Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Human-written alt text is usually the safer choice when the author understands an image’s purpose and the page around it. AI-generated descriptions can improve coverage and provide a useful first draft, but they can omit the detail that matters, misidentify content, or describe an image without explaining its function. The accessible choice depends on the image role, the information a reader needs, and whether a person reviews the result.
Accessibility is not a contest between fluent sentences and awkward ones. It is about giving a screen-reader user the same relevant information or action that the image gives other visitors. That requires a context decision for every image, even when AI helps with the wording.
As an Amazon Associate I earn from qualifying purchases.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →What makes alt text accessible?
Accessible alt text communicates an image’s relevant information or function in its page context. It does not have to list every visible feature, and it should not describe details that do not help the reader understand the content or complete an action.
W3C WAI’s Images Tutorial puts the decision with the author: “The text alternative needs to be determined by the author, depending on the usage, context, and content of an image.” The same picture can therefore need different alternatives on different pages.
#1 Best Overall
Identify the image’s role first
| Image role | What the alternative should do | Typical HTML treatment |
|---|---|---|
| Decorative | Contribute no information that the page needs. A border, flourish, or purely atmospheric background normally belongs here. | alt="" (a null alternative) |
| Informative | State the essential information conveyed by the image, such as the subject, result, or key condition. | A concise description in the alt attribute |
| Functional | Describe the action or destination of a control, link, or button. The wording should tell a user what happens when it is activated. | Alt text that names the function, not merely the appearance |
| Complex | Provide the important relationships, trends, values, or instructions in a complete text equivalent. A short alt attribute alone may not be enough. | Short identifying alt text plus nearby or linked long-form text |
Include important text inside an image
If an image contains words that matter—such as a warning, label, quote, or data point—those words need to be available in the alternative or in an equivalent text on the page. Saying “a poster with text” does not preserve the information.
Use the page, not just the pixels
A bird photograph on a species-identification page may need the species name. The same photograph used as a decorative banner may need an empty alternative. A product image beside a “Buy” control may need a functional label, while a detailed photograph in a lesson may need the visual fact that supports the lesson’s claim.
Is human-written alt text more accessible than AI-generated text?
Human authors have the strongest opportunity to match wording to purpose because they know why the image is present, what the surrounding copy already says, and which detail the reader must not miss. That advantage is not an automatic guarantee: a person can also write vague, redundant, or inaccurate text. AI can help when descriptions would otherwise be omitted, but its output needs a context-aware human review.
A 2022 comparison found that participants preferred descriptions written by Wikipedia contributors on average over the automatic systems tested. Some systems performed well for particular image categories, yet the study authors still reported better perceived correctness for the human-authored descriptions in that test set. The result applies to that study’s tools, images, and raters—not to every newer generative-AI system.
What the published comparisons actually measured
| Study or dataset | Scope | Reported figure or finding | How to interpret it |
|---|---|---|---|
| 2022 automatic-captioning comparison | 60 Wikipedia images; Azure Computer Vision Engine, Amazon Rekognition, Cloudsight, and Auto Alt-Text for Google Chrome | 76 computer-science students blindly rated descriptions; Wikipedia contributor descriptions were preferred on average | Perceived correctness in a bounded test, not a current benchmark for all AI systems and not a blind- or low-vision user study |
| 2022 blind- and low-vision participant study | Participants judged descriptions with and without page context | Context was highly valued; context-free referenceless metrics did not align reliably with participant judgments | Fluency or a generic caption score cannot substitute for judging usefulness on the actual page |
| 2024 AI scene-description diary study | 16 blind and low-vision participants using one AI-powered scene-description application for two weeks | Average satisfaction was 2.76 out of 5; average trust was 2.43 out of 4 | Useful evidence about one application and short study period, not a universal score for web alt text |
The evidence therefore supports a workflow rather than a universal winner: use human judgment to define the information requirement, use automation where it reduces omission, and edit before publication.
Where can AI-generated descriptions help?
AI is most useful as an assistive layer for scale. It can flag images that lack alternatives, create a draft for a large backlog, suggest visible text or objects for an editor to verify, and make it less likely that an image receives no description at all. Those benefits address a practical production problem without proving that the generated sentence is sufficient.
Recommended Free Tools
Good candidates for an AI draft
- Large libraries in which many images have no alt attribute.
- Routine informative photographs whose purpose is already clear from the page template.
- Editorial workflows in which a content creator can accept, rewrite, or reject a suggestion before publishing.
- Initial discovery of possible text, people, objects, or spatial relationships that an editor can check against the source image.
