The Evidence Lineage

The Evidence Chain, Part 9 of 10. Part 7 showed how evidence gets contaminated, and Part 8 how contaminated evidence can agree with itself. This part asks what it would take to see through both, for one consequential decision.

On 10 April 2026, South Africa published its draft National Artificial Intelligence Policy in the Government Gazette for public comment (Department of Communications and Digital Technologies [DCDT], 2026). Sixteen days later the minister responsible, Solly Malatsi, withdrew it. The reason was not a political row or a change of strategy. The reference list contained sources that do not exist. “The most plausible explanation,” the minister said, “is that AI-generated citations were included without proper verification,” and he added that this was not a mere technical issue but had compromised the integrity and credibility of the draft (South African Government News Agency [SAnews], 2026). A country’s plan for governing artificial intelligence had been undone by the thing it set out to govern.

It would be easy to laugh, and wrong. The officials who drafted that policy did what almost everyone reading this now does at work: they asked a machine to help, it handed back something fluent and well referenced, and nobody walked each reference back to the shelf. Here is the uncomfortable part for you and me. A reference list is the only lineage most documents ever carry. It is the trail that says where a claim came from. When the trail itself can be generated, a document can look perfectly sourced and rest on nothing at all.

From “show me the log” to “show me the lineage”

Banks learned this lesson the hard way. After the 2008 crisis, supervisors found that many large banks could not quickly and accurately add up their own exposures, because the numbers in their risk reports had passed through so many systems that nobody could trace them to their origin. The Basel Committee’s 2013 principles, known as BCBS 239, required banks to aggregate risk data accurately and to understand how it flows from source to report (Basel Committee on Banking Supervision [BCBS], 2013), and the European Central Bank now expects complete lineage for individual data attributes, from capture to report (European Central Bank [ECB], 2024). The rule underneath is simple: a number you cannot trace is a number you cannot trust, however authoritative the report it sits in.

An audit log tells you what a system did and when. Evidence lineage tells you what it relied on and where that came from, traced as far back as it goes: from the decision to the evidence used, from the evidence to its source, from the source to whether a person or a machine produced it, through every transformation, to an original observation of something that happened in the world. My argument is that AI decisions are heading for their own BCBS 239 moment, and that the question regulators, courts and readers ask is moving from “show me the audit log” to “show me the lineage.”

The forged trail is not a one-off

South Africa is the clearest public case, not the only one. Part 7 told the story of a consulting report for the Australian government whose invented quotation from a court judgment was spotted by a single academic reader (McGuirk, 2025). In May 2026 The Lancet published an audit, led by the Columbia University researcher Maxim Topaz, that screened millions of biomedical papers and some 97 million citations and found about 4,000 fabricated references. The share of papers with at least one went from one in 2,828 in 2023 to one in 458 in 2025, and to one in 277 in the first seven weeks of 2026 (Topaz et al., 2026; Oza, 2026). The absolute numbers are small. The direction is not. And the study’s origin story is the most honest description of the problem I have read: Topaz told STAT it began when a chatbot slipped a bad citation into his own editorial after he had already checked the references (Oza, 2026). Careful people, checking, still miss it.

Notice what caught each of these. Not the authors’ tools, not the publishers’ systems, not the commissioning department. A South African newsroom, an academic in Sydney, a journal editor, and a researcher who turned his own embarrassment into a census. Lineage, when it was checked at all, was checked by someone independent of the people who produced the document.

Why this is about power, not tidiness

It is tempting to file this under sloppiness. I think it is about who gets to make claims stick. A national AI policy decides which skills a country’s young people are trained in, which industries get support and how public services will use automated systems. Daily Maverick reported that the fabricated material touched three of the policy’s six pillars, including capacity and talent development (Schutters, 2026). When a document like that rests on sources no one can retrieve, the people it governs lose the ability to argue with its evidence, because there is no evidence to argue with. And the ability to check lineage is unevenly spread. Verifying references takes access to journals, archives and databases, which well-funded institutions in rich countries buy and many others cannot. A world in which AI makes documents cheap to produce and lineage expensive to check hands the advantage to whoever can afford the checking, or whoever can afford to skip it and absorb the embarrassment.

Most of the parts of a solution already exist in separate places. The World Wide Web Consortium standardized a general model for provenance in 2013, describing how things are derived from other things through activities carried out by people or software (Moreau & Missier, 2013). Content Credentials let media carry a signed record of its ingredients (Coalition for Content Provenance and Authenticity [C2PA], 2025). The EU AI Act requires providers of high-risk systems to document the origin of their training data (Regulation (EU) 2024/1689, art. 10). What does not exist is the connection between these pieces at the moment a specific decision is made, or a specific document is relied on.

