A mainstream AI social app may offer the same AI personality through text, synthesized voice, photographs, short clips and real-time digital-human video. Multiple modes do not automatically make an experience accessible. A good assessment asks whether someone can complete the same meaningful task when hearing, vision, motor control, bandwidth or device settings change. This six-gate test plan pairs established web accessibility principles with practical mobile product evidence.
Start with outcomes and input alternatives
Define one task before evaluating interfaces: open an authorized AI profile, start a conversation, ask a question, understand the reply and end the session. Then test it using a screen reader, a keyboard or switch-access equivalent, enlarged text, reduced motion and voice-free operation. A user should not be forced into audio or video simply to reach a feature that is available in a text path. When a feature inherently requires media, state that dependency plainly.
The W3C WCAG 2.2 Recommendation, first published October 5, 2023, offers useful success criteria for operability, understandable content and robust interfaces. It is a reference point rather than proof that a particular app is compliant. Define a result for each gate: pass, partial with documented workaround, or fail with a named owner. Attach reproducible device and operating-system settings to each finding.
Gate one: readable text and semantic control labels
Increase text size to the highest supported accessibility setting and check whether a conversation can still be read and composed without clipped content. Confirm that send, back, mute, close and report controls have distinct programmatic names. Focus must remain visible and logical when a keyboard or assistive technology moves through the interface. A decorative avatar image should not be announced as a primary action unless it truly performs an action.
Test error messages after a failed send and ensure that they are communicated both visually and programmatically. The status ‘message is being retried’ is more helpful than a spinner with no label. Measure whether task completion improves after fixing confusing labels, not merely whether a static screenshot looks cleaner. Record the exact screen-reader utterance and focus sequence for every regression test.
Gate two: voice output that has a text fallback
Speech synthesis can make an AI personality expressive, but users may be unable or unwilling to play audio in public spaces. Provide readable text for the generated reply, an obvious audio mute, and the ability to switch from voice back to text without losing the conversational context. Avoid automatically routing sensitive sound to Bluetooth devices without a visible output indicator. On poor connectivity, keep the text response available if video streaming fails.
The W3C’s captions guidance distinguishes prerecorded and live media, with differing success criteria and user needs. The team should not label an inaccurate generated transcript as a perfect caption. Review samples containing names, accents, code-switching and background noise. Mark uncertainty, provide correction controls where appropriate and retain the option to communicate by text instead.
Gate three: live video with useful captions and clear exit
For a live digital-human conversation, captions must remain readable against changing backgrounds, avoid blocking the face or primary controls, and persist long enough for someone to follow the exchange. The user should be able to end capture independently of closing the entire app. Display microphone and camera state without relying on color alone. Make the end-call control reachable with assistive technology and verify that media capture stops promptly.
Test a call under simulated packet loss and delayed captions. A speech reply may arrive before the video animation, and a caption may appear after the person has already stopped speaking. Define acceptable synchronization windows based on user comprehension rather than a technology-only latency target. Test with real people who use captions or screen readers; internally generated test transcripts are not substitutes for user participation.
Gate four: shared images and short videos with context
Digital-human selfies, user-sent images and creator short clips add information that may not be spoken. Describe what an image contributes to the task without inventing details about a person’s identity, feelings or private traits. Avoid copying an entire decorative caption into alt text. When a creator uploads a short clip, provide a title, a description of the important action and captions for relevant speech. A photo-upload control should clearly describe which files are being shared.
Ask whether someone can understand a transaction, safety warning or consent prompt when the image fails to load. The W3C guidance emphasizes alternatives appropriate to the medium; that does not mean a single automatic description can replace human review in every situation. Sample outputs for privacy leaks and material errors before making them searchable. Offer a report mechanism for misleading AI-generated descriptions.
Gates five and six: resilient recovery and repeatable evidence
Gate five tests interruption: suspend the app, revoke microphone permission, switch networks, increase text size mid-session and return after a failed login. The app should state whether the conversation was saved, which input was lost and how to recover without revealing private content on the lock screen. A failed video call should not silently restart capture. Keep a user-facing fallback path and log only diagnostic events necessary to investigate the failure.
Gate six turns accessibility into release evidence. Use a versioned matrix of device, OS, assistive setting, test task, result, issue owner and retest date. Pair automated checks with assisted user sessions and define blocking defects before launch. The NIST AI RMF, released January 26, 2023, provides a useful govern-map-measure-manage cycle for assigning risks and verifying corrective actions. An accessible AI social experience is the ability to make a meaningful choice, communicate and leave, regardless of mode.
Sources
- W3C WCAG 2.2 Recommendation (first published 2023-10-05)
- W3C WAI Captions/Subtitles guidance (living guidance, checked 2026-11-08)
- NIST AI Risk Management Framework (released 2023-01-26)
