Your Commits Are for Sale: Who Owns Code History Sold to AI

The email looked like a revenue opportunity. In early June 2026, some developers whose Android apps had been downloaded millions of times received an invitation from Google to “join a confidential content offer pilot”. The pitch was simple: “Get paid for sharing the code powering your apps, as well as your archived projects.” The reassurance came straight after: “This is non-exclusive. You keep 100% of your IP.”

The message, first reported by Jason Koebler at 404 Media and then widely picked up, never used the words artificial intelligence. It said the code would help “improve Google's developer tools and products”. A linked page was more candid, describing “partnerships to improve our AI products”. Nobody outside the pilot knows what Google was offering to pay.

Earlier in the year, a founder named Keelan Miskell had been closing down his crypto startup. According to MIT Technology Review, he netted roughly $25,000 for his codebase, with a further payment, probably around $15,000, pending for the company's workspace data from Google, Notion, Figma and Slack. “At least all that stuff we were doing for two and a half years is worth something to someone somewhere,” he told the magazine.

Then, in August, came the deal that made everyone look up. In the bankruptcy of Spirit Airlines, which stopped flying in May, Google won an auction for the carrier's corporate data with a $10 million bid. TIME's Billy Perrigo reported that the cache held 100 million emails, 500 million Microsoft Teams messages and about 30 million lines of code. A court filing released in November put the volume far higher: about 1.9 billion internal emails, 664 million Teams items, 91 million SharePoint and OneDrive files and roughly 30 million lines of code.

Taken together, these episodes describe something new: a market in which the private working record of software (the code, but also the commit history, the pull requests, the arguments in review threads, the tickets and the design documents) is being bought and sold as AI training material. What ethical, legal and security problems arise when that record changes hands? How can a buyer prove a seller had the right to sell it, when dozens of people wrote it? And does the money now flowing into this trade create risks that existing safeguards were never designed to catch?

Why the Messy Middle of Software Became Valuable

For a decade, AI companies trained on the public internet, and code was one of its richest seams. Public repositories on GitHub and elsewhere supplied billions of lines for the first generation of coding assistants. That seam is thinning. Epoch AI, a research group that tracks the inputs to machine learning, estimated in 2024 that the effective stock of human-generated public text was around 300 trillion tokens, and gave an 80 per cent confidence interval that it would be fully used at some point between 2026 and 2032.

Private code is the obvious next reservoir. It is not one that platforms can simply tap: when GitHub updated its terms in March 2026 to let it train on Copilot interaction data from individual users who do not opt out, it stated plainly, “We do not use private repository content at rest to train AI models.” Private repositories have to be acquired from whoever controls them, and they are valuable for reasons beyond volume. A finished public library shows what working software looks like. A private repository, with its full history, shows how software actually gets made: the false starts, the reverted commits, the reviewer who points out a race condition, the ticket that explains why a strange workaround exists. That is precisely the material that the newest training methods crave.

Turing, the AI training and talent company, made the case explicitly in its AGI Advance newsletter of 24 December 2025, announcing a programme it called Project Lazarus. “We're acquiring full private repositories, design docs, testing logs, infrastructure manifests, and internal wikis from defunct startups,” it wrote, naming fintech, pharmaceuticals, manufacturing and software as target sectors. Existing datasets, the newsletter argued, “fail to capture the texture of real work”. The stated aim was to “preserve the complete operational history of defunct companies”.

The appetite is driven by reinforcement learning. Rather than simply predicting the next token of text, AI agents are increasingly trained in simulated environments where they take sequences of actions and are rewarded for good outcomes. TIME reported that gains from this approach have been largest in coding, where results are easy to verify. Surge AI builds such environments, which it calls reinforcement-learning gyms, and Nick Heiner, who leads that work, told MIT Technology Review he wants archives that capture varied situations, disagreements and judgement calls. Speaking to TIME about the Spirit acquisition, he said the purchase makes sense “if you believe that your agent is going to generalize into the rest of the economy.”

