CRAIONS Methodology — How the Observatory Was Built

An Observatory, not a List or Report

In the Spring of 2025, at the start of this project, we made a list of major issues of how AI was impacting the arts: ownership and authorship, training data and fair use, voice and likeness, platform labeling, guild contracts, market structure, practical steps, and what to watch. These reflected issues we had observed regarding the state of AI and creative industries.

It was accurate then, but within a few months the state of play had changed enormously. Every line in the memo was a moving piece, and that’s the difficulty of tracking AI policy by issue. An issue list is organized around cases and statutes, but because AI is impacting the creative world at a systemic level in which everything is shifting or challenging status quos, tracking by issue is too confining a view.

Another problem: Most of those working to define the rules for AI and creative work, through lawsuits, legislation, licensing deals and platform policy, have large stakes in existing systems: technology companies, major labels, studios, big publishers, platforms. Nonprofit arts organizations and the vast majority of working artists don’t have the time or the background to follow it, let alone take part. This has happened before. During the digital transition of the late ’90s and early 2000s technology companies wrote new rules for digital marketplaces by themselves, putting themselves in the middle and setting tolls. We’re still living with the consequences.

We endeavored to create something more useful than a report that would quickly be out of date, or an enumeration of lobbying positions argued issue by issue. Instead we decided to build a clearinghouse or observatory, uncolored by lobby positions or interests that had a stake. The intent is a neutral journalistic reporting, in hopes of keeping an eye on the bigger picture. Our goal is to track what is being decided, across commercial culture, nonprofit arts and the tech industry, make it freely available and built to be corrected rather than republished. CRAIONS, the Creativity & AI Observatory, is that record. This essay explains how it’s organized, what goes into it, how it’s checked, and what it can’t tell you.

Organizing by function

The way to see the bigger picture is to stop sorting by issue and instead sort by function, tracking what a development does to creative life, rather than which statute it invokes. While statutes come and go, the questions they act on change more slowly. How do we decide who made something? How does creative work get paid for? What keeps a creative field producing new work and connecting with audiences? If the categories are these questions, then perhaps an overall shape can be discerned.

This led to three systems:

Attribution and Accountability asks who or what gets recognized, trusted or held responsible for a creative act.
Value and Exchange asks how creative labor and creative work are priced and paid for once AI enters the market.
Cultural Infrastructure and Ecology asks what happens to the conditions that produce creative work in the first place: how new artists are trained, how audiences find work, the institutions that fund and interpret it, and the public resources it draws on.

An entry can be placed in one, two or all three systems. A single court ruling can settle an authorship question, set a damages figure and change employment in a field.

An earlier attempt to classify the playing field resulted in three different spheres: “Systems & Value,” “Tools, Skill & Practice,” and “Culture, Identity & Public Trust.” The middle category about how AI changes what happens inside a creative studio, eventually morphed into the third system, which for a while was called “Creation and Cultural Ecology.”

Why sort this way?

The project started out by identifying and tracking sources. ArtsJournal.com had been tracking news stories on AI and creativity day by day, providing a base of hundreds of sources. We hand-built a database of news stories and primary sources. And we found several trackers specific to industries and private interests. Law firms track litigation. Legislative services track bills, usually as a compliance product. The University of Glasgow’s CREATe tracker codes commercial licensing deals (274 of them by January 2026) better than we could, and Edward Lee’s tracker follows US copyright suits within days of filing. Each exists because a profession has a reason to watch it. You can learn more about these various efforts on the CRAIONS Resource Page.

But nobody owns the whole systemic shift, and the whole problem is exactly what an arts leader faces. A voice-cloning statute in Tennessee, a digital-replica clause in a SAG-AFTRA contract and a platform’s likeness-detection tool are three instruments in three different trackers, but they are also three answers to the same question. Sorting by instrument tells you what a profession is watching, but sorting by function attempts to let you know what’s at stake.

By fall 2025, we had tested this way of sorting and the systemic approach held.

