Skip to Fulmerica links
Worth Building · Chapter 175 October 2026A draft for public use

What public AI and data could be for

Every AI system runs on things people made. Posts, books, recorded speech, trip logs, health records, and the open web were gathered mostly without asking, and most of the value went to the companies that collected them. Between 2025 and 2027, consent standards, court settlements, state privacy laws, public models, and public computing change who can say yes, who gets paid, and who can build. This chapter applies the catalog's ideas to public AI and data, then maps eleven territories with current facts, things to build, who is already building them, scores, and where a reader could start.

01The idea, in public AI and data

Data that people can see, steer, and share in

A more capable system that leaves people more isolated and more dependent is a step backward, whatever its benchmark scores.

The protections in this domain already have names. The cooperative that pools its members' records and votes on who may use them. The license a community writes for its own language. The robots.txt file a small publisher edits by hand. The deletion request. The public list of software an agency uses to decide things about people. Each is being bypassed or buried. In July 2024 the Data Provenance Initiative audited 14,000 web domains behind C4, the Colossal Clean Crawled Corpus used to train many models. Within one year about 5 percent of its text had become fully restricted, and more than 28 percent of its most active sources. People were saying no with the only blunt tool they had. The work ahead is software that makes terms machine-readable, lets members own their records, and trains open models on data with a paper trail. It also keeps a public record of what automated systems did.

SceneThe default pathWhat we could build instead
The small publisher's websiteCrawlers copy every page. The only tool is a robots.txt file that blocks everyone or nobody, and visits from search keep falling.The site states its terms in a machine-readable license, chooses which uses are paid, credited, or refused, and can check a public record of which crawlers honored them.
The rideshare driverThe platform holds every trip. The driver cannot prove what an hour of work really paid, and the city buys trip data from the platform or goes without.Drivers pool their own trip records in a cooperative, sell the analysis to the city, elect the board, and share the profit.
The small nonprofitA discounted seat on one vendor's assistant. Staff learn by trial and error, the records stay with the vendor, and the price changes when the discount does.A shared set of tested uses and evaluations, export by default, a public model as one option, and a record of what the tool changed for the people served.
The language communityRecordings of the language are scraped, a company trains a speech model on them, and the community is sold its own voice back.The community records its own speech under its own license, decides who may use it, and keeps the benefit of the models built on it.
The data broker's listA person must find and petition hundreds of brokers one at a time, and nobody checks whether any of them complied.One deletion request reaches every registered broker, the person sees each broker's status, and compliance is public by month.
The automated denialA letter declines an application. It does not say that software helped decide, and there is no route to a person who can explain.The notice names the system, the public inventory lists it, a reader explains the decision plainly, and the correction clock starts the same day.
02Seven commitments, seven examples

What each commitment means when the subject is data

CommitmentIn public AI and data, it meansA concrete instance
1 · Judge by what happens to peopleJudge a data project by what members, writers, and communities can now do and what they receive. Model size and dataset size are inputs.A data cooperative that reports income returned to members and the decisions members made, alongside the number of records it holds.
2 · Keep people in chargeTerms a machine can read, a person who can answer, and exit. People can withdraw consent, export their records, and move to another service.California's single deletion platform, where one request reaches every registered broker, paired with a tracker the person controls.
3 · Build capability rather than dependenceAfter the first project, a nonprofit can test the next tool itself, and a community can train models on its own data.A shared evaluation kit that a food bank runs on three models, one of them public and open, before choosing.
4 · Make slow harm visibleSurface the gradual losses. Sites closing to all crawlers. Contributors unpaid. Incidents repeated. Agencies adding automated decisions without notice.A monthly public record of which crawler operators honor machine-readable terms.
5 · Design for groupsThe cooperative, the language community, the newsroom, and the nonprofit coalition are the users. Their shared terms matter more than any single click.Te Hiku Media's Kaitiakitanga license, under which the community decides as a group how its speech data may be used.
6 · Remove busywork, keep judgmentRemove the hundreds of deletion letters and the hand-edited crawler lists. Keep the community's vote on who may use its data and the human review of automated decisions.A notice reader that drafts the correction request after an automated decision while the person decides whether to send it.
7 · Name choices as choicesScraping without consent, closed training data, and vendor lock-in are decisions with owners and dates. Point at the projects that chose otherwise.Apertus publishes its training data, code, and recipes under an open license, so a closed model now has a working comparison.

