and describe people, so in the EU licensing them for training is a processing of personal data that needs a legal basis, and private messages sit further out of reach. This section is information, not legal advice.
The EDPB adopted Opinion 28/2024 on 2024-12-17. It assesses legitimate interest by a three-step test, and it treats a model as anonymous only if the likelihood of extraction and of obtaining personal data from queries is insignificant.
The Irish Data Protection Commission wrote in its AI Insights Report of September 2026 that it is "unlikely that a controller will be able to demonstrate that data subjects expect the usage of their private messages or other private data for AI training", a position that weighs against putting direct messages into a forum dataset.
Usernames, signatures and quoted replies carry identifiers into a thread, and PII removal before licensing covers removing usernames and identifiers. The EDPB opinion adds that a corrective measure may include erasure of the dataset or model (para. 114).
Forum Pages Already Crawled: Reservations, Removal Limits and Access Suits
Public forum pages can be collected by crawlers without any license, so an operator's controls are crawler rules, a machine-readable reservation in the EU, access limits and, for Reddit, lawsuits.
Common Crawl is the upstream source of C4, FineWeb, Dolma and RedPajama, and GPTBot is OpenAI's training crawler, while OAI-SearchBot and ChatGPT-User are not (OpenAI documentation, checked 2026-11-09).
In the EU, Article 4(3) of Directive (EU) 2019/790 lets a rightsholder reserve its rights "in an appropriate manner, such as machine-readable means in the case of content made publicly available online".
Article 53(1)(c) of Regulation (EU) 2024/1689 requires providers of general-purpose AI models to keep a copyright policy that complies with such reservations, and those obligations apply from 2025-08-02.
A reservation reaches later collection, not earlier copies. Common Crawl states, on 2025-11-04, that its WARC files are immutable and that affected URLs are removed or filtered only from subsequent crawls and public indexes.
Reddit asserts contract claims against crawlers in court. As of 2026-11-09, Reddit v. Anthropic (San Francisco County Superior Court, CGC-25-625892) is pending after remand on 2026-03-30 and, according to Anthropic's S-1/A, narrowed in September 2026 to contract, interference and unfair competition claims, and these are Reddit's allegations.
In S.D.N.Y. 1:25-cv-08736-PAE, filed 2025-10-22, the court predominantly denied the motions to dismiss of SerpApi and Perplexity on 2026-07-31, and the Oxylabs motion was undecided as of 2026-11-06.
Community Chat Servers and Company-Run Forums
A company's own community, such as a Discord server or a support forum it runs, is a different holder from a public platform, and the chat platform's terms can bar licensing what was posted there. Two intermediaries name Discord as a source, while Discord's terms bar commercializing its content, and both statements stand.
Turing's Project Lazarus lists Discord among the chat systems it takes, on a page re-read on 2026-11-16, and Turing sells the record to AI labs on a revenue share. Mercor's Game Studio listing, checked 2026-11-09, names Discord among tools. No license of a small, self-hosted or company-run forum archive is on the record.
Chat archives as a type are covered under chat archives licensed for AI training, including company workspaces.
Licensing a Forum Archive: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
Licensing a forum archive can turn years of community writing into multi-year contract revenue, but the operator licenses what users wrote, the content is often already crawled, and a model trained on it cannot be untrained. The four lists below set out both sides.
Benefits
- Multi-year terms: Reddit's S-1 describes terms of two to three years.
- Attribution: Stack Overflow states it expects attribution.
- A consenting subset: Automattic shares only content from sites that have not opted out.
Risks
- A terms change made after the fact may be unfair or deceptive, per FTC staff.
- Reddit disclosed an FTC non-public inquiry into licensing user content, with no outcome on the record.
- Personal data in threads brings EU legal-basis duties.
- Unlicensed collection undercuts the asset, as Reddit's suits allege.
Downsides
- Concentration: Reddit states substantially all licensing contract value comes from two counterparties.
- No per-buyer price exists to benchmark a forum archive.
- Past crawls cannot be recalled, since Common Crawl removes only from later crawls.
Alternatives
- License an opted-in subset, as Automattic describes.
- Restrict crawlers and logged-out access, and treat GPTBot separately from search bots.
- Reserve rights under Article 4(3) of the DSM Directive.
- Choose not licensing at all and keep the archive closed.
Each option is a legitimate outcome.
Forum Archives Beside Text Corpora, Platform Sellers and the Seller Steps
This page covers forum and community archives as data; text corpora in general, the platform's licensing decision and individual deals each have their own page. Neighboring pages:
- robots.txt rules for AI crawlers
- text corpora for AI training
- news archives as AI training data
- AI data licensing deals
- Stack Overflow AI licensing deals
- Automattic (Tumblr, WordPress.com) data licensing
- Wikimedia Enterprise AI deals
- Discord server data and AI licensing
- Reddit v. Perplexity, SerpApi and others
- services that train on their own users
- opting out of AI training
- OpenWebText
- EU text and data mining opt-out
The steps an operator follows to license an archive:
- Read how to sell or license company data for the full process.
- Check GDPR and AI training data for legal basis and notices.
- Review AI data licensing agreements for grant, attribution and deletion clauses.
- Compare written terms from several AI data buyers in parallel.
Frequently Asked Questions
How much is forum data worth to AI companies?
No per-post or per-buyer price for forum data is on record. Reddit's USD 203.0 million is an aggregate contract value for unnamed January 2024 arrangements (S-1, 2024-02-22), not a payment for one archive. See what company data is worth to AI companies.
Can a small forum operator license its archive for AI training?
Only as far as its user terms grant it rights. No license by a small or self-hosted forum is on the record, so the operator's own terms and its users' rights are the starting point.
Is Stack Overflow data licensed for AI training?
Stack Overflow states continuous access for training and fine-tuning, with attribution expected (checked 2026-11-09). It announced a collaboration with OpenAI on 2024-05-06, and no value is disclosed.
Can deleted posts be removed from data already licensed or crawled?
Not always. Automattic asks the AI companies to remove newly opted-out content from past sources, while Common Crawl states its files are immutable and removal reaches only later crawls (2025-11-04).