Cases that need especially careful human authorship
- Charts, diagrams, maps, medical or scientific figures, and any image whose relationships or trends carry the meaning.
- Images containing important words, numbers, warnings, or instructions.
- Images used as links, buttons, or controls, where the required wording is an action or destination.
- News, legal, safety, educational, and cultural content where a wrong identity or omitted detail changes interpretation.
- Images whose significance comes from the surrounding paragraph, caption, or sequence rather than from objects visible in isolation.
How should I compare AI and human alt text?
Judge both outputs against the same six questions. A polished sentence should not pass merely because it sounds natural.
- Accuracy: Does it identify the subject and important details correctly? Check names, numbers, colors when they matter, locations, and visible words.
- Completeness: Does it include the information the image contributes, or has it omitted the central point?
- Purpose: Does it explain why the image is on this page rather than reciting unrelated visual details?
- Function: If the image is interactive, does the text tell the user what the control does or where it leads?
- Relationships: For a chart, diagram, or group scene, does it preserve comparisons, sequence, labels, and connections?
- Workflow: Can a named content creator edit the suggestion, verify it against the original, and record approval before publication?
Human writing generally performs best when the author has strong subject and editorial context. AI can be faster and more consistent at producing a starting point across thousands of files. Neither characteristic removes the need to test the final alternative in context.
Can AI write accessible alt text without editing?
It should not be published unchecked. Treat generated text as a draft whose claims must be verified and whose focus must be adapted to the page.
A practical review workflow
- Classify the image. Mark it decorative, informative, functional, or complex before asking whether a description is needed.
- Read the nearby content. Note what the heading, caption, body copy, and adjacent controls already communicate. Do not make alt text repeat information the reader already has unless repetition prevents confusion.
- State the reader’s need. Write a short internal brief such as “identify the bird species shown” or “tell the user this button opens the annual report.”
- Generate or draft the alternative. If AI is used, provide the page purpose or editorial brief when the system supports that context. Do not assume a pixel-only prompt can infer the page’s intent.
- Verify every claim. Compare names, text, counts, directions, and relationships with the original. Remove guesses about identity, emotion, profession, or location unless the source establishes them.
- Choose the right length. Keep a simple informative alternative concise. Move a complete explanation of a complex graphic into visible or linked text rather than forcing all meaning into one short attribute.
- Check the HTML. Decorative images need
alt="", not a missing attribute. Informative and functional images need an appropriate value. Complex graphics need their extended equivalent connected to the image. - Review in context. Use a screen reader or accessibility review with the page’s surrounding content. Confirm that the sequence makes sense and that the alternative does not create duplicate or contradictory announcements.
- Publish only after approval. Assign responsibility to the content creator or editor, not to an automated process alone, and retain a way to revise the text when the image or page changes.
What are the most common AI alt-text failures?
Generated descriptions can be grammatically excellent while failing the image’s job. These are the failure patterns to look for during review.
- Object inventory instead of meaning: “A person, a table, and a laptop” may omit that the image is a screenshot of an error message or a demonstration of a task.
- Wrong function: “Blue arrow icon” does not tell a user that the control opens the next page.
- Missing text: “A sign” leaves out the warning or instruction printed on it.
- False certainty: The system guesses a person’s identity, occupation, age, emotion, or location from appearance.
- Lost relationships: A chart description names bars but omits which category is highest, the direction of a trend, or the comparison the article discusses.
- Redundancy: The alternative repeats a visible caption word for word, creating unnecessary verbosity without adding access.
- Decorative pollution: A decorative image receives a long description and interrupts a screen-reader user with irrelevant content.
- Template drift: A model produces plausible text for one page type but misses a changed purpose when the same asset is reused elsewhere.
How should complex images be described?
Complex images need a text equivalent that preserves their information, not merely a visual label. Start with a short identifier in the alt attribute, then provide the detailed explanation in nearby HTML or a clearly associated link.
Charts and graphs
Include the subject, the variables, the important comparison or trend, and any conclusion the surrounding text draws from the graphic. Listing every plotted point may be unnecessary when the page already supplies a data table; omitting the relationship that the graph is meant to reveal is not.
Diagrams and processes
Explain the components and the connections between them, including direction or sequence where those relationships carry meaning. “Diagram of a network” is an identifier, not a text equivalent.
Screenshots and interfaces
Include the visible words and the state or result that the screenshot demonstrates. If the screenshot accompanies instructions, describe the control or message the reader must find rather than narrating every decorative element.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →How do automated accessibility checks fit into the workflow?