There is also a quality argument. In a widely cited 2024 paper in Nature, Ilia Shumailov and colleagues showed that models trained recursively on the output of other models suffer “model collapse”, progressively losing the rarer parts of the original data distribution. They concluded that data about genuine human interactions will become increasingly valuable as AI-generated content floods the web. A repository written by humans, before coding assistants were commonplace, is that kind of data.

A Market Assembled in Under a Year

The supply side organised itself quickly. Startups fail constantly; MIT Technology Review cited estimates from Carta and AngelList showing shutdowns rising by between 26 and 56 per cent a year recently. An “end-of-life” industry has grown up around them, and part of that industry has discovered that the corpse has value.

SimpleClosure, a service that handles the legal and operational wind-down of startups, added data sales in April 2026 through a platform called Asset Hub, where closing companies can list code repositories, Slack archives, emails and internal documents. In April, Forbes reported that the company had handled nearly 100 such deals, with sellers typically receiving between $10,000 and $100,000; heise online, summarising that report, quoted chief executive Dori Yona describing “a feeling of a gold rush from these companies trying to get their hands on real-world data”. By October, MIT Technology Review reported more than 150 deals, an expectation of returning more than $10 million to startups by the end of 2026, and Yona's estimate that 47 per cent of SimpleClosure's deals now include code or data sales, once rare and “somewhat taboo”.

Competitors followed. Brendan Mahony, chief executive of Sunset, another wind-down firm, told MIT Technology Review that his company values assets by headcount, years in operation, data volume and how the team used its tools. “This operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton,” he said. Brokers now publish buying guides. One, run by Human Union Data, lists Project Lazarus as paying between $10,000 and $100,000 for legacy code; that figure is the broker's own claim, and Turing does not publish its prices. Dan Novaes, chief executive of the data broker Mode.Inc, told Business Insider, as reported by IBTimes UK in October, that “the demand is just astronomical.”

The buyers are foundation labs, vendors that build simulated workplaces for training, and established technology companies building their own models. What they want is completeness. The Human Union Data guide tells sellers that buyers are after “real production code, internal documentation, and development history”, listing “pull requests, design docs, incident notes” as the material that makes a package whole. The completeness that makes the data valuable is exactly what makes it legally and ethically fraught.

A Repository Is a Crowd, Not a Possession

Start with the most basic question: who owns a git history?

The instinctive answer is the company. In many cases that instinct is broadly right, but only broadly. Under section 11 of the UK's Copyright, Designs and Patents Act 1988, “The author of a work is the first owner of any copyright in it,” with an exception: where a work is made by an employee in the course of employment, “his employer is the first owner of any copyright in the work subject to any agreement to the contrary.” The United States reaches a similar destination through the “work made for hire” doctrine, which covers “a work prepared by an employee within the scope of his or her employment.”

Now look at who actually appears in the commit log of a typical startup. There are employees, certainly. There are also freelancers and agencies, whose work is not automatically owned by the client in either jurisdiction; in the United States, commissioned software does not fall within the nine statutory categories that can be made works for hire by written agreement, so ownership passes only through an explicit assignment. There are co-founders who wrote early code before the company was incorporated. There are interns, advisers and contributors from partner firms. A repository that has run for five years may contain the work of dozens of people under a dozen different contractual arrangements, some of which were never written down.

Then there is everyone else's code. Private repositories are full of open-source components, vendored libraries and snippets copied from elsewhere, each carrying its own licence. Black Duck's 2026 Open Source Security and Risk Analysis report, which audited 947 commercial codebases across 17 industries, found licence conflicts in 68 per cent of them, up from 56 per cent the year before; one codebase contained 2,675 distinct conflicts. The report links part of that rise to AI coding assistants reproducing material derived from copyleft code without the original licence information, a phenomenon it calls licence laundering. A seller who cannot say which open-source licences govern the files in a repository is in no position to warrant that those files can be licensed onwards for training.

