Learn · Accessibility descriptions
Verify the description, not just the writing.
An image or video description can be fluent, detailed and wrong. If it invents an action or misses an important detail, the reader may receive a different account of the material from someone who can inspect it directly.
In developing No Default, its creator reported discussing an AI-generated accessibility description for a short video with a blind person. The video had no narration, only bird sounds. The generated description contained inaccuracies and invented details about what happened visually. That prompted this guide. It is a reported discussion, not a published interview or a claim about what all blind people prefer.
Work from the original image or video
- Identify the purpose. What information does someone need to understand the image or video in this context? A social post, instructional demonstration and news report may call for different detail.
- Verify people, objects and actions. Confirm what is actually present. Don't turn a guess about intention, identity or cause into a factual description.
- For video, check the sequence. Watch the whole recording, revisit uncertain moments and verify what happens at each timestamp. Distinguish sounds that are actually audible from visual events that need describing. Never invent narration or changes between frames.
- Separate observation from interpretation. “The bird spreads its wings” is an observation when visible; “the bird is frightened” may be an unsupported explanation.
- Review relevant omissions. A description can contain no false statements and still leave out the feature that makes the content meaningful.
- Check the final wording. Remove unnecessary repetition, speculative adjectives and over-specific details that aren't supported.
Example: observation versus invention
Source observation: A bird approaches another bird, which then moves away.
Unsupported addition: “The first bird angrily chases the other away.”
Better description: “One bird approaches. The other turns and moves away.”
The source might support a stronger description in a particular case, but the observer must establish that from the actual recording rather than from the model's confident wording.
What should be checked especially carefully?
- Visual specifics: number of people or objects, colours, positions, clothing, text on screen and changes over time.
- Audible information: distinguish narration, dialogue, music and ambient sounds. A video with bird sounds but no narration still needs an accurate account of significant visual action; don't invent speech or imply that the sounds explain what is happening visually.
- Movement and timing: direction, sequence, duration and whether a timestamped description matches the recording.
- Emotion and intent: avoid claiming that a person or animal feels or intends something based solely on appearance or movement.
- Relevance: keep essential meaning without flooding the reader with details that don't help them understand the content.
Who should review it?
Someone who can access the original visual or audiovisual source should perform a source-accuracy check. People who use descriptions can also give valuable feedback on clarity, relevance, length and usefulness. These are complementary forms of review, not interchangeable checks. Ask what works for that individual rather than assuming one person speaks for a whole community.
When you cannot verify something
Omit unsupported details, use appropriately qualified language, or say that a detail cannot be established if it matters. Do not fill gaps just to make a description sound complete.
After publication
Provide a practical way for readers to flag problems, and correct descriptions when errors are identified. Preserve the difference between what was observed, what someone reported and what was inferred.