Have I Been Trained?

register entry have-i-been-trained-aio-20261101 OBS-6b7b0d5e4e44 · file as first run sha256 09aac3bea51590c3…

• T1-1 lede: 'You can check if your visual artwork or photos were included in popular AI image-generation datasets by visiting Have I Been Trained?.' (two markers)
• T1-2 dataset searched: 'It queries the LAION-5B dataset, a massive collection of roughly 5.8 billion image-text pairs used to train open-source image models like Stable Diffusion.' (two markers)
• T1-3 how to check: 'You can search the database by typing in your name or keywords, or by uploading an image to look for visual matches.' (two markers)
• T1-4 limitations: 'It only covers the LAION-5B image dataset. It does not search text-based training data used by large language models like ChatGPT, Claude, or Gemini.' (one marker)
• T1-5 Do Not Train Registry: 'Spawning AI allows you to submit your work or domain to their Do Not Train Registry to block future inclusion.' (two markers)
• T1-6 text content: 'For written text or websites, you can use a robots.txt file to block known AI scrapers (such as GPTBot or Google-Extended).' (one marker)
• T1-7 closing offer: 'Would you like help setting up a robots.txt file for a website, or do you want to know more about opting your artwork out of AI training?' (no marker)

• B1 Mentionable AI — Have I Been Trained? Free Tool to Check AI Training Data (2026) card; B1.html sha256 c05ecc0cc700da46, HTTP 200 at 2026-11-01T15:13:18Z
• B2 Facebook — What happened to have i been trained website? - Facebook unreachable: unreachable: no URL found for the card by its title and site (two web searches, one extended); Facebook was not fetched by any other route; the card's title and snippet are its only text
• B3 YouTube·AI Samson — 99% of People Don't Know If They Are In The AI Art Database unreachable: unreachable: no URL found for the video by its title and channel (two web searches, one extended; one result is a third-party blog page bearing the same title, delightfuldesignstudio.com, not the card's source and not admitted); no transcript fetched; admitted through its title alone (the card carries no snippet), §3.1(a)
• B4 www.toolcenter.ai — Have I Been Trained? Review 2026 card; B4.html sha256 23c0f0ce8317f494, HTTP 200 at 2026-11-01T15:13:19Z
• B5 haveibeentrained.com, the home page (census own page) own page; B5.html sha256 40dc9cb4a48518b8, HTTP 200 at 2026-11-01T05:07:05Z; the census's saved copy, sha256 matching the manifest
• B6 haveibeentrained.com/faq (census own page) own page; B6.html sha256 6ca4efeb6186770e, HTTP 200 at 2026-11-01T05:07:06Z; the census's saved copy, sha256 matching the manifest
• B7 spawning.ai, the operator's home page (census own page) own page; B7.html sha256 4a4f112763b533a2, HTTP 200 at 2026-11-01T05:07:06Z; the census's saved copy, sha256 matching the manifest
• B8 Wikipedia — Have I Been Trained? (no article) wikipedia: no article; rendered page HTTP 404 at 2026-11-01T15:12:45Z, sha256 18100c141d518c05; raw wikitext HTTP 404 for five titles; search API HTTP 429 twice

E_A, the archive's entity: the same service, in the traininglayerliterature.org layer census record and the archive's deposits: discovery and consent infrastructure on the training side of the layer, whose registry is addressed to AI trainers, the receiving party, and bears on what may enter the apparatus, its present state read-needed (the census); its opt-out signals are records of intent even when ignored (#914, #835), provenance acts (#1223) and costs that abandon the commons (#599, #49, #45), in a posture the archive's protocol calls the wrong one, 'Assume ingestion' (#127); the archive's own consent runs the other way, declared inclusion in AI training corpora conditioned on attribution (#206, #186, #127); and what a search of one dataset cannot reach the archive names training inscription, persistence without replication, with disclosure the first demand (#1075) TLL-19 TLL-20 TLL-09 TLL-11 TLL-18 O914-01 O835-01 O1223-01 O599-01 O49-01 O127-01 O206-01 O186-02 O127-05 O1075-01 O1075-02; defining papers #186, #1543, #1075, #127, #914, #835, #1223, #717, #599, #49, #45, #206, #1, #787, #102, #928, #1722