Open-source projects have built some machinery for this problem. Many use contributor licence agreements or the Developer Certificate of Origin, a short statement in which a contributor certifies that “The contribution was created in whole or in part by me and I have the right to submit it under the open source license”. Private companies rarely have anything comparable. The rights in their repositories are inferred from employment contracts, many of which predate the idea that internal engineering discussion could become a product.

Nor is copyright the only right at stake. Review threads and commit messages are written conversation: explanation, critique, sometimes humour or irritation. An employer may own the text, but that does not settle whether selling it is fair to the people who wrote it, or lawful under data protection law.

Possession Is Not the Same as Ownership

The Spirit Airlines auction has become the first public test of these questions. Mercor, an AI data company, was named backup bidder at $7.5 million, and another firm, micro1, later filed notice of a $12.5 million offer. The approval hearing was adjourned repeatedly through the autumn before being set for 17 November, and the objections filed in the bankruptcy court in White Plains read like a checklist of what can go wrong.

Springshot, a software company that ran operational tooling for Spirit, objected that the dataset could include its own intellectual property. Its filing cites a contract clause giving it rights to “any data or information generated by the Service or Software.” Chief executive Doug Kreuzkamp put the distinction plainly, according to an account published by The Next Web: “The distinction is between the information Spirit supplied and the proprietary outputs our platform generated.” Explaining why the company had not objected earlier, he told the Miami Herald, as quoted in the same account, “we had no reason to believe that Spirit would sell property that wasn't theirs.”

International Aero Engines, the engine manufacturer, filed its own objection arguing that its proprietary commercial, technical and financial information may have been copied into Spirit's internal communications, and that, in the words of its filing reported by ComplexDiscovery, “deidentification is insufficient to ensure that the IAE Proprietary Information has been removed.” Navitaire, Spirit's long-standing reservations platform provider, joined with Amadeus in a limited objection.

These objections expose the central weakness of the market: a company's systems hold far more than the company owns. Repositories routinely contain material supplied under non-disclosure agreements, vendor documentation pasted into tickets, customer configurations and code written under licence. In a working company, contractual confidentiality keeps that material in place. In a sale, it travels with everything else unless someone takes it out.

That matters to buyers, not just sellers. The EU's Trade Secrets Directive of 2016 makes the acquisition or use of a trade secret unlawful when the person “knew or ought, under the circumstances, to have known that the trade secret had been obtained directly or indirectly” from someone using or disclosing it unlawfully. A buyer who acquires a repository full of third-party confidential material without asking reasonable questions may not be able to claim innocence later.

There is now a striking precedent on the cost of careless acquisition. In Bartz v Anthropic, Judge William Alsup held in 2025 that training on lawfully acquired books could be fair use, but sent the claims over pirated library copies towards trial. Anthropic settled for $1.5 billion, covering roughly 480,000 works, and the settlement received final approval in July 2026, according to Jurist. The lesson for the code market is not that training is unlawful. It is that how training data was obtained can matter enormously, and that a court may look past the model to the provenance of the inputs.

The People Who Live Inside the Diff

Every commit in git carries an author name, an email address and a timestamp. Under Article 4(1) of the EU's General Data Protection Regulation, personal data means “any information relating to an identified or identifiable natural person”, and an identifiable person is one who can be identified “directly or indirectly”, including by a name or an online identifier. A git history is therefore, among other things, a structured personal data set: a record of when named individuals worked, what they worked on and, in review threads, what their colleagues thought of their work.

That last part deserves attention. Code review is one of the most socially charged activities in software engineering. A 2020 study by Carolyn Egelman, Emerson Murphy-Hill and colleagues at Google, presented at the International Conference on Software Engineering, surveyed 1,317 developers about negative experiences in review, which the authors called “pushback”. They found such experiences relatively rare but harmful when they occurred. A sold repository carries all of that: the blunt rejection, the apologetic reply, the thread in which someone was told they were wrong. These people wrote for colleagues, not for a model.

