Context Poisoning

Signs it might be this

• Edit histories, logs or forensic review show a deliberate insertion that serves an identifiable interest.
• The same false detail appears in unrelated places soon after one origin.
• Text intended for automated readers is found hidden in a page, file or message.
• Citations do not support the claim, or point to pages that appear to have been created for this purpose.
• Security researchers or the platform have reported tampering with a specific dataset or source.

Signs it is probably something else

• The error appears once, in one place, and is fixed when pointed out.
• The wrong information is old, and was once correct.
• Different sources disagree because the evidence itself is unclear.
• The only sign is that an AI tool gave a wrong answer, which can have many causes.
• You cannot trace who edited or posted the material, and are guessing at motive.
• The error gets corrected quickly once it is reported, with a visible record.
• The claim is easy to cross-check against a primary source you can reach yourself.
• The mistake is minor and affects nothing you are deciding.

Questions to ask yourself

• Where did this claim first appear, and does the original source actually say it?
• Do the supposedly separate sources trace back to a single origin?
• Is there an edit history or log that shows when and by whom the content changed?
• Would I accept this claim if I had to defend it with a primary source?
• Am I relying on one tool to check another that learned from the same material?

What it is

Context poisoning means slipping false, skewed or booby-trapped content into places that people and software draw on. These include wikis, reviews, shared documents, search results, databases, and the material an AI system reads or is trained on.

Shared information spaces also fill with ordinary error, stale pages and honest disagreement. Deliberate insertion is a claim that needs forensic evidence, and finding something wrong is not enough to support it.

Why it matters

Decisions get made on a corrupted record. The corruption sits in the source, so both the person and the tools they use to double-check return the same wrong answer.

That makes the error hard to catch.

How it shows up

• False or distorted entries placed in reference pages, review sections, shared files or search-indexed pages.
• Material that targets the sources AI systems and search tools draw on.
• Content that tries to steer automated readers as well as human ones, which a human reader may never notice.
• It works well because the copy spreads. Once the material is cited, reposted or summarized, the original source of the error is buried.
• Different from a one-off lie in that it targets the shared source rather than a single conversation.

What you can do

Pick what fits. You do not owe anyone a confrontation.

A check to run before you share or act

“Everything I can find says the same thing, but I have not yet seen where it began. Do the sources really agree, or are they one claim repeated, and what does the original document actually say?”

Ask about it gently

Trace the claim back to its first appearance and compare it with a primary source before treating it as settled.

Set a clear limit

Do not forward, cite or act on a claim until you have found support that does not trace back to the same origin.

Step back

Step back from a source that keeps resurfacing a claim you cannot verify, and use another one you can check yourself.

Get support

Report suspected tampering to the page’s maintainers or the platform. Share what you found with a librarian, journalist or fact-checking service.

At work, school or with services

If you manage data, documents or AI systems at work, log the suspicious content and restrict what automated tools may act on. Then escalate to your security or data team.

Using this label fairly

• Calling an inconvenient source ‘poisoned’ without evidence is a way to dismiss facts one dislikes.
• Treating every AI mistake as sabotage hides the more common causes of error.

How it differs from look-alikes

Algorithmic Manipulation leaves the content alone and changes what gets shown and in what order. Context Poisoning changes the content or the data itself.

Synthetic Persona Networks create fake voices that speak. Here the damage is placed in a source that people and tools later consult.

AI-Driven Social Engineering targets one person’s action, but this contaminates what many will read.