The three questions, applied

ThingMore capable or more dependent?Closer to each other or to the system?See, steer, and leave?
Cloudflare's default blocking of AI crawlers and pay per crawlCapable; more than one million customers chose to blockCloser to one intermediary that sits between sites and crawlersSite owners steer; the rules live with one company
Reddit's licensing deal with Google, reported at about $60 million a yearNeutral for the people who wrote the posts; the platform gainsCloser to the platformThe terms are not public to the writers, who cannot steer the use
Apertus, the Swiss public modelCapable; data, code, and checkpoints publishedCloser; researchers in many languages build on itApache 2.0 license; anyone can run it, change it, or leave
Nonprofit discounts from OpenAI, Anthropic, Google, and MicrosoftCapable at entry, at 15 to 75 percent offCloser to a vendorLeaving depends on each vendor's export tools
Te Hiku Media's Kaitiakitanga licenseCapable; more than 2,500 people contributed labeled speechCloser; the community is the userThe community steers every use
California's Delete Request and Opt-Out PlatformCapable; one request in place of hundredsNeutral; nobody has to petition aloneThe person sees and steers; broker compliance still needs a public record
03Where to push

Where to push in public AI and data

Place to pushWhat it meansIn public AI and data
ParadigmAI as a meansThe purpose of public AI is people and communities who can build, decide, and share in the value. Larger models and larger scrapes are tools toward that purpose.
Goals and metricsCapability, trust, connectionValue returned to contributors. Share of training data with a known source and license. Nonprofits that can evaluate a tool themselves. Deletion requests honored on time. Communities holding licenses to their own language data.
Information flowsMake drift visible; the most practical leverWhich crawlers honor which terms. Which brokers deleted what. Which agencies use which automated systems for which decisions. Which incidents repeat. Most of this is held privately, and almost none of it is public by month.
RulesLegibility, portability, interoperability, exitMachine-readable licenses. Deletion through one request. Cloud switching without fees in the European Union from January 2027. Transparency about automated decisions in Colorado from January 2027. Public inventories of government AI.
Self-organizationCooperatives, civic groupsData cooperatives that hold members' records. Publisher collectives that license together. Language communities with their own licenses. Nonprofit coalitions that test and negotiate together. Open model projects run by universities and public supercomputing centers.
Why public AI and data, specifically

This is the domain where the catalog's precedents came from. Linux and Wikipedia showed that people working together could build shared infrastructure that no company owned, and the open web they helped make became the raw material for today's models. Credit unions and agricultural cooperatives showed that people with little bargaining power could pool what they had and keep the returns. MIDATA in Switzerland has applied that model to health records since 2015. The domain also shows every cause of failed scale. Rules differ by country and state (F1). Trust is hard to rebuild after years of scraping (F2). Nobody pays at the moment a person's data is taken (F3). Licenses and laws are dense (F4). Most contributors are too small to bargain alone (F5). Courts and contracts have started to put prices on training data. A September 2025 settlement in Bartz v. Anthropic pays authors about $3,000 per book, and Reddit's deal with Google was reported at about $60 million a year. In both cases the money flows to rights holders or platforms. The shared layer nobody owns yet is the record of who contributed what, on which terms, and what came back. The standards and laws of 2025 to 2027 make this the season to build it.

04Eleven territories

The data map

Scored on the catalog's rubric (shared layer nobody owns, buyer or deadline, outcome data, the three questions, fit). Each card names which of the nine reusable parts it reuses. Several cards share one ledger of terms, requests, and outcomes held by the people the data describes.

How to read a card: its score adds five judgments, each from 0 to 2 (shared layer nobody owns, buyer or deadline, outcome data, the three questions, fit), for a total out of 10.