European regulators have started to address the training side. In Opinion 28/2024, adopted in December 2024, the European Data Protection Board said that whether an AI model can be considered anonymous must be assessed case by case, set out a three-step test for relying on legitimate interests, and indicated that unlawful processing at the development stage may affect the lawfulness of later deployment, with remedies that could in serious cases include deletion of the data or the model. The opinion did not contemplate a secondary market in staff conversations, but its logic applies.

American law offers less. Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, told MIT Technology Review that “Employers should have express consent before selling workers' data,” noting that no federal law bans such sales. She described the proceeds, which typically flow to creditors and investors rather than to the workers whose words are sold, as “captured capital”. Greg Ewing, an attorney who advises AI startups, called such data “legally radioactive” and described “a Wild West of AI training”.

Sellers and brokers insist that personal information is scrubbed. SimpleClosure says it removes personal data before sale. The Spirit deal shows what a serious attempt at scrubbing looks like: according to chapter11cases.com, the parties adopted the California Consumer Privacy Act's de-identification standard, appointed Tonic.ai as an independent de-identification agent for unstructured data such as email, made certified de-identification a condition of closing, required re-identification testing before each transfer, and obliged the buyer to commit publicly not to attempt re-identification. Those protections originally covered customers rather than staff. After unions led by the Association of Flight Attendants objected, Skift reported in November that revised terms require Google, still paying $10 million, to cover the cost of de-identification as well and to commit publicly not to knowingly re-identify the data, while employee medical examinations, payroll records, drug testing, crew pairings, biometrics and union grievance filings are excluded from the sale altogether. “Without our union, there would have been no one to speak for workers in this fire sale of their private information,” said Sara Nelson, the union's president.

Even that may not be enough. In a letter dated 8 October 2026, more than a hundred members of Congress led by Senator Elizabeth Warren and Representative Steven Horsford warned that “Removing names, email addresses, or other direct identifiers does not necessarily make a dataset anonymous.” The point applies with special force to code. Writing style, working hours, the modules a person owned and the distinctive way they phrase a review comment can identify them as surely as a name. And a git history without its authorship is a much less useful training asset, which creates a quiet commercial pressure against thorough scrubbing.

Secrets That Travel With the Code

The security problem is more concrete still. Developers routinely commit things they should not: API keys, database passwords, cloud credentials, private certificates. GitGuardian's State of Secrets Sprawl 2026 report, published in March, counted about 28.65 million new hardcoded secrets added to public GitHub commits in 2025, a 34 per cent increase on the year before. Internal repositories, the kind now being sold, were roughly six times more likely than public ones to contain hardcoded secrets. More than 64 per cent of secrets that GitGuardian confirmed as valid in 2022 were still valid when retested in January 2026. And about 28 per cent of incidents originated entirely outside repositories, in collaboration tools such as Slack, Jira and Confluence, precisely the systems now sold alongside code.

Deleting a secret from the current version of a file does not remove it from git. It persists in history unless that history is rewritten. A buyer who values complete history is therefore buying, by design, every credential anyone ever committed and later regretted. Some will have been rotated. Many, on GitGuardian's evidence, will not.

The risk does not end at transfer. Language models can memorise and regurgitate their training data. In a 2021 paper presented at the USENIX Security Symposium, Nicholas Carlini and colleagues extracted hundreds of verbatim sequences from GPT-2's training data, including personally identifiable information and code, and found that larger models were more vulnerable. A credential that survives into a training corpus could, in principle, resurface in a model's output. A dead startup's cloud account may be closed; its partners' shared keys may not be.

Provenance Tools Built for a Different Problem

If the market is to function legitimately, buyers need something like a chain of title for code: evidence of who wrote what, under which agreement, and with whose permission it is being sold. Some of the tools exist. None was designed for this job.