The prediction

My prediction is that, for consequential AI-assisted decisions and documents in regulated and public settings, provenance obligations will extend from the output back through the chain of evidence transformations, following the path data lineage took in banking: first as a supervisory or editorial expectation, then as a condition of trusting an institution’s own reporting, and eventually as a normal feature of how serious systems are built. If you work anywhere that produces documents others rely on, the practical version arrives sooner: someone will ask you to show where a claim came from, and “the tool gave it to me” will not be an answer. The research question for this part is where to draw the line. What counts as sufficient lineage for a consequential decision?

For Mercury Intelligence members: the section below proposes a working answer, a proportionate lineage standard in three tiers, tests it against the South African withdrawal link by link, maps the rules clause by clause, and sets out a sampling method you can run this quarter.

---

Support independent investigation. Every fabricated trail in this part was caught by someone who went back to the shelf, and that work is slow, unglamorous and paid for by nobody who benefits from the shortcut. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron: fairbydesign.org/patrons-2.

References

Basel Committee on Banking Supervision. (2013). Principles for effective risk data aggregation and risk reporting (BCBS 239). Bank for International Settlements. https://www.bis.org/publ/bcbs239.htm

Basel Committee on Banking Supervision. (2023). Progress in adopting the Principles for effective risk data aggregation and risk reporting. Bank for International Settlements. https://www.bis.org/bcbs/publ/d559.htm

Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation, & Office of the Comptroller of the Currency. (2026, April 17). Revised guidance on model risk management (SR 26-2). https://www.federalreserve.gov/supervisionreg/srletters/SR2602.pdf

Coalition for Content Provenance and Authenticity. (2025). Content Credentials: C2PA technical specification (Version 2.2). https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html

Department of Communications and Digital Technologies. (2026, April 10). Draft South Africa National Artificial Intelligence (AI) Policy (General Notice 3880 of 2026). Government Gazette, No. 54477. https://www.gov.za/sites/default/files/gcis_document/202604/54477gen3880.pdf

European Central Bank. (2024). Guide on effective risk data aggregation and risk reporting. ECB Banking Supervision. https://www.bankingsupervision.europa.eu/ecb/pub/pdf/ssm.supervisory_guides240503_riskreporting.en.pdf

Hamilton v Post Office Ltd [2021] EWCA Crim 577. https://www.judiciary.uk/wp-content/uploads/2022/07/Hamilton-Others-v-Post-Office-judgment-230421.pdf

International Organization for Standardization & International Electrotechnical Commission. (2023). Information technology — Artificial intelligence — Management system (ISO/IEC Standard No. 42001:2023). https://www.iso.org/standard/42001

McGuirk, R. (2025, October 7). Deloitte to partially refund Australian government for report with apparent AI-generated errors. Associated Press (via WDIV ClickOnDetroit). https://www.clickondetroit.com/business/2025/10/07/deloitte-to-partially-refund-australian-government-for-report-with-apparent-ai-generated-errors/

Moreau, L., & Missier, P. (Eds.). (2013, April 30). PROV-DM: The PROV data model (W3C Recommendation). World Wide Web Consortium. https://www.w3.org/TR/2013/REC-prov-dm-20130430/

Oza, A. (2026, May 7). Fraudulent citations, blamed on AI hallucinations, are becoming more common in research papers. STAT. https://www.statnews.com/2026/05/07/lancet-study-finds-steep-rise-fraudulent-citations-academic-papers/

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union, L, 2024/1689. http://data.europa.eu/eli/reg/2024/1689/oj

Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI). (2026). Official Journal of the European Union, L, 2026/1744. http://data.europa.eu/eli/reg/2026/1744/oj

Schutters, L. (2026, April 27). Malatsi withdraws draft AI policy after hallucination revelations. Daily Maverick. https://www.dailymaverick.co.za/article/2026-04-27-malatsi-withdraws-draft-ai-policy-after-hallucination-revelations/

South African Government News Agency. (2026, April 26). Minister announces withdrawal of draft AI Policy. SAnews. https://www.sanews.gov.za/south-africa/minister-announces-withdrawal-draft-ai-policy

Topaz, M., Roguin, N., Gupta, P., Zhang, Z., & Peltonen, L.-M. (2026). Fabricated citations: An audit across 2·5 million biomedical papers. The Lancet, 407(10541), 1779–1781. https://doi.org/10.1016/S0140-6736(26)00603-3

What to keep, and why: Appendix B of Operational AI Readiness (free). To test whether your own records would let you rebuild an AI decision, take the Reconstruction Test.