D1Score 9 of 10Reuses: member-owned data, outcome commons, standards cohort

Data cooperatives that return value to members

Drivers, patients, and other people whose daily activity produces valuable records, the cities and researchers who want those records, and the platforms that hold them today.

Why now

MIDATA, a Swiss cooperative founded in 2015, lets members collect their health records and grant selective access for research, and it supports regional and national sister cooperatives. Driver's Seat Cooperative is owned by drivers, who share their driving data with local governments and transportation agencies, elect the board, and share the profits. Salus Coop in Barcelona organizes citizens around health data for research. The Mozilla Data Collective launched on 6 November 2025 so that communities can set terms for their datasets and sell them. The Open Data Institute's data trust pilots in the United Kingdom found that trusts fit best where interests conflict. It called for more research on cooperatives and trusts led by the people in the data.

Plays

  • A cooperative starter kit for one kind of worker or patient, with bylaws, consent records, a data schema, and a dividend rule that members vote on.
  • A first public buyer. Offer a city transportation office the drivers' pooled analysis of pay per hour and pickup times, with the price and the dividend published.
  • A standards cohort among the existing cooperatives to agree one consent record, so a member can move from one cooperative to another with their records and permissions intact.

Already there

Driver's Seat Cooperative, MIDATA and its sister cooperatives, Salus Coop, the Mozilla Data Collective, the Open Data Institute. Each runs alone. A shared consent format and a shared buyer list do not exist.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 2 · Fit 2

D2Score 8 of 10Reuses: place configuration, standards cohort, outcome commons

Machine-readable consent for the open web

Small publishers, libraries, archives, and nonprofits whose pages are copied by crawlers, and the AI companies that depend on a web that stays open.

Why now

The Data Provenance Initiative found that robots.txt restrictions on C4 rose from about 1 percent of its text in mid-2023 to about 5 percent a year later. Among its most critical sources the share passed 28 percent, and terms of service restricted 45 percent of C4. Since July 2025 Cloudflare blocks AI crawlers by default for new domains. More than 1 million of its customers have turned on blocking, and its pay per crawl option uses Hypertext Transfer Protocol (HTTP) status code 402, payment required. Really Simple Licensing (RSL) was first published on 10 September 2025 and is stewarded by the nonprofit RSL Collective. Version 1.0 became an official specification in December 2025, with Yahoo, Reddit, and O'Reilly among adopters and Cloudflare, Akamai, Creative Commons, and the Interactive Advertising Bureau (IAB) Tech Lab announcing support. Media Copilot, an industry newsletter, counted more than 1,500 publishers on board. Whether major AI companies honor RSL terms could not be verified in this pass.

Plays

  • A public crawler ledger. Consenting sites share their logs, and the ledger reports each month which operators honored robots.txt and RSL terms.
  • An RSL terms writer for small sites and libraries that explains the 1.0 categories plainly, such as allowing search but refusing AI answers, and versions each site's terms.
  • A collective license for small publishers, through the RSL Collective or a cooperative, with every payment to members visible.

Already there

The RSL Collective, Cloudflare and Akamai, Creative Commons, the IAB Tech Lab, and the Data Provenance Initiative's audits. Nobody publishes compliance by crawler operator.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 1 · Three questions 2 · Fit 2

D3Score 8 of 10Reuses: evidence file, member-owned data, outcome commons

Value return when people's work trains a model

Authors, forum posters, and encyclopedia editors whose work is licensed or litigated, the platforms and rights holders who receive payment, and the AI companies who pay.

Why now

On 25 September 2025 Judge William Alsup preliminarily approved a $1.5 billion settlement in Bartz v. Anthropic, a class action in which authors alleged that about 7 million books were downloaded from pirate libraries. It pays about $3,000 per book for the 482,460 works on its final list, requires the company to destroy the downloaded files, and was expected to pay out during 2026. In February 2024 Reddit agreed to license its content to Google for model training in a deal reported at about $60 million a year. The Wikimedia Foundation's Enterprise service has struck paid access deals with Amazon, Meta, Microsoft, Mistral AI, and Perplexity. In each case the payment goes to a rights holder, a platform, or a foundation. The people who wrote the posts and edited the articles are not parties.

