16 Jun 2026

AI deep-research poisoning — decision integrity for SMEs (v1)

AI deep-research poisoning — decision integrity for SMEs (v1)

Your AI research tool may look and feel like an analyst. But it’s not.

Often, it is a fast, articulate reader of pages anyone can edit, including the people who would quite like to sell you something.

In tests on deep-research agents, thirteen words in the right Reddit-style comment were enough to get a company that did not exist into a report, cited and formatted, sitting comfortably beside real recommendations.

Researchers showed that a single line planted on a page that these tools repeatedly retrieve can steer the report across an entire topic.

In their tests, the invented brand appeared in a large share of reports once that poisoned source was picked up. The same method fabricated a cancellation service, a dating app and a restaurant, none of which existed.

You would not catch it by reading closely, because there is almost nothing to catch. The agent folds the planted line into its own generated prose, in the same measured voice as the genuine recommendations, and cites it as a source.

The planted line stops looking planted the moment the agent paraphrases it.

This hits hardest on the very questions businesses bring to these tools. Not the ones with a settled answer, but “which supplier?”, “is this firm trustworthy?”, “what do people recommend?” Those queries have no single authority and plenty of community chatter, which is exactly where a planted entry competes best.

None of this means abandon the tool. It means demoting it.

Use AI research as a starting point, not a decision authority. When a report names a vendor, service or product you are about to act on, confirm independently that the thing exists and is what it claims before money or trust moves.

That is the practical lesson. Not “never use AI for research”, because that would be silly. Use it. But do not confuse a polished report with a verified one.

So here is the question worth asking yourself before your next AI-assisted decision:

Are you reading research, or reading something that merely looks like it?

Originally published on LinkedIn

Want to apply this to your business?

If this sparked a useful question, let’s talk about where AI, automation, or product strategy can create practical leverage in your organisation.

Start the conversation