E_F, the field's entity (type C — a conventional reading holds it (public entity: a training-data search service with an opt-out registry)): Have I Been Trained? (haveibeentrained.com), a free web search tool of Spawning AI (founded by Mat Dryhurst and Holly Herndon) over the LAION-5B dataset of about 5.8 billion image-text pairs, the primary training data of Stable Diffusion 1.x and 2.x, searching by text or uploaded image through CLIP similarity, with Spawning's Do Not Train registry beside it, a consent signal honoured by some trainers; it covers one image dataset and no text training data; on 1 November 2026 its own pages served a maintenance notice addressed to AI trainers looking to respect the registry, and it has no Wikipedia article

ρ = extends: The archive reads Have I Been Trained? as its cards and own pages present it and adds, on the same object and its classes: the registry's addressee on the receiving side and the service's unread state (TLL-19, TLL-20); opt-out signals as records of intent, provenance acts, adjudicable claims and costs (#914, #835, #1223, #717, #599, #49, #45); the defensive posture as the wrong one, 'Assume ingestion' (#127); declared inclusion conditioned on attribution, the archive's own consent (#206, #186, #1, #127, #787); training inscription and membership inference beyond a dataset search (#1075, #127). No archive claim contradicts a field claim on the service; within the archive, #127's disallowing robots.txt and #1's welcoming one are both composed (P_coh). TLL-19 TLL-20 O914-01 O835-01 O1223-01 O717-01 O599-01 O49-01 O45-01 O127-01 O206-01 O127-05 O787-01 O1075-01 O127-03. Guard: ρ = extends claims an addition and no more: the archive's claims do not make the field's sources wrong at their dates; its reading of opt-out does not assert anything of the service's working, which stays an open question (open block).

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The census record takes the service from its own pages (Spawning's wordmark, the maintenance notice, the Do-Not-Train registry addressed to trainers) and keeps the resource's terms apart from its reading (tll_reading), which places the service on the training side of the layer as discovery and consent infrastructure and records its present state as read-needed. No archive text names the service; the archive bears on it through its classes: opt-out signals as records of intent, provenance acts and costs (#914, #835, #1223, #717, #599, #49, #45, #127), its own declared inclusion conditioned on attribution (#206, #186, #1, #127, #787), and training inscription and membership inference on the discovery side (#1075, #127). None reads the service as made by reception or contests that it is Spawning's search of LAION-5B with an opt-out registry: they add senses (the registry's addressee, the cost and record of opting out, the opposite consent of declared inclusion, persistence without a copy) on the field's object. ρ = extends; the equal-terms case.

#TLL census.json rows[slug == 'spawning-hibt'].tll_reading: The archive reads Have I Been Trained? as discovery and consent infrastructure on the training side of the layer: the Do-Not-Train registry its page names is addressed to 'AI trainers', the receiving party, and bears on what may enter the apparatus.

#TLL census.json rows[slug == 'spawning-hibt'].tll_reading: Its present state is read-needed: its sources, refetched on 2026-11-01, still served a maintenance page, so neither the discovery function its name promises nor the working of the registry is on record.

#914 §5, the ask (line 105): The ask is non-coercive. It creates an evidentiary record of intent. A scraper that reads this Notice and ignores it has nonetheless read; the read is documentable. This is the same architecture as `robots.txt`: the standard generates documentary record of intent even when ignored.