Plays

  • A contributor ledger for a writers' cooperative or a community forum, recording what each member contributed and which licenses it entered, so any payment can be split by a rule members set.
  • A plain-language claim helper for creators when a settlement opens, with an evidence file of authorship and the claim deadline.
  • A public tracker of licensing deals and settlements with price per work and date, so communities bargain from facts.

Already there

Wikimedia Enterprise, the RSL Collective, authors' organizations and class counsel, Fairly Trained. No record of value returned to individual contributors exists.

Score

Shared layer 2 · Buyer or deadline 2 · Outcome data 1 · Three questions 2 · Fit 1

D4Score 8 of 10Reuses: standards cohort, outcome commons; new: provenance records

Training data with a paper trail

Open model builders, the researchers and public funders who need to know what a model learned from, and the writers and sites whose work is in the data.

Why now

EleutherAI released the Common Pile v0.1 in June 2025 with partners including Poolside and Hugging Face. It holds 8 terabytes of public domain and openly licensed text from 30 sources and took about two years to build. Its authors report that their Comma v0.1 models, with 7 billion parameters trained on 1 and 2 trillion tokens, perform on par with models of similar size trained on unlicensed data. Fairly Trained, a nonprofit, began certifying in January 2024 models trained only on data used with consent or under license. Apertus publishes its training data along with its code. The Data Provenance Initiative audits licenses and consent signals across widely used datasets.

Plays

  • A provenance record for every dataset, listing source, license, the consent signal in force on the day of collection, and the date, adopted by a cohort of open model projects.
  • A snapshot service that records a site's robots.txt and RSL terms at the moment a page is copied, so training sets can prove what was allowed.
  • An evaluation commons comparing licensed-only models with others on the tasks nonprofits and public agencies care about.

Already there

EleutherAI, Hugging Face, the Data Provenance Initiative, Fairly Trained, the Swiss AI Initiative. A shared provenance format is the gap.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 1 · Fit 2

D5Score 8 of 10Reuses: member-owned data, standards cohort, outcome commons

Language data governed by the people who speak it

Speakers of languages with little text online, the community groups that steward them, and the researchers and companies that want speech and text data.

Why now

Te Hiku Media in New Zealand licenses its te reo Māori speech data under its Kaitiakitanga license, which allows use only for the benefit of Māori. A 10-day campaign drew more than 2,500 people and produced more than 300 hours of labeled speech, according to NVIDIA's account. Mozilla's Common Voice covered more than 250 languages as of December 2025 and runs in-country programs with local partners. Kathleen Siminyu, a researcher in the Masakhane community, has worked on the Kiswahili Common Voice dataset. The Mozilla Data Collective, launched on 6 November 2025, lets communities set terms and sell access to their datasets.

Plays

  • A library of community data licenses modeled on Kaitiakitanga, which other communities can adapt with their own counsel.
  • A recording campaign kit with prompts, validation, and a dashboard the community owns.
  • A ledger of uses and revenue for community datasets licensed through the Data Collective, reported back to the community each quarter.

Already there

Te Hiku Media, Mozilla's Common Voice and Data Collective, Masakhane. The work must be led by people from each community; outside builders supply tools and templates.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 2 · Fit 1

D6Score 7 of 10Reuses: standards cohort, place configuration

Public models and public inference

Nonprofits, public agencies, schools, and researchers who need a capable model they can inspect and run, and the universities and public supercomputing centers that can host one.

Why now

Apertus was released on 2 September 2025 by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre under the Apache 2.0 license, in 8 billion and 70 billion parameter versions. EPFL says it covers more than 1,000 languages, with about 40 percent of its training data in languages other than English. It was trained on the Alps supercomputer and is distributed through Hugging Face and Swisscom. It publishes its data, code, alignment recipe, and intermediate checkpoints. OpenEuroLLM, a pan-European project with a reported budget of €37.4 million, had published experimental checkpoints, and whether its first full model shipped in 2026 could not be verified. A public inference service open to nonprofits could not be confirmed in this pass.

