Watch: Grounding Bias
A form of secondary bias: how much an AI model defers to, or discounts, retrieved sources once they are grounded into the context, rather than leaning on what it already believed.
Transcript
When we use artificial intelligence, we often feed it search results or documents to help it give us accurate answers. This process is called grounding. But how the AI actually treats those sources depends on something called grounding bias.
Grounding bias is a secondary bias. It is the tug-of-war inside the model between the new information it just read and what it already believed before the search.
To understand it, look at its opposite: primary bias. Primary bias is the model's gut instinct, firing off before it reads any external sources. Grounding bias, on the other hand, happens after. It is the measure of how much the model defers to the new documents, or how much it stubbornly clings to its pre-existing knowledge.
When retrieved snippets fail to shift the model's answer, grounding bias is at play. To find and measure this bias, researchers run paired tests, comparing how the AI responds both with and without the grounding documents. By analyzing the difference between these two states, we can see exactly how much the AI trusts the facts we give it, versus how much it trusts itself.