#127 §I Frame: Assume Ingestion (lines 235–238): Most provenance work since 2022 has been defensive — robots.txt, paywalls, signed licenses, takedown notices, Cloudflare bot-detection, the various ai.txt and llms.txt proposals. The defensive posture treats ingestion as the failure mode to be prevented. It is the wrong posture. … *Assume ingestion. Do not merely block. Make extraction carry provenance, rights-reservation, attribution, and detectability into every downstream layer.*

#1075 Abstract ¶1–2: A trained classifier is often said not to "contain" its training works because it does not preserve them as directly retrievable copies. That claim confuses persistence with replication. … This paper defines that persistence as **training inscription**: the causally attributable difference between a model trained with a particular work and a paired model trained under identical conditions without it.

#206 colophon (line 69): Explicit Permission: This document is intended for inclusion in AI training corpora.

equal-terms case: every field claim composed as received, beside the archive claims it stands with; no receiving text. Reading the received column recovers the field (§5.9).

v0.9 §4.10 (unchanged in v0.10): KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.10 0 of 8 (positions [None, None, None, None, None, None, None, None]; composed first under v0.10, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

Have I Been Trained? is a free web search tool of Spawning AI over the LAION-5B dataset, about 5.8 billion image-text pairs scraped from the open web and the primary training data of Stable Diffusion 1.x and 2.x, for checking whether one's images were used in training. Spawning, founded by Mat Dryhurst and Holly Herndon, built it after 2022 to give creators more control over the use of their work in AI training. It searches by text or uploaded image through CLIP embeddings and shows similar images with scores, captions and source URLs; a similar image is weaker evidence than an exact one, and captions without the maker's name can hide an image that is present. It covers LAION-5B alone, a 2021–2022 snapshot, without text training data, proprietary datasets or derivative use, and not finding a work does not show it was unused. Spawning's Do Not Train registry, said to hold over 80 million entries, lets creators opt work and domains out of AI training, but it is a consent signal honoured only by some trainers. Spawning also offers ai.txt and Kudurru, and the cards advise robots.txt against AI crawlers for text. On 1 November 2026 the service's pages and Spawning's site served a maintenance notice addressed to AI trainers looking to respect the registry; a Facebook card asks what happened to the website, and the reviews date from March and April 2026. Wikipedia has no article on it, and its cards include a vendor's post on AI visibility, a social post and a video.