Plays

  • A public inference pilot. A university or public supercomputing center hosts an open model for nonprofits and agencies at cost, and publishes usage and outcomes.
  • A deployment kit for public models with configuration by country and language, so a ministry or a city can run one without a vendor.
  • Fine-tuning for small-language communities on public models, with the community holding the adapted model (D5).

Already there

The Swiss AI Initiative, OpenEuroLLM, EleutherAI's Comma models, Hugging Face. Hosting for people who cannot run models themselves is the gap, and it is operations-heavy.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 1 · Three questions 2 · Fit 1

D7Score 6 of 10Reuses: standards cohort; new: shared compute allocation

Public computing for researchers and civic builders

University researchers, students, and small teams who cannot rent large clusters, and the state and federal programs deciding who gets public computing time.

Why now

The National Artificial Intelligence Research Resource (NAIRR) pilot launched in early 2024, and Lawfare reported 26 corporate and nonprofit partners. By June 2026 it had supported more than 700 projects and 7,000 students in all 50 states, according to HPCwire. The Senate's fiscal 2026 spending bill for the National Science Foundation (NSF) proposed $30 million for the pilot, and NSF solicitation 25-546 sought an operations center for the full resource. California's Senate Bill 53, signed on 29 September 2025, creates a consortium within the Government Operations Agency to plan CalCompute, a public computing cluster. New York's Empire AI joins universities in a shared computing consortium.

Plays

  • An allocation track for nonprofits and community groups that pairs computing time with a technical mentor and measures tools shipped.
  • A public dashboard of who receives public computing time, by state, type of institution, and purpose.
  • Shared datasets and evaluation sets hosted next to the computing, so small teams start with more than hardware.

Already there

The NAIRR pilot and its partners, NSF, Empire AI, the CalCompute consortium. Running computing is operations-heavy, and a track for civic builders is the gap.

Score

Shared layer 1 · Buyer or deadline 2 · Outcome data 1 · Three questions 1 · Fit 1

D8Score 9 of 10Reuses: standards cohort, outcome commons

Nonprofit access to models, with the means to choose and leave

Small nonprofits, the coalitions and funders that support them, and the four large vendors now offering discounts.

Why now

Fast Forward's 2025 AI for Humanity Report, released on 30 September 2025, surveyed nearly 200 nonprofits in April 2025. Among the AI-powered nonprofits surveyed, 30 percent had budgets under $500,000, 40 percent had used AI for a year or less, and 84 percent said funding would help most. Vice compared the vendor programs. OpenAI offers 20 percent off ChatGPT Business and 25 percent off Enterprise. Anthropic offers up to 75 percent off Claude Team and Enterprise from 2 December 2025, with Benevity, Blackbaud, and Candid as partners. Google offers 70 to 75 percent off Workspace tiers. Microsoft offers 15 percent off Microsoft 365 Copilot. Eligibility is checked through Goodstack or TechSoup.

Plays

  • A nonprofit evaluation commons with real tasks, test sets, and results across vendors and public models, so a food bank can compare before it commits.
  • An exit checklist for each vendor, versioned by month, covering export formats, data retention, and the price after the discount.
  • An outcome record of what a tool changed for clients, such as hours saved or cases resolved, pooled with consent across organizations.

Already there

Fast Forward, TechSoup, Goodstack, Candid, and the vendors' own programs. A neutral comparison and a shared outcome record do not exist.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 2 · Fit 2

D9Score 10 of 10Reuses: notice reader, deadline engine, place configuration, records-request pipeline, outcome commons

Deletion, portability, and the right to leave

Everyone whose records data brokers trade, the small organizations locked into one cloud provider, and the regulators who must check compliance.

Why now

A January 2026 roundup by the law firm Hunton lists 18 states with comprehensive privacy laws in effect. Under California's Delete Act, the Delete Request and Opt-Out Platform (DROP) went live in January 2026, and registered data brokers must process requests from 1 August 2026. The European Union's Data Act has applied since 12 September 2025, giving customers rights to switch cloud providers and move their data. Switching fees are being phased out, and none may be charged from 12 January 2027.