Automated checks are valuable for finding missing alt attributes and other detectable markup errors. They cannot reliably decide whether a description is appropriate for the image’s purpose, whether a chart’s key relationship is preserved, or whether the wording conflicts with the surrounding page. A person must make that contextual judgment.
Use an automated check as a coverage report, not as an approval stamp. A useful review queue separates images with no alternative, empty alternatives that may be incorrect, generated drafts awaiting approval, and complex images requiring an extended equivalent.
What does current accessibility guidance say about generated descriptions?
W3C’s established image guidance is purpose-based: informative images need essential information, decorative images use a null alternative, functional images describe function, important text in images must remain available, and complex images need a complete text equivalent. W3C’s H37 technique likewise says an HTML img alternative should serve the same function as the image and does not necessarily need to describe visual characteristics.
A WCAG 3.0 image-alternatives outcome draft includes a supplemental requirement that auto-generated descriptions be editable by a content creator. That is draft guidance, not a normative WCAG 2.2 requirement. It nevertheless points to a sound governance rule: generated copy should remain editable and should not be treated as immutable machine output.
How can a team scale alt-text production without lowering quality?
Scale comes from assigning decisions to the right stage, not from removing review.
Rank #4
Set an image-purpose field in the content model
Require authors to select decorative, informative, functional, or complex. That choice determines whether an empty alternative, a concise description, an action label, or an extended text equivalent is needed.
Give AI the editorial brief
Where possible, pass the page topic, caption, or intended use to the generation step. A model that sees only pixels cannot reliably infer why an image matters in a particular article.
Route high-risk images to specialists
Send charts, diagrams, screenshots with instructions, images containing critical text, and content with legal, medical, safety, or educational consequences to a knowledgeable editor.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsMeasure coverage and corrections separately
Track how many images have an alternative, how many drafts are edited, and which error types recur. A rising completion rate is useful only if review does not reveal a corresponding rise in omissions or incorrect descriptions.
Recheck reused assets
An image can be informative in one placement and decorative or functional in another. Store the alternative with the page use when the context changes, rather than assuming one description is correct everywhere.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What do user studies say about context and trust?
Studies with blind and low-vision participants show why context and trust belong in the workflow. In a 2022 study, context was highly valued, while metrics that judged captions without page context failed to align with participant judgments. In a separate 2024 two-week diary study of one AI-powered scene-description application, 16 participants reported average satisfaction of 2.76 out of 5 and average trust of 2.43 out of 4. Those figures describe that application and study period; they are not a universal rating for web alt text.
Best Value
The distinction matters because a general scene-description tool answers “what might be in front of me?” A web alternative must answer “what does this image contribute here?” The systems may overlap technically, but their editorial objectives are different.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Which approach should an organization choose?
Choose based on volume, risk, and available editorial knowledge.
- Small site or high-context content: Have the author write the alternative while creating the page. The author is already making the purpose decision.
- Large archive with many omissions: Use AI to create a queue of drafts, then require human approval and prioritize images that carry information or perform a function.
- Highly technical or consequential content: Use a subject-matter-informed human writer, with AI limited to suggestions that are checked against source data and terminology.
- Mostly decorative imagery: Improve the content model so authors can correctly mark decorative assets with
alt=""instead of generating unnecessary prose. - Complex visualizations: Plan a text equivalent as part of the content, not as an afterthought to an image-captioning tool.
The key decision is not whether a sentence came from a person or a model. It is whether the final alternative conveys the image’s essential information or function to the reader who cannot see it, and whether someone accountable has checked that claim in context.
Frequently Asked Questions
Why do so many online images still have no description?
Historical platform data helps explain the production problem: a 2019 analysis of 1.09 million image tweets found descriptions in only 0.1% of them. In interviews, 20 people who had written descriptions cited forgetting, lack of time, and uncertainty about what to include as barriers.
Does a scene-description app’s satisfaction score measure the quality of website alt text?
No. The 2024 diary study covered 16 blind and low-vision participants using one AI-powered scene-description application for two weeks. Its averages—2.76 out of 5 for satisfaction and 2.43 out of 4 for trust—should not be read as a score for every web authoring system or page-specific alternative.
Free tools Windows power users keep installed
One-click scans. No signup required.
Can an automated checker approve my alt text?
It can identify missing attributes and some markup problems, but it cannot reliably decide whether the wording serves the image’s purpose or preserves a chart’s key relationship. Contextual review by a person remains necessary.
Does the WCAG 3.0 draft make AI-generated alt text mandatory?
No. The image-alternatives outcome is draft material, not a normative WCAG 2.2 requirement. Its supplemental language about keeping generated descriptions editable is better treated as a governance signal than as a current conformance rule.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
