When AI Gets the Source Wrong
AI systems do not only create risk when they produce incorrect facts. They also create risk when they separate information from its source, timestamp, jurisdiction,
or issuing authority.
Attribution failures are often invisible until something breaks. A document may look polished. A summary may sound confident. A citation may appear complete. But if the source cannot be verified, the authority cannot be confirmed, or the publication context is lost, trust begins to weaken.
This problem matters in government communication because public information depends on more than factual accuracy. It also depends on who issued the information, when it was issued, where it applies, and whether the source can still be verified.
Documented Examples of Source Failure
South Africa Withdraws Draft National AI Policy After Fabricated Citations
South Africa withdrew its draft national artificial intelligence policy after fictitious sources were discovered in the reference list. Reporting indicated that several academic citations appeared to reference papers, journals, or authors that could not be verified.
The issue was not simply a drafting mistake. It showed how official documents can lose credibility when references appear authoritative but cannot be traced back to real source material.
Failure type: Citation verification failure
- References appeared complete but could not be verified.
- The policy process was disrupted after publication.
- The incident became a public example of AI-related source failure inside government documentation.
Source: Public reporting from Rest of World and African Business.
Deloitte Australia Report Included Nonexistent References and an Invented Legal Quote
A Deloitte report prepared for the Australian government was revised after errors were identified, including references to nonexistent academic research and a quote attributed to a court judgment that did not exist. Deloitte agreed to provide a partial refund after the issues were found.
The report’s recommendations were not the only issue. The larger trust problem was that supporting citations and source references could not be relied upon without independent verification.
Failure type: Source attribution and verification failure
- Nonexistent sources were presented as supporting references.
- A legal quotation was attributed to a judgment that did not contain it.
- The government-facing report required revision after publication.
Source: Public reporting from The Guardian, Fortune, and Fast Company.
Deloitte Canada Report Raised Similar Concerns About Fabricated Research Citations
Reporting later identified another Deloitte government-facing report that allegedly contained false citations, including references to academic papers that did not exist and citations attributed to researchers who had not authored the claimed work.
These incidents point to a broader pattern: AI-assisted work can appear well-sourced while the underlying citations, authorship, and source trail may not withstand verification.
Failure type: Provenance and citation integrity failure
- Research support appeared authoritative but was reportedly unverifiable.
- Named researchers were associated with work they had not authored.
- The failure centered on the source trail behind the claims.
Source: Public reporting from Fortune and The Independent.
Attribution Can Break Even When the Information Is Not Fabricated
Not every attribution failure involves fake citations. In many cases, the information itself may be accurate, but the source context becomes unclear as AI systems summarize, reformat, combine, or redistribute public information.
An AI System Attributes a Statement to the Wrong Agency
A public statement may be accurate in general but attributed to the wrong department, office, city, county, or jurisdiction. In government communication, that distinction matters.
Failure type: Authority attribution failure
- Accountability becomes unclear.
- Residents may rely on the wrong issuing authority.
- Government agencies may be associated with statements they did not issue.
An AI System Surfaces Superseded Guidance
AI systems may retrieve or summarize older information when timestamps, version history, or update status are not explicit enough for machine interpretation.
Failure type: Timestamp and version failure
- Obsolete information may appear current.
- Corrections may occur after the outdated information has already spread.
- Public confidence can decline even when the original agency later clarifies the record.
Local Information Is Interpreted Through the Wrong Geographic Context
A city, county, state, or federal source may be summarized without enough jurisdictional context. The result can be an answer that sounds useful but applies to the wrong place.
Failure type: Jurisdictional attribution failure
- Residents may receive guidance intended for another jurisdiction.
- Local authority can be blurred by state or national context.
- AI systems may prioritize larger or better-indexed sources over the correct local source.
A Statement Moves Through Summaries Until the Original Source Disappears
Government information often moves from websites to PDFs, social posts, newsletters, vendor systems, news articles, search results, and AI-generated answers. Each step can weaken the connection to the original source.
Failure type: Attribution continuity failure
- The original issuing authority becomes harder to identify.
- Verification requires manual investigation.
- AI-generated summaries may preserve the claim while losing the source context.
The Problem Is Bigger Than Hallucination
AI hallucination is only one visible form of source failure. A broader problem emerges when information is separated from the authority, timestamp, jurisdiction, verification status, and provenance signals that make it trustworthy.
In public-sector communication, the question is not only whether a statement is true. The question is whether an AI system can reliably identify who issued it, when it was issued, where it applies, and whether it remains authoritative.
Why Neutral, Shared Infrastructure Matters
Attribution failures do not affect only one website, one vendor, or one agency. They occur across an ecosystem of public websites, publishing tools, notification systems, resident engagement platforms, search engines, and AI systems.
That is why source attribution increasingly requires neutral, shared infrastructure. A shared layer allows independent GovTech providers and verified government authorities to preserve attribution, provenance, timestamps, jurisdiction, and verification continuity without each organization building a separate trust system alone.
Aigistry operates as neutral, shared infrastructure for the National AI Feed, helping preserve the source context that AI systems need in order to interpret government information with greater attribution clarity.