Direct answer
Source or data poisoning contaminates the information environment from which a system learns or retrieves with misleading content. Many pages copying one another can create the appearance of a consensus that does not exist. NOMOS GEO calls this 'synthetic consensus'. A weak claim made to look credible through copies and citations is 'source laundering'. Audit the earliest source, independence, change history and whether each source truly supports the claim—not the page count.
In plain language
Imagine photocopying one false rumour a hundred times. There are a hundred sheets on the table, but not a hundred independent witnesses. If an AI system mistakes copies of one sentence for independent confirmations, it sees a false majority.
Why this matters
Poisoned information may cause the wrong entity match, a faulty recommendation, fabricated reputation or the disappearance of genuine sources. The error may look trustworthy and persist because the number of copies is high.
Do not confuse
- Data poisoning intentionally corrupts training, fine-tuning or vector-store data.
- Source poisoning contaminates a retrieval or research surface with misleading content.
- Synthetic consensus makes non-independent repetition look like independent agreement.
- Source laundering makes a weak or invented claim look strong through a chain of copies and citations.
- False information is not always an attack; when intent cannot be proved, intent remains UNRESOLVED.
What should you do?
- For every material claim, identify the earliest source that can be found and the later copying relationships.
- Create provenance clusters using text similarity, publication time, ownership, links and citation traces.
- Measure independent origin, authority, recency and direct claim support rather than URL count.
- Quarantine suspect sources before deletion and stop their influence on retrieval.
- Reverify the claim against an official record, primary document or genuinely independent source.
- Identify and rebuild affected indexes, vector records, caches and derived records in a controlled way.
- After correction, retest with the same question and source set.
How do you audit it?
- How many sources have a genuinely independent origin?
- Does the earliest source prove the claim or merely assert it?
- Were copies or circular citations counted as independent verification?
- Were source ownership, publication date and change history preserved?
- Which models, indexes, vector records and outputs did the suspect content affect?
- Was the correction made only on the visible page or across every derived surface?
- Was malicious intent stated as certain without evidence?
Limit
High textual similarity alone is not proof of copying or malicious intent; independent sources may describe the same fact similarly. Conversely, source diversity alone does not establish truth. Where intent or provenance is uncertain, keep the result explicitly UNRESOLVED.
Remember in one sentence
One hundred copies are not one hundred independent items of evidence.
Sources for this record
- S35W3C, *PROV-O: The PROV Ontology*Standard
- S39NIST AI 100-2 E2025, *Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations*Voluntary standards-oriented institutional guidance
- S41NCSC and partner agencies, *Guidelines for Secure AI System Development*Official multi-agency security guidance
- S43NIST SP 800-161 Rev. 1 Update 1, *Cybersecurity Supply Chain Risk Management Practices for Systems and Organizations*Voluntary standards-oriented institutional guidance
- S44NIST, *AI Risk Management Framework Playbook* and AI RMF CoreVoluntary implementation guidance