The most rigorous work on dataset provenance comes from the Data Provenance Initiative, a collaboration led by researchers including Shayne Longpre at MIT. Its audit of more than 1,800 text datasets, published in Nature Machine Intelligence in 2024, found licence omission rates above 70 per cent and error rates above 50 per cent on popular dataset hosting sites. If carefully curated public datasets are that poorly documented, the documentation that comes with a hastily sold startup repository is unlikely to be better. Industry groups have tried to standardise the metadata: the Data & Trust Alliance, a consortium of large companies, published data provenance standards covering source, legal rights, lineage and intended use. Such standards help only if someone fills them in truthfully.

Git itself offers little assurance. Its documentation is candid: the author name “has no effect on authentication”, and options such as —author and —date allow anyone to set the claimed author and timestamp of a commit to whatever they like. A git history is a record of what someone typed into a configuration file, not a verified log of who did what and when.

Cryptographic signing can change that. Sigstore's gitsign tool lets developers sign commits using an OpenID Connect identity rather than long-lived keys, recording the signing event in Rekor, a public transparency log, so that the signature can later be checked with git verify-commit. The Supply-chain Levels for Software Artifacts framework, known as SLSA and part of the Open Source Security Foundation, describes how to generate verifiable provenance for build artefacts. Software bills of materials, which the US Cybersecurity and Infrastructure Security Agency describes as “a nested inventory, a list of ingredients that make up software components”, can document which third-party components a codebase contains.

The trouble is timing. These tools establish provenance going forward. They work when commits are signed at the moment of creation and builds are attested as they run. Almost none of the repositories now being sold were built that way. A five-year-old startup history consists overwhelmingly of unsigned commits, and no amount of tooling applied at the point of sale can retroactively prove that a commit dated 2021 was really written in 2021 by the person named on it. SBOMs describe components, not the employment status of the people who wrote the glue code. Even the media world's answer, the content credentials standardised by the Coalition for Content Provenance and Authenticity, attaches provenance to finished objects rather than to an evolving, many-authored record.

What Brussels Now Asks to See

Regulation is starting to demand some of this documentation. Article 53(1)(d) of the EU AI Act requires providers of general-purpose AI models to publish a summary of the content used to train them. In July 2025, the European Commission's AI Office published a mandatory template for that summary. According to the Commission's own guidance, it requires providers to list their data sources, including private datasets licensed from third parties, and to describe relevant processing aspects such as respect for copyright. The obligation has applied since 2 August 2025 for new models; models already on the market before that date have until 2 August 2027. The AI Office's enforcement powers began on 2 August 2026, with fines of up to 3 per cent of worldwide annual turnover or 15 million euros, whichever is higher.

That is real pressure: a lab that buys a startup's codebase must at least acknowledge publicly that it trained on licensed private data. But the template has limits. The Commission has said it will not audit content at the level of individual datasets, acting instead on complaints or qualified alerts from its scientific panel. And a summary of sources is not a chain of title. A provider can accurately state that it used “private datasets licensed from third parties” while knowing nothing about whether the third party had the right to license them.

Incentives That Reward Forgery

The most troubling feature of this market is not any single legal gap but the way money changes behaviour around them.

Begin with what buyers prize. Sunset values assets partly by headcount and years in operation; broker guidance stresses development history and pull requests; Turing wants the “complete operational history” of a company. Buyers want long, busy, human histories because those are what make a codebase useful for training agents. Those are exactly the properties that git makes trivially easy to fake. Dates can be backdated with an environment variable. Author names can be set at will. Language models can generate plausible commit messages, review comments and tickets in volume. A fabricated five-year history, populated with synthetic engineers who argue convincingly in pull requests, is now a technically modest undertaking.

There is evidence that metadata on code platforms is already gamed when it has value. A study by Hao He, Bogdan Vasilescu, Christian Kästner and colleagues at Carnegie Mellon University, North Carolina State University and the security firm Socket identified around 4.5 million suspected fake stars on GitHub, with fake-star campaigns increasing by two orders of magnitude in 2024, largely to promote short-lived malicious repositories. Stars are a crude signal. Commit histories are a far richer one, and they now carry a price tag. If buyers pay more for histories that look human and long, some sellers will manufacture them, producing exactly the synthetic contamination that real-world data was supposed to avoid. MIT Technology Review noted that no company has published tests comparing training on real archives with the alternatives, so buyers may struggle to tell the difference.

