Automated testing tools currently detect only about 30% of accessibility errors. This leaves Quality Assurance teams and developers to close the remaining gap through manual testing, expert reviews, and often, a lot of guesswork. With Large Language Models (LLMs) now embedded in daily workflows, browsers, and IDEs, Artificial Intelligence seems like the long-awaited “Holy Grail” that could finally solve digital accessibility at scale. But is entrusting accessibility evaluation to AI models truly safe for product quality, or are teams simply outsourcing a highly nuanced problem to a confident, yet unpredictable, black box?
In this 45-minute session, Emilia Krzemińska-Komenda provides a no-nonsense, evidence-based analysis of where AI genuinely accelerates WCAG compliance and where it quietly introduces severe risks. Instead of theoretical discussions or vendor hype, this presentation is built around concrete, real-world use cases. Attendees will see a side-by-side comparison of how different AI assistants - specifically Google Gemini, Anthropic’s Claude, and ChatGPT - handle the same accessibility prompts.

To provide a clear roadmap, the presentation is structured into three practical segments:
The Accelerator: Where AI Excels
Emilia will explore real examples of how AI can bridge the communication gap in tech teams. Attendees will see how to successfully prompt models to translate dense, highly technical WCAG success criteria into plain, actionable language for developers and product owners.

The Double-Edged Sword: Generating Alternative Text
A deep dive into how Gemini, Claude, and ChatGPT handle image recognition for accessibility. Emilia will compare their outputs for UI screenshots, informational diagrams, and decorative content. The audience will see concrete examples of when AI-generated alt text is genuinely “good enough” and when it provides misleading context that a sighted reviewer might easily miss, ultimately failing visually impaired end users.

The Danger Zone: Accessibility Hallucinations
AI cannot yet be trusted as the final judge in an audit. This segment introduces “accessibility hallucinations” — specific patterns in which LLMs generate plausible-sounding yet completely incorrect advice. Emilia will examine examples where AI invents non-existent ARIA roles, misinterprets WCAG 2.2 criteria, and displays dangerous overconfidence by passing inaccessible code snippets.

Participants will leave this session with a ready-to-use “AI-as-an-Assistant” framework. This includes a decision checklist for when to let AI draft, when to let it suggest, and when to keep humans firmly in control. By the end of the talk, attendees will have a clear strategy for safely integrating AI into their testing workflow - using it to build quality and provide intellectual support, without replacing essential human empathy and expert knowledge.