Did it hold?

The database got bigger and more complex as we added tags, and individual entries needed to be frequently revised, updated and cross-referenced, and we needed a better way to monitor and iterate new versions.

So we built a model that focused on primary sources, trackers and news stories and was guided by the systemic lens framework we had developed. The model went through 35 version tweaks in testing. In July 2026, at 308 entries, we ran a structural audit of Observatory entries against the framework for tracking developments that we had settled on. The durability bet was holding. Eighteen months of statutory churn had produced no creative-sector instrument the three functions couldn’t slot.

What gets in

Entries are added to the Observatory when they are a new development that is additive to an issue. We don’t list every individual, duplicative case. A development gets an entry if it introduces a new legal theory, a new category of plaintiff, a new empirical mechanism or a new policy instrument. Related entries already in the map get appended as an update. But there are more than a hundred US copyright suits against AI companies currently pending, and the map doesn’t need a hundred entries that say the same thing. The leading litigation is included but subsequent filings that are similar are not included.

Entries are living records. A case gets updated as it moves; a bill’s status changes when it’s signed or dies; an entry that has been overtaken by new developments can be frozen as a dated record and marked superseded rather than rewritten. Alongside the dated events there are a smaller number of arguments that generalize from several entries (the work-for-hire trap, the implementation gap), and they trace what evidence they generalize from.

What each entry records

Field	What it records
Systems	Which of the three functions the development operates on. The rule is to tag the function acted on, not every downstream effect.
Type	The kind of instrument: litigation, legislation, regulation, market, labor, platform, standard, framework, initiative, advocacy, research or multi-vector.
Creative sectors	Music, visual art, film and TV, publishing, journalism, performance, games and others.
Region and date	Where and when. Region is derived from the data.
Status	Active, pending, emerging, resolved, decided but not final, dormant, superseded.
Actors	Named organizations, registered and grouped by sector (98 at present), so a reader can follow one company or union across the map.
Watch level	The curator’s judgment of momentum, not importance. Includes receding and settled, so the map can record that a concern went away.
Representation tier	What kind of organized voice, if any, the affected creators have, on a seven-step scale that ends at “no organized voice.”.
Evidence themes	For research entries: these studies explore workforce impact, content flood, detection gap, trust penalty, creativity comparison, cognitive effects.
Decided by	Whose decision produced the development: a court, legislature, regulator, platform, rightsholder, union or guild, AI developer, funder, creators. A lawsuit filed is coded to the filer; only a ruling is coded to a court.
Creators at the table	Yes (a party to a decision); Consulted (a formal input channel such as testimony, but no vote); No, or Not applicable. Informal input channels such as advocacy, amicus filings or public comment are not marked as being a party to the decision.
Sources, source quality, coding confidence	Links to primary documents where they exist; a rating of the sources; and a flag where an entry rests on something thin, such as a press release describing an unpublished agreement.
Cross-references	Links to related entries, across system lines.
Curator notes	Caveats, counter-evidence and the reasoning behind contested calls.

How it’s built and checked

In February, 2026 we began using an AI model for research collection, pointed at sources we had manually collected. We built a verification layer of adversarial agents trained to challenge, with every entry approved by the human curator before it is listed in the tracker.

Each version moves through the same steps:

Sweep. Research passes cover four stages over a window of two to four weeks: litigation dockets; legislation and regulation; deals and platform policy; and non-English-speaking jurisdictions and thinly covered sectors.
Verify, adversarially. Every potential finding goes to independent passes instructed to refute it, to read the primary document (the docket, the bill text, the agreement) and to default to “unproven” when not reached.
Composed against the source. The entry is included based on the original source documents wherever possible, rather than relying on news accounts. Every “first” claim gets a search for an earlier instance.
Decide. The curator approves each new entry and every change to an existing one.
Build and regress. A script applies the version to the spreadsheet that holds the full record and to the source file the website reads, in one pass. A regression check confirms that no entry was lost, that only the intended entries changed, that the two files agree field by field and that no cross-reference points nowhere. Then an update to the site is rendered as a test before anything publishes.
Tested. Each new version is reviewed and challenged by the curator before updating previous versions.
Record. The verification record for the sweep, by both the AI search engine and the curator, including what was refuted and why, is kept with the version.