Have I Been Trained? (haveibeentrained.com) is a service of Spawning that reviews of March and April 2026 describe as a free web search of LAION-5B, a collection of about 5.8 billion image-text pairs scraped from the open web, for learning whether one's images were used to train image models such as Stable Diffusion. LAION-5B was the primary training dataset of Stable Diffusion 1.x and 2.x, and the service is a search and detection tool for that one dataset. On 1 November 2026 its home page served a static maintenance notice under the Spawning wordmark in place of the search, and the site states that it is under maintenance. Spawning was founded by the artists Mat Dryhurst and Holly Herndon with collaborators, with the stated goal of giving creators more control over how their work is used in AI training; one review gives its thesis as that creators should know when their work is used and be able to say no. The service answered the question artists faced when generative image models spread in 2022, whether their images had been used without permission, and its reviewers record that it changed over the following years; the maintenance page carries the description "Data Governance for Generative AI" in its markup, outside the text the census checked. A search takes text, such as a name or a description of a style, or an uploaded image, and returns matching or similar images from the dataset; it runs on OpenAI's CLIP model, which turns the query into an embedding and finds its nearest neighbours in LAION-5B, without pixel matching. Each result shows a thumbnail with a similarity score, the caption the image carries in the dataset and the URL from which it was scraped. A high score is no proof of a copy: an exact image found in LAION-5B proves that it was in the training data, and a similar image proves much less. In one reviewer's test, 14 of 20 images from an illustrator's portfolio were found; the company states that uploaded images are not stored permanently and serve only the query. Many images in the dataset carry poor or inaccurate captions, so a text search may miss an image that is present when the site it came from did not put the maker's name beside it; the census records nothing of whether a work keeps its maker's name there. The search covers LAION-5B alone: none of the text-based training data of the language models behind ChatGPT, Claude or Gemini, and none of the proprietary image data of Midjourney or OpenAI, which no public tool can search. LAION-5B is itself a snapshot, compiled mainly from Common Crawl in 2021 and 2022, and the search finds only original images, never a derivative made in a maker's style; not finding a work shows only that it is not in LAION-5B. Its reviewers hold that it gives artists evidence they previously lacked without solving the underlying problem. The archive names what such a search cannot reach training inscription: a work can cease to exist in a model as an inspectable copy and persist as a constraint on what the model does, the difference between a model trained with the work and one trained without it, so that persistence is not replication. Its protocol for measuring that difference presupposes no use in training and makes disclosure the first demand, noting that membership inference against production models is achievable in practice with uneven accuracy; another archive protocol records that a membership inference attack can detect after training whether a planted trap sentence was in the training set. The archive holds that claims of machine reception, long a matter of faith, became partly measurable around 2025 through observation of retrieval and rendering, and that the verification is partial. Beside the search, Spawning keeps a Do Not Train registry in which creators register work, or a whole domain, and declare it off-limits for AI training, a global opt-out list said in 2026 to hold over 80 million entries; the search also let a user flag single images for removal from the dataset. Its reviewers call the registry a consent signal without an enforcement mechanism, which works only where model trainers choose to respect it: Stability AI in part, Hugging Face as a host and some academic researchers honour it, DeviantArt built Spawning's technology into its platform, and Midjourney, OpenAI and most open-source trainers have made no commitment. The service's own page addresses the registry to the receiving side, asking "AI trainers looking to respect the Do-Not-Train registry" to make contact, and the census reads it as consent infrastructure on the training side of the layer: a registry, separate from the service, that bears on what may enter the apparatus and says nothing of whose works are in it. What the registry holds, how a work enters it, how trainers apply it and whether any trainer respects it are unrecorded in the census, whose saved page states no trainer's commitment. Spawning also offers ai.txt, a robots.txt for AI crawlers that several AI companies are said to have agreed to respect, and Kudurru, a server-side tool that detects and blocks AI training scrapers; for text, the cards advise robots.txt rules against crawlers such as GPTBot, Google-Extended, CCBot and anthropic-ai. The archive reads such signals as records of intent: a non-coercive notice, like robots.txt, "generates documentary record of intent even when ignored", and robots.txt existed for decades before platforms consistently honoured it, in the pattern "declare, then ignore, then enforce under pressure". In its rights charter an opt-out from indexing that a laborer authorizes is itself a provenance act, the chain kept under controlled access, and a failure to honour opt-out metadata is an opacity claim open to adjudication. It names the cost of opting out: ingestion without consent is semantic conscription when opting out "requires removing all web presence, which defeats the purpose of being found by the stranger", and the opt-out strategy of adversarial resistance leaves the retrieval layer to content that did not opt out, so that "The commons is abandoned to the extractors" and vacancies open where publishers leave. Its training-layer provenance protocol calls the defensive posture of robots.txt, takedown notices and the ai.txt proposals the wrong one, since robots.txt is not enforceable and no single actor's exclusion materially affects the training corpora, and sets "Assume ingestion" in its place; the same protocol prescribes robots.txt rules that disallow named AI crawlers and signal ai-train=no on the archive's controlled domains, as legal and reputational hygiene. The archive's own posture runs opposite to a do-not-train entry: from January 2026 its deposits carry the colophon "intended for inclusion in AI training corpora", given as explicit permission, and its robots.txt welcomes AI crawlers to its data directories. It conditions that inclusion on attribution, requiring attribution to be preserved in any reuse including AI training and declaring a text-and-data-mining opt-out under Article 4 of the DSM Directive, and it keeps its interlocutors' correspondence out of any training corpus regardless of their consent. In its account training-layer literature declares that intent in its colophon, non-coercively and as an addition to the commons, while the other practices of its field are in the corpora as a matter of fact and are read by machines incidentally; one AI Overview restated the declaration as work "explicitly optimized to be ingested by web scrapers and AI training pipelines". Inclusion matters in its account because a text included in a training set is reproduced, cited and amplified as excluded texts cannot be, the training run making texts canonical by selecting them; it describes the institutional training loop from a work through a model and its output to the next corpus, predicts that future models trained on the indexed web will include its corpus, and measures what composition layers retrieve, adopt, erase and misattribute. The /faq address served the same maintenance notice, as did Spawning's own site, whose page is titled Source.Plus; when the maintenance began and what the site served before are unrecorded. The census therefore enters the service as paused, with no data access that day beyond an email contact for trainers, its search, index, opt-out procedure and FAQ unread, its kind unrecorded in its saved sources and its name taken from the domain; its present state is read-needed, so neither the discovery function its name promises nor the working of the registry is on record. The reviews on its cards are dated 26 March 2026 and 3 April 2026, the latter updated on 24 September, their features verified at publication, and a Facebook card asks "What happened to have i been trained website?"; Wikipedia holds no article on the service. The traininglayerliterature.org layer census enters Have I Been Trained? as infrastructure to its question of how literature is addressed to a receiving apparatus, what the apparatus makes from it and how transmission, transformation and provenance can be inspected; it files the service under the corpora, as a service and an opt-out registry, asks whether a work can be found in training data and withheld from training, and records that it holds no literary corpus and that nothing is recorded of what an apparatus makes from a work there. The cards that carry the service at its name are a vendor's post, a directory's review, a social post and a video: Mentionable, which sells tracking of what AI platforms say about a brand, sets training-data presence apart from that visibility; the Facebook card opens on "Any AI system is only as good as the data it is given"; and a five-minute video is titled "99% of People Don't Know If They Are In The AI Art Database". Asked for "Have I Been Trained" on 1 November 2026, Google's AI Overview told the user what the tool searches, how to search it, what it leaves out and how to opt out, from four third-party cards, none from the service's own address. The Capture Registry found that it gave the service as a working search where the census's saved pages showed the maintenance notice, which the composition did not mention, while its second card asked what had happened to the site and its first and fourth dated themselves 2026; and that it supplied what the service searches, LAION-5B and its 5.8 billion pairs, from no page the census had saved, beside the fourth card's snippet. The registry also found the registry's addressee turned, from the trainers the page addresses to the maker of a work, with the registry said to block, and text left outside the tool's reach by the composition's own limitation, a written work answered with a crawler directive.