Plays

  • A deletion tracker the person owns, which logs each DROP request and each broker's response and counts the days allowed under state law.
  • An outcome commons of broker compliance by month, from consented tracker records, shared with the state regulator and the press.
  • A switching kit for nonprofits and small firms in the European Union, ready before January 2027, and place configuration for each American state's privacy law.

Already there

California's DROP, state privacy regulators, law firm trackers of state laws. Nobody publishes which brokers deleted what, or when.

Score

Shared layer 2 · Buyer or deadline 2 · Outcome data 2 · Three questions 2 · Fit 2

D10Score 9 of 10Reuses: notice reader, evidence file, deadline engine, records-request pipeline

Automated decisions, listed and explained

People denied housing, jobs, credit, or benefits by systems they cannot see, the legal aid offices and libraries that help them, and the agencies and companies that deploy the systems.

Why now

On 3 April 2025 the Office of Management and Budget (OMB) issued memo M-25-21 on federal use of AI and M-25-22 on buying it. Both build on an executive order of 23 January 2025. The 2025 federal AI use case inventory listed 3,611 uses across 56 agencies, up 105 percent from 1,757 in 2024, according to Nextgov. Colorado's 2024 AI Act was first delayed from 1 February 2026 to 30 June 2026. On 14 May 2026 the governor signed Senate Bill 26-189, which repeals it and replaces it with a narrower law on automated decision-making technology used for consequential decisions. The new law centers on transparency and consumer rights, drops the earlier duty of reasonable care, risk programs, and annual impact assessments, and takes effect on 1 January 2027 after the attorney general's rulemaking.

Plays

  • A readable federal inventory that links each listed use to the program, the decisions it touches, and the notices people receive, refreshed with each annual release.
  • A notice reader for Colorado's automated-decision notices from January 2027 that explains the decision, builds the evidence file, and helps the person use whatever correction rights the final rules give.
  • A records-request pipeline for state and local agencies, which the federal inventory does not cover.

Already there

The federal inventory itself, Colorado's attorney general's rulemaking, law firm and press trackers, legal aid offices. The person-side reader and the state and local inventory are open.

Score

Shared layer 2 · Buyer or deadline 2 · Outcome data 1 · Three questions 2 · Fit 2

D11Score 8 of 10Reuses: outcome commons, standards cohort, records-request pipeline

AI incidents as a public record

People harmed by AI systems, the journalists and researchers who document harm, and the standards bodies and regulators that need patterns.

Why now

The AI Incident Database, run by the Responsible AI Collaborative, held about 6,000 incident reports, according to a presentation on the database at a National Institute of Standards and Technology (NIST) workshop on AI incident management on 14 May 2026. The Center for AI Standards and Innovation (CAISI) at the National Institute of Standards and Technology (NIST) is carrying out parts of the federal AI Action Plan, including recommendations on incident response. California's Senate Bill 53, signed on 29 September 2025, requires frontier model developers to publish safety frameworks and report critical safety incidents to the state's Office of Emergency Services.

Plays

  • A community intake for incidents, so legal aid offices, libraries, and nonprofits can report harms from automated systems in a format the database can take.
  • A monthly view of repeated incidents by sector and type of system, aimed at the agencies that buy them.
  • A standards cohort to agree one incident record across the database, researchers, and public agencies.

Already there

The Responsible AI Collaborative, NIST's CAISI, California's reporting channel under SB 53. Reports from the people harmed, as opposed to the press, are thin.

Score

Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 1 · Fit 2

The extraction trap, and what not to build

The years that bring consent standards and deletion rights also bring a market for squeezing more out of personal data. Apps that pay a few dollars for browsing history. Crawlers that rotate addresses to slip past blocks. Discounts that make one vendor the default before anyone compares. The pattern worth building runs the other way. The people in the data hold the record and vote on its use, terms are honored in public, and every automated decision can be seen and corrected.