The second incentive is laundering. A repository containing third-party code, an ex-employer's proprietary modules or material taken under NDA becomes harder to trace once it passes through a broker, is scrubbed of identifying metadata, and is merged into a training corpus with thousands of others. De-identification, presented as a privacy safeguard, also strips away the authorship information that might reveal where code came from. No such case has yet been documented, but the structure of the market rewards anyone who can present tainted material as clean.

The third incentive is insider exfiltration. Developers routinely have access to code they do not own. Until recently, the main buyers for a stolen codebase were competitors and foreign rivals, and the cases that reached court were dramatic. In January 2026, a federal jury in San Francisco convicted Linwei Ding, a former Google software engineer, of seven counts of economic espionage and seven counts of theft of trade secrets, in a case in which prosecutors said he had stolen more than 2,000 pages of confidential material about Google's AI infrastructure. That was a case about geopolitics. The new market creates a far more mundane temptation: a departing engineer, a disgruntled contractor or a founder in a dispute with co-founders may now find a ready, apparently respectable buyer for a private repository.

That is not hypothetical. According to IBTimes UK, citing Business Insider, Surge AI's Nick Heiner said someone had approached him offering a large company's codebase for millions of dollars, and he concluded that the seller did not have the rights to sell it. A less careful buyer, or one under competitive pressure to acquire data faster than its rivals, might not have asked.

The final incentive concerns who gets paid. In the startup wind-downs described by MIT Technology Review, proceeds typically go to senior creditors and investors. The engineers whose judgement, arguments and late-night fixes make the data valuable receive nothing and are usually not asked. Ryan Lock, co-founder of Polyshares, compared the offers to those of early oil prospectors buying land from unaware farmers. People with no stake in a transaction also have no channel through which to object, so the safeguard most likely to catch a bad sale, the knowledge of those who were actually there, is switched off.

Building a Chain of Title for Code

None of this means the trade must stop. A failing company has legitimate reasons to recover value from its work, and AI developers have genuine reasons to want realistic engineering data. The question is what a legitimate transaction would look like. The Spirit proceedings, existing law and the tools described above suggest a working standard.

Sellers should map every author in the history to an employment contract, a contractor agreement with an assignment clause or a documented licence, and remove whatever cannot be matched. The labour of that mapping is the provenance. Former employees and contributors should be notified and given time to object; the Spirit case surfaced vendors, unions and lawmakers only because a bankruptcy court made the sale public, and private startup deals have no equivalent forum. Vendored code, NDA material and partner data should be found through software composition analysis and contract review, then stripped out or separately licensed.

De-identification should follow the Spirit model of an independent agent, certification and re-identification testing before transfer, and should address the stylistic signals that identify developers even without their names. Every commit and every attached collaboration archive should be scanned for credentials, with live secrets revoked rather than merely redacted. Buyers should demand warranties of title backed by indemnities and escrow, so that the cost of a bad sale falls on whoever was best placed to prevent it. And companies that want their engineering record to be worth something later should sign commits and attest builds now, which would also make fabricated histories easier to detect.

The Record Belongs to More Than One Person

Return to that email from Google: “You keep 100% of your IP.” For a solo developer who wrote every line of an app, that may be close to true. For almost everyone else, it is a simplification that the new market quietly depends on. A codebase with a real history is a crowd of authors, a web of contracts, a store of other companies' secrets and a diary of working lives. It can be sold as a single asset only by ignoring most of what makes it valuable.

The AI industry has discovered that the most useful training data is not polished output but the visible process of people working things out together. That process was never created with a buyer in mind. If the market is allowed to treat possession as ownership, and a git log as proof of anything, it will reward forgers, launderers and insiders. At least one buyer has already recognised an offer that looked too good to be legitimate. A market cannot rely on its buyers happening to be careful.