The Observatory is currently at version 85, with 456 entries, as of 16 September 2026.

Counting who was at the decision-making table

After months of monitoring entries we decided to test against original premises – how actively involved are artists in decision-making around new rules of the road for AI? As a result, two fields were added in July 2026: who decided and whether creators were present. Here’s how the determination was built:

Our criteria:

Unions and guilds are creators. A contract bargained by the WGA, SAG-AFTRA or the AFM is coded as creators present and creators holding the pen.
Rightsholders are not creators. A label licensing its catalogue to an AI company is coded as creators absent, unless a creator organization was also a party. The Observatory itself shows why the line has to sit there: Sometimes creators are also the rights holders. More often, the catalog owner is not the creator, and deals made with the catalog rights owner will not flow back to the artists.
Being heard isn’t being a party. Testifying, filing comments, sitting on a working group or sponsoring a bill is coded Consulted. Artists or their unions hold no vote; the legislature decides.
Nothing decided, not counted. Research entries don’t result in policy decisions and standing frames do not include decision-making designations in the Observatory.

Given the criteria, the map clearly shows that, so far, collective representation is almost the only means by which artists have a meaningful voice in decisions.

Is anyone else counting this? We looked. There’s an atlas of participatory AI projects, an index that scores AI companies on participatory governance, and a 2026 study of who actually submitted to the EU AI Act consultations. None focus specifically on the creative sector.

The argument and the record

CRAIONS has an interpretive layer, represented in the current version by way of how entries are organized into systems and how they are tagged. An earlier version of the site also contained an extended introduction and three framing essays, one per system, arguing what developments might mean. Each of the essays named the ways institutions are under stress: disruption in attribution, capture in value and exchange, substitution in cultural infrastructure. The entries’ summaries and curator notes carried these arguments down to the individual record, saying why a development matters and what it connects to.

That, in our pre-launch feedback, proved useful to some readers and an obstacle to others. A legislative staffer, a lobbyist, a researcher or someone who disagrees with the framing may want the record without the reading of it. So we built the current version of the site with the interpretive framing pared back. The interpretive work will find a voice in another space. The framing essays are replaced by plain descriptions of the three systems. And each entry carries a basic rather than interpretive summary: what the development is, its cross-references and its source links.

What CRAIONS can’t tell you

A few limitations:

It’s hand-coded. Systems, watch level, representation tier and who was at the table are judgments made against a fixed vocabulary. The rules are published so you can disagree with a specific call.
It’s not a census. It’s a structured record of developments. The 140th suit on a settled theory won’t have its own entry.
It leans English-language. The sources are largely in English, and the coverage of Africa, Latin America and South and Southeast Asia is thinner than the world deserves.
Some fields are sparse. Evidence themes are set on a minority of entries, mostly research, so filtering by theme shows a small slice of the map.
Live litigation dates fastest. Case statuses can change in a week. Check an entry’s date and coding confidence before quoting it, and read the source.
On the map view, position means something but distance doesn’t. Entries sit inside the region for their system, but two dots being close together carries no information.

The list of changes we’ve considered and deferred is part of the method too: a fourth system for who decides; splitting Type into the arena and the instrument used; a column for the issuing jurisdiction alongside region; separating dated events from standing frames as a field; and a way to mark which entries support, complicate or contradict the map’s own arguments.

The best way to use the Observatory is to start from a function rather than a sector, pick an entry and follow its cross-references outward, then go to the sources. And if you find an entry that’s wrong, tell us. A large share of the work on this project has gone into finding our own mistakes. Wouldn’t it be better to find them with more people looking?