05An ordinary day

A Thursday in March 2027, with the data serving the people it describes

In Sacramento a home health aide checks her deletion tracker over breakfast. Forty-one brokers have confirmed deletion since August, three are late, and the tracker has drafted the complaint for her to read. In Denver a man receives a letter declining his rental application, and it says an automated system helped decide. At the library, the notice reader shows which records the system used, and he sends a correction that afternoon. In Chicago a food bank's staff test three models, one of them public and open, on fifty real client questions before renewing a contract. They pick the best value and keep the test set for next year. A drivers' cooperative sends the city its monthly analysis of pay per hour and pickup times, and the dividend reaches members on Friday. A small history blog in Ohio receives its first payment from a crawler that honored its terms, and the public ledger shows which crawlers did not. A language group records two hundred hours of speech under its own license and votes on the first research request. In the evening, the incident log shows a cluster of benefit denials from one vendor's system in three states, and a legal aid office files the first records request. In each place the people described by the data can see it, steer it, and leave.

Signs we are getting there

  • The largest AI crawler operators state publicly whether they honor RSL and robots.txt terms, and an independent monthly ledger confirms what they do.
  • California publishes, by broker, how many deletion requests each registered data broker received and completed on time.
  • Each new open model release ships with a provenance record for its training data, and more than one widely used model is trained only on licensed or public domain data.
  • Data cooperatives publish member counts and dividends each year, and at least one city buys workers' pooled data from a cooperative.
  • Colorado's automated-decision law produces a public count of corrections requested and granted in 2027, and state and local agencies publish AI inventories like the federal one.
06Where you could start

First steps for a weekend, a season, or a career

A weekend

  • Add machine-readable terms to your own or your organization's website with RSL or robots.txt rules, then log which crawlers arrive over the next month (D2).
  • Record and validate sentences in your language on Common Voice, or help a community group run a recording session on its own terms (D5).
  • If you live in California, file a request through DROP and document each broker's response and timing for a shared tracker (D9).
  • Read one agency's entries in the federal AI use case inventory and write one plain paragraph per use on which decisions it touches (D10).

A season

  • With one nonprofit coalition, build an evaluation set of fifty real tasks, test three models including one public open model, and publish the results and costs (D8, D6).
  • With one city transportation office and a group of drivers, pilot a pooled trip-data analysis and pay a dividend at ninety days (D1).
  • With one legal aid office or library in Colorado, build a notice reader before 1 January 2027 and record every automated-decision notice it sees in the first quarter (D10).
  • With a publishers' group, run a 90-day crawler ledger across fifty consenting sites and publish compliance by operator (D2, D11).

A career

  • Found the member-owned data cooperative for one kind of worker or patient, then carry its consent format to the next cooperatives (D1, D3).
  • Convene the standards cohort for provenance records across open model projects, and win a public funder that requires them (D4, D6).
  • Lead a public inference program at a university or supercomputing center that serves nonprofits and agencies at cost and publishes outcomes (D6, D7, D8).
  • Run the public record of AI incidents and automated decisions for a decade, so the harms that repeat are named by system and year (D10, D11).

Who is already doing it

  • Driver's Seat Cooperative is owned by drivers who share their driving data with local governments and share the profits.
  • MIDATA is a Swiss cooperative, founded in 2015, where members hold their health records and grant access for research.
  • Te Hiku Media stewards te reo Māori speech data under its Kaitiakitanga license.
  • The Mozilla Foundation runs Common Voice in more than 250 languages and launched the Mozilla Data Collective in November 2025.
  • EleutherAI released the Common Pile, 8 terabytes of openly licensed text, with partners including Poolside and Hugging Face.
  • The Data Provenance Initiative audits the licenses and consent signals behind widely used training data.
  • The RSL Collective is the nonprofit steward of the Really Simple Licensing standard.
  • The Swiss AI Initiative of EPFL, ETH Zurich, and the Swiss National Supercomputing Centre built Apertus.
  • The Responsible AI Collaborative runs the AI Incident Database.
  • Fast Forward publishes the AI for Humanity Report on how nonprofits build with AI.
07Scorecard

Eleven territories on one page

Score key: SL shared layer nobody owns, BD buyer or deadline, OD outcome data accrues, 3Q the three questions, FT fit for a builder who convenes. Each is scored 0 to 2, so the total is out of 10.

