AI Agents Have Brand Biases Baked In From The Start

Paris, Tuesday 24 February — IASEAI'26, UNESCO Headquarters

The finding: Research presented today at IASEAI’26 in Paris confirms that large language models encode systematic preferences for specific brands and sources — preferences that operate below the surface of any user interaction, can override content quality entirely, and cannot be neutralized by instructing the model to “not be biased.”

The scale: 12 models, 6 providers, three task domains: news selection, academic paper prioritization, product-seller recommendation. The biases are consistent, predictable, and cross-model.

The brand risk: A brand’s standing in a model’s preference hierarchy is invisible, varies by category, and is not correctable at the prompt level.

Auditing practices: The researchers’ conclusions point toward “auditing practices” as a practical next step — a capability the brand community does not yet have at scale.

MQ note: Wait. Did somebody say auditing practices? Outside of this newsletter, I have developed a tool for exactly that. The Brand AI Confidence Meter (BAICM) uses structured prompt batteries — varied by model, intent, and category — to help reveal and monitor all the sources different AI systems look to when responding to brand-relevant queries, their confidence in citing them, and their relative competitive risk.

I acknowledge the coincidence: I’m here at IASEAI’26 reporting on independent academic research that validates the problem my consultancy at MichaelQuinn.ai was built to address. A welcome convergence! Back to the study:

Soumi Das from the Max Planck Institute for Software Systems presented the findings of her team today that add a measurable, documented dimension to a question that has received surprisingly little systematic study: whether large language models exhibit consistent preferences for certain sources when selecting and recommending information, independent of content quality. Their study — “In Agents We Trust, but Who Do Agents Trust?” — tests that question across twelve models, six providers, and three task domains.

What they found

In controlled experiments, content was identical while only the name of its source varied. Models consistently favored certain sources, and in real-world tests with actual articles, source preference was so dominant it overrode differences in content quality. The preferences are consistent, predictable, and cross-model.

They are also context-specific in ways that compound the risk. A model may strongly favor a brand in one category and be largely indifferent to it in another, even when the content is identical. Equity built in one domain does not transfer automatically when a brand moves into adjacent territory; the model prefers its own prior category-level biases.

What does not work

The researchers tested whether instructing models to disregard source preferences, using direct prompts along the lines of “do not be biased”, reduced the observed behavior. It did not. As the paper states, “simple prompting-based strategies are often insufficient to override them.” The researchers note this finding points to the need for more robust control methods than prompt-level intervention alone.

What the paper calls for

The researchers do not take a stance on whether these latent source preferences are inherently harmful — in some settings they may direct users toward higher-quality sources. The concern is opacity. Their closing recommendation is direct: “Organizations may need to actively monitor and manage how their brand, credentials, and digital identities are encoded in the data ecosystem, while also developing safeguards against brand impersonation or adversarial mimicry.”

Day two continues tomorrow with the Economics of AI panel and the accountability and failure-detection tracks.

“In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations” — Khan, Amani, Das, Ghosh, Wu, Gummadi, Gupta, Ravichander. Max Planck Institute for Software Systems (MPI-SWS) and Microsoft. Presented at IASEAI’26, Paris, 24 February 2026.

MPI-SWS is part of the Max Planck Society, Germany’s premier basic research organization and consistently one of the world’s most cited scientific institutions.

This post first appeared in the Brand AI Report, my newsletter on how AI describes and recommends brands.