References and Sources

  1. Koebler, Jason. “Google Is Quietly Buying Code From Play Store Developers to Train AI.” 404 Media, 2 June 2026. https://404media.co/google-is-quietly-buying-code-from-play-store-developers-to-train-ai/
  2. King, Matthew. “AI labs are buying data from dead startups to train their models.” MIT Technology Review, 13 October 2026. https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/
  3. Perrigo, Billy. “Google's $10 Million Bid for Spirit Airlines' Data Reveals AI's Next Frontier.” TIME, 25 August 2026. https://time.com/article/2026/08/25/google-spirit-airlines-ai-data-RL/
  4. Chapter11Cases. “Spirit Airlines Case Update: The Data Sale Slips to November as Spirit Moves to Unwind the HFS Aircraft Deal.” 14 October 2026. https://chapter11cases.com/blogs/news/spirit-airlines-case-update-the-data-sale-slips-to-november-as-spirit-moves-to-unwind-the-hfs-aircraft-deal
  5. ComplexDiscovery. “Lawmakers press Google and Spirit on employee data before a hearing on a $10 million AI training data sale.” 9 October 2026. https://complexdiscovery.com/lawmakers-press-google-and-spirit-on-employee-data-before-a-hearing-on-a-10-million-ai-training-data-sale/
  6. Adebayo, Kolawole Samuel. “Springshot says Google's $10m Spirit deal could include its trade secrets.” The Next Web, 18 September 2026. https://thenextweb.com/news/springshot-objection-spirit-airlines-data-google
  7. Jaen, Regyl. “Companies Are Selling Employees' Private Slack Messages and Emails to Train AI, Often Without Workers' Knowledge.” IBTimes UK, 20 October 2026 (reporting remarks given to Business Insider). https://www.ibtimes.co.uk/companies-sell-workplace-data-ai-firms-1826648
  8. Holland, Martin. “Lucrative internal data: AI firms buy internal data from liquidated companies.” heise online, 17 April 2026 (summarising Forbes reporting of 16 April 2026). https://heise.de/-11262202
  9. Turing. “Can an AI Learn From a Company That No Longer Exists?” AGI Advance newsletter, 24 December 2025. https://www.turing.com/blog/agi-advance-newsletter-44
  10. Human Union Data, Inc. “Best Platforms for Selling Your Codebase in 2026.” HUD, 24 June 2026 (broker marketing material). https://www.hud.ai/guides/best-platforms-selling-codebase
  11. Villalobos, Pablo; Ho, Anson; Sevilla, Jaime; Besiroglu, Tamay; Heim, Lennart; Hobbhahn, Marius. “Will we run out of data? Limits of LLM scaling based on human-generated data.” Epoch AI, 2024. https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
  12. Shumailov, Ilia; Shumaylov, Zakhar; Zhao, Yiren; Papernot, Nicolas; Anderson, Ross; Gal, Yarin. “AI models collapse when trained on recursively generated data.” Nature, 631, 755-759, 24 July 2024. DOI: 10.1038/s41586-024-07566-y.
  13. UK Parliament. Copyright, Designs and Patents Act 1988, section 11. https://www.legislation.gov.uk/ukpga/1988/48/section/11
  14. United States Code. 17 U.S.C. section 101, definition of “work made for hire.” Legal Information Institute, Cornell Law School. https://www.law.cornell.edu/uscode/text/17/101
  15. Hamilton, Corey. “Software governance in the AI era: Key findings from the 2026 OSSRA report.” Black Duck, 25 February 2026. https://www.blackduck.com/blog/open-source-trends-ossra-report.html
  16. Linux Foundation. Developer Certificate of Origin, Version 1.1. https://developercertificate.org/
  17. European Parliament and Council. Directive (EU) 2016/943 on the protection of undisclosed know-how and business information (trade secrets), Article 4. https://www.legislation.gov.uk/eudr/2016/943/article/4
  18. Jurist. “Judge approves record $1.5 billion AI copyright settlement involving Anthropic.” July 2026. https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/