CodeTerritoryShared layerBuyer or deadlineOutcome dataThree questionsFitTotalHorizon
D9Deletion, portability, and the right to leave2222210Now, with DROP processing since August 2026
D1Data cooperatives that return value2122292027 with a city buyer
D8Nonprofit access to models212229Now, while the discounts are new
D10Automated decisions, listed and explained221229Before 1 January 2027
D2Machine-readable consent for the web211228Now, with RSL 1.0
D3Value return to contributors2212182026 to 2027, as settlements pay
D4Training data with a paper trail2121282027 cohort
D5Community-governed language data212218Ongoing, led by each community
D11AI incidents as a public record2121282027 with legal aid partners
D6Public models and public inference2112172027 pilot with a host
D7Public computing121116After the CalCompute plan
08Picks and declines

Six to pursue first, and five to leave unbuilt

  1. The deletion tracker and broker ledger (D9). California's platform is live and brokers have had to process requests since August 2026. A tracker the person owns and a public record of compliance turn a new right into a checked one.
  2. The automated-decision reader (D10). Colorado's law starts on 1 January 2027 and the federal inventory doubled in a year. Build the reader with a library or legal aid office before the first notices arrive.
  3. The data cooperative kit (D1). Driver's Seat and MIDATA show the model works. Give the next cooperatives a shared consent format and a first city buyer.
  4. The nonprofit evaluation commons (D8). Four vendors now offer discounts from 15 to 75 percent. A shared test set and an exit checklist let nonprofits choose with evidence and leave when they need to.
  5. The crawler ledger (D2). RSL 1.0 and default blocking give sites a voice. A public record of which crawlers honor it is the missing piece.
  6. Community language data (D5). Supply license templates and recording tools to the communities that lead this work, starting with those already in Common Voice.
A data dividend app that sells personal data. It pays people a few dollars and turns them into inventory. The only version worth building is a cooperative where members vote on each use and see every payment.
A crawler built to slip past opt-outs. It teaches every site to block everyone, and the open web closes faster. Build crawlers that read terms and publish their compliance.
A nonprofit assistant tied to one vendor's discount. The discount makes one company the default before anyone compares, and leaving gets harder each month. Build the comparison and the exit path first.
A model called public with closed training data. Without the data and the recipe, nobody can check what it learned or rebuild it. Apertus shows the fuller version is possible.
Automated eligibility scoring sold to agencies without notice. It moves the decision out of sight just as Colorado moves toward transparency. Build the notice, the inventory, and the correction path instead.
Sources for this chapterChecked 5 October 2026

Assumptions and limits. Scores are judgments. Figures were checked on 5 October 2026 against search result text, because the source pages could not be opened here, and each figure is attributed to the organization that reports it. This draft rests on 20 web searches, 15 in an earlier session and 5 in this one. No public inference service open to nonprofits, such as the Public AI Network, turned up in search. Whether OpenEuroLLM's first full model shipped in 2026, and whether major AI companies honor RSL terms, could not be confirmed. A reported 2026 change to Cloudflare's pricing is omitted. Empire AI's funding figures are omitted because the only source link found was incomplete. Member counts for MIDATA, Driver's Seat Cooperative, and Salus Coop were not found. Reddit's deal value is as reported by the press, and the terms are not public. The Fast Forward survey covers about 200 nonprofits that build with AI, not the whole sector. Vendor discounts come from Vice's comparison and can change. Counts of state privacy laws differ by source. Sources differ on how many languages Apertus covers, and EPFL's figure is used. The Comma models' parity with models trained on unlicensed data is the authors' claim. The AI Incident Database counts reports, which outnumber distinct incidents. The NAIRR figure of $30 million was a Senate proposal, and the final amount was not verified. Colorado's new law depends on rules the attorney general had not finished. Google.org's AI programs, the AI Alliance, and state AI inventories were searched for without usable results. A separate spot-check on 5 October 2026 used 7 more searches and also relied on search results, because the source pages could not be opened. It corrected the Bartz works count to 482,460. The AI Incident Database's count of 6,173 reports did not appear in search results, so it is given as about 6,000, and the share of C4 text restricted by robots.txt is given as about 5 percent because that figure did not appear in search results either.