Playing telephone with generative models: "verification disability," "compelled reliance," and accessibility in data visualization

Generative models used to describe visualizations for blind and low-vision users can introduce or amplify bias in ways that are difficult for users to detect. The paper demonstrates through a "telephone game" experiment that AI models can misinterpret data visualizations and propagate these errors when generating descriptions, creating a compounding problem where users cannot easily verify accuracy.

People with visual impairments face "compelled reliance" on AI systems without the ability to independently verify outputs. Unlike sighted users who can check an AI-generated description against the actual visualization, blind and low-vision users must trust the model's interpretation, meaning algorithmic failures become particularly consequential for this population.

Current use of generative models for accessibility in data visualization may create a "verification disability"—a situation where users cannot adequately interrogate or challenge algorithmic bias. This represents a new form of accessibility barrier where the tool meant to increase access actually reduces user agency and control.

The risks of AI-assisted visualization description extend beyond individual errors to systematic bias reproduction. The paper shows that biases can be embedded in how models interpret and re-interpret visualizations, potentially affecting many users simultaneously when these systems are deployed at scale.

Responsible deployment requires input from multiple stakeholders: technologists must design for verifiability, users with disabilities need awareness of limitations, and researchers must continue studying how algorithmic bias specifically impacts accessibility. No single group can solve this problem alone.