Source Contamination

Where can contamination enter?

Contamination can enter at two stages, and the response differs. The table compares them.

Stage	What happens	What a brand can do
Training data	Bad content published on the open web is collected into the data used to train a model	Little directly. Publish accurate, well-sourced material and report clear abuse to the platforms involved
Retrieval at answer time	A search-connected assistant pulls in a wrong, outdated or manipulated page and uses it	Find the page, correct or request correction at the source, and strengthen accurate pages

Not all wrong answers come from contamination. Models can also make errors on their own, and outdated pages can be accurate for their time.

What does the research show?

Researchers have shown that manipulating the open web can influence AI systems, though the findings are narrow. A joint study published in October 2025 by Anthropic, the UK AI Security Institute and the Alan Turing Institute found that about 250 poisoned documents were enough to implant a backdoor in models of several sizes up to 13 billion parameters, regardless of how much clean data they were trained on. The test used a narrow, low-risk behavior, producing gibberish after a trigger phrase, and its authors noted caveats. A report from the Atlantic Council's DFRLab in April 2026 described how a pro-Kremlin network of sites that publishes large volumes of content appears in a major public web archive used in AI training, a tactic sometimes called "LLM grooming". These are early findings, and how much real-world answers are affected is still being studied.

How do you find contaminated sources?

Start from the wrong answer and trace back to its sources. A practical sequence:

1. Test a fixed set of questions about the brand across the main assistants, and record every wrong statement. See prompt audit.
2. Note the sources each assistant cites or appears to rely on.
3. Check each source for errors, age, manipulation and who runs it. See source quality.
4. Look for patterns, such as many low-quality pages repeating the same claim.
5. Monitor new mentions, since contamination can appear after a news event. See media monitoring.

Illustrative contamination log (example only)

• Wrong claim, engine and date
• Source page and who owns it
• Type: outdated, honest error, copied claim or deliberate manipulation
• Action taken and result at re-test

How should a brand respond?

1. Fix what you control: correct your own pages, profiles and press materials, and keep facts consistent. See entity home.
2. Ask third-party sources to correct clear errors, with evidence, using their published correction process.
3. Publish an accurate, well-sourced page that answers the disputed question directly, so there is a strong source to retrieve.
4. Report deliberate manipulation or defamation to the platform or host, and take legal advice where harm is serious.
5. Use the feedback tools the AI products provide, and re-test later, since changes take time.
6. Prepare statements for serious cases. See crisis communications.

What are the limits?

A brand cannot edit what a model has already learned, and it cannot control which sources an assistant retrieves. Correction at the source can help, but engines update on their own schedules, and results are not guaranteed. Avoid fighting contamination with more manipulation, such as creating bulk content to drown out a claim, since that can break platform rules and add to the problem.