AI Photo Readers Get Distracted by Random Text – Predictably

Many modern AI systems can do both: look at a picture and understand text. For example, you might ask one, “Is there a cat in this photo?” But what if you add an unrelated sentence, like “The coffee is hot”? Would the AI's answer change? A new study shows that yes – even completely irrelevant text can push these models to give different answers. And the effect isn't random. It follows a predictable, straight-line shift in the AI's confidence. Researchers call this an affine transformation, which basically means the distraction works like a consistent nudge in one direction.

The team tested several AI models on visual questions, keeping the question identical and only swapping the added text. They measured how strongly the model preferred one answer over another. The results were striking: the stronger the irrelevant text, the more the model's opinion moved – almost like the text is adding weight to one side of a scale. This means the bias is not chaotic noise but a systematic distortion that can be studied and estimated.

Based on this pattern, the researchers created two useful measurements. The first, “visual commitment preservation,” shows how well an AI sticks to what it actually sees in an image. The second, “directional answer bias,” reveals which way the irrelevant text tends to push the model. These tools could help developers diagnose weak spots in AI systems and build safeguards.

This matters because these image-and-text AI models are already used in real-world tools: scanning medical images, analyzing social media photos, helping self-driving cars understand their surroundings. If we can predict when an AI is being distracted, we can correct it – making these systems more trustworthy and safe for everyday use.