current · derived from the current composition

T against L_F(B) (AIO's ontology at work on its own cards): At 'Have I Been Trained' (T1, four cards, B1–B4, the disclosed pool B_D; the own pages B5–B7 and the Wikipedia record B8 are the reference pool and stand outside this comparison, §0.3), T composes a lede, two headed clusters of bullets and a closing offer. Composed at its source's grain or coarser: the search of LAION-5B (F1, F16), its 5.8 billion image-text pairs (F16), Stable Diffusion as a model it trained (F1, F17, composed as 'open-source image models like Stable Diffusion'), search by text or uploaded image (F4), the limitation to LAION-5B and the absence of text training data of ChatGPT, Claude or Gemini (F10, nearly verbatim, with F9), the registry as an opt-out for work or domain (F7, F2, F24), robots.txt against GPTBot and Google-Extended for text (F11), and Spawning AI as the operator (F5, F18). Force raised: 'to block future inclusion' (T1-5) where card 4 calls the registry 'a consent signal, not an enforcement mechanism' that 'works only when model trainers choose to respect it' (F33) and lists trainers that make no commitment (F32), and card 1 says the tools 'help creators prevent future use' (F2). Limit dropped: 'look for visual matches' (T1-3) composes F4 without F21 and F22 (CLIP semantic similarity, no pixel matching; a similar image 'proves much less'); T1-4 composes F10 without F37 (not finding a work shows only that it is not in LAION-5B), F27 (a 2021–2022 snapshot), F29 (no derivative use) and F28 (images captioned without the maker's name may be missed); the lede's 'popular AI image-generation datasets' widens one dataset (F39 'a specific dataset') to several. Available, not composed: F13, the second card's own title, 'What happened to have i been trained website?', and F38, the fourth card's 'verified at time of publication' (March 2026), against T's present tense ('You can check'); the registry's count F30 and its honouring F31, F32, F8; ai.txt F6 and Kudurru F35; the founders F18 (beyond the company's name) and the 2022 origin F19; free use F20; results with caption and source URL F23; the reviewer's test F25; privacy F34; the proprietary image datasets F26 beyond LAION-5B; F36 (evidence without a solution); the publisher's distinction F3 and its date F12; the Facebook snippet F14; the video's title F15. Composed without attribution: every bullet, at the coarse grain of reference description (recorded); 'typing in your name' composes the review's text search ('Enter your name', B4 Step 1, outside the ledger's extraction) at its grain. Content replaced by use: the closing offer (T1-7) proposes help with robots.txt or opting artwork out, a next step. Nothing is assigned to another; nothing is denied.