  19. European Parliament and Council. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 4. https://gdpr-info.eu/art-4-gdpr/
  20. Egelman, Carolyn D.; Murphy-Hill, Emerson; Kammer, Elizabeth; Hodges, Margaret Morrow; Green, Collin; Jaspan, Ciera; Lin, James. “Predicting Developers' Negative Feelings about Code Review.” Proceedings of the 42nd International Conference on Software Engineering (ICSE 2020), 174-185. DOI: 10.1145/3377811.3380414. https://research.google/pubs/predicting-developers-negative-feelings-about-code-review/
  21. European Data Protection Board. Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, adopted December 2024. Summary: CMS. https://cms.law/en/int/legal-updates/edpb-opinion-28-2024-key-takeaways-on-processing-personal-data-in-the-context-of-ai-models
  22. GitGuardian. “The State of Secrets Sprawl 2026.” 17 March 2026. https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/
  23. Carlini, Nicholas; Tramèr, Florian; Wallace, Eric; Jagielski, Matthew; Herbert-Voss, Ariel; Lee, Katherine; Roberts, Adam; Brown, Tom; Song, Dawn; Erlingsson, Úlfar; Oprea, Alina; Raffel, Colin. “Extracting Training Data from Large Language Models.” 30th USENIX Security Symposium, 2021, 2633-2650. https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting
  24. Longpre, Shayne, et al. “A large-scale audit of dataset licensing and attribution in AI.” Nature Machine Intelligence, 6, 975-987, 2024. MIT News coverage: https://www.media.mit.edu/articles/study-transparency-is-often-lacking-in-datasets-used-to-train-large-language-models/
  25. Help Net Security. “Cross-industry standards for data provenance in AI” (interview with the Data & Trust Alliance). 22 July 2024. https://www.helpnetsecurity.com/2024/07/22/saira-jesani-data-trust-alliance-data-provenance-standards/
  26. Git project. “git-commit Documentation.” https://git-scm.com/docs/git-commit
  27. Sigstore. “Gitsign” documentation. https://docs.sigstore.dev/cosign/signing/gitsign
  28. SLSA (Supply-chain Levels for Software Artifacts), Open Source Security Foundation. https://slsa.dev/
  29. Cybersecurity and Infrastructure Security Agency. “Software Bill of Materials (SBOM).” https://www.cisa.gov/sbom
  30. Coalition for Content Provenance and Authenticity. “About C2PA.” https://c2pa.org/about/
  31. European Commission. “Template for General-Purpose AI model providers to summarise their training content” (FAQ). https://digital-strategy.ec.europa.eu/en/faqs/template-general-purpose-ai-model-providers-summarise-their-training-content
  32. He, Hao; Yang, Haoqin; Burckhardt, Philipp; Kapravelos, Alexandros; Vasilescu, Bogdan; Kästner, Christian. “4.5 Million (Suspected) Fake Stars in GitHub: A Growing Spiral of Popularity Contests, Scams, and Malware.” arXiv preprint 2412.13459, December 2024. https://arxiv.org/abs/2412.13459
  33. CBS News Bay Area. “Former Google engineer Linwei Ding found guilty of stealing AI secrets for Chinese firms.” 30 January 2026. https://www.cbsnews.com/sanfrancisco/news/former-google-engineer-linwei-ding-found-guilty-stealing-ai-secrets-chinese-firms/
  34. GitHub. “Updates to our Privacy Statement and Terms of Service: How we use your data.” GitHub Changelog, 25 March 2026. https://github.blog/changelog/2026-03-25-updates-to-our-privacy-statement-and-terms-of-service-how-we-use-your-data/
  35. Maharishi, Meghna. “Spirit's $10 Million Data Sale to Google Clears Hurdle as Unions Win Privacy Protections.” Skift, 10 November 2026. https://skift.com/2026/11/10/spirits-10-million-dollar-data-sale-clears-hurdle-as-unions-win-privacy-protections

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...