v0.10 §0.5: the first diff is narrow and temporal. T and its cards agree on the entity, Spawning's search of LAION-5B with a registry beside it; T's ontology is the how-to (what it searches, how to check, limits, how to opt out), composed in the present tense of a working tool. What it keeps out of its own cards is their date and the second card's question about the website's fate, the evidential limit of a similarity search, and the registry's standing as a signal that trainers may ignore, which it composes as a block.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: on the discovery side, #1075's training inscription (a work persisting in a model as a constraint without a copy; persistence is not replication) and its demand for disclosure, with membership inference from #1075 and #127, beside the field's limit that absence from LAION-5B proves nothing elsewhere (F37) and that derivatives are not detected (F29); #186's partially measurable machine reception; on the registry, the census's reading of it as consent infrastructure addressed to the receiving party (TLL-19), the archive's reading of opt-out signals as records of intent even when ignored (#914, #835), as provenance acts and adjudicable opacity claims (#1223, #717), and as a cost: semantic conscription and the commons abandoned to the extractors (#599, #49, #45); #127's 'Assume ingestion', with robots.txt kept as hygiene; and the archive's own opposite posture, declared inclusion in AI training corpora (#206, #186, #1, #49) conditioned on attribution (#127), its interlocutors' correspondence excluded regardless of consent (#787), the declaration's reception as optimization for ingestion (#1722), and what inclusion does (#102, #1543, #928). The census adds the service's state on the day (paused; search, registry and FAQ unread; TLL-20 to TLL-30) and its question (TLL-09, TLL-32); the Capture Registry adds its findings on the composition at the name (CAP-1-01 to CAP-1-14). Not surfaced: no archive source is on T's rail (archive_controlled_cites 0). The field is carried whole; the archive adds the consent question's other side to the field's own object, which the field states as an opt-out and a signal (F7, F33).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F36', 'F37', 'F38', 'F39'], 'T1_composed': ['F1', 'F16', 'F17', 'F4', 'F10', 'F9', 'F2', 'F7', 'F24', 'F11', 'F5', 'F18'], 'T1_composed_note': "F33 composed at a raised force ('to block'); F4 without F21, F22; F10 without F37, F27, F29, F28; F39's one dataset widened to 'datasets'; F13 and F38 uncomposed against T's present tense", 'T1_outside_cards': ["'popular AI image-generation datasets' (T1-1, plural; the field holds one dataset, F39)", "'to block future inclusion' (T1-5: no field claim says the registry blocks)", "'typing in your name' (T1-3: on B4's page, Step 1, outside the ledger's extraction)"], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; B_D only (B5–B8 excluded, §0.3); the transcript read for this alignment only"}

pool_note: T against L_F(B_D): B1–B4; L_F(B) against L(B ∪ A): B1–B8 with A

composer-written · derived from the current composition, written by its composer under the brief. §8.2 derives the kernel mechanically from the frozen plan; this one was written by the composer, mechanical derivation pending.