Skip to the report

Research report · AI tool accessibility 2026

We tested 30 AI tools with a blind user. Only 6 passed.

Our 2026 AI Tool Accessibility study evaluated how usable 30 popular AI tools are for people who rely on a screen reader and keyboard navigation.

The goal was practical: can a blind or low-vision user independently complete the core tasks these tools are built for, including signing in, prompting, reading and copying a response, managing chats, changing settings, and signing out?

  • 6 highly accessible tools, grade A
  • 13 usable with gaps tools, grades B to C
  • 11 largely inaccessible tools, grades D to F

The headline finding

80% of the AI tools we tested fell short of basic accessibility for a blind screen reader user.

Twenty-four of thirty tools left a blind tester unable to finish at least one core task without a workaround, or unable to finish it at all.

Results by category

Every tool we tested, grouped by what it is built to do and graded on the same ten tasks.

Chat assistants
4 highly accessible, 5 usable with gaps, 1 largely inaccessible
Image generation and editing
2 highly accessible, 1 usable with gaps, 4 largely inaccessible
Video generation
2 largely inaccessible
Coding and app builders
2 usable with gaps
2 usable with gaps
Companions
1 usable with gaps, 1 largely inaccessible
Agentic
2 largely inaccessible
Productivity
1 usable with gaps
Accessibility AI
1 usable with gaps
Music and audio
1 largely inaccessible

The gap is hard to miss. The big labs’ chat assistants, including ChatGPT, Claude, Gemini and Copilot, took accessibility seriously. Image, video, coding and search tools mostly didn’t. The rule of thumb: the further a tool gets from text in, text out, the worse it works for blind users.

The leaderboard

Every tool tested, ranked by how independently a blind screen reader user could complete ten core tasks. Sort any column, or filter by tier.

Category and tier columns appear on wider screens.

All 30 AI tools ranked by accessibility grade, with category, tier and the key finding for each.
Tier Key finding
1AGrade AGeminiChat assistantHighly accessibleBest in class. Every task completed cleanly with only cosmetic label gaps.
2AGrade AChatGPTChat assistantHighly accessibleExcellent screen reader support; settings only navigable via arrow keys.
3AGrade AClaudeChat assistantHighly accessibleStrong across the core journey; no copy confirmation, an unnamed email field and duplicated state cues.
4AGrade ACopilot (Microsoft)Chat assistantHighly accessibleVery strong; only missing an interim “generating” announcement.
5AGrade ADALL·E (via ChatGPT)Image generationHighly accessibleInherits ChatGPT’s strong accessibility; same settings Tab-key limitation.
6AGrade ACanva (Magic Studio)Design assistantHighly accessibleExcellent with clear status announcements; only prompt and response separation is missing.
7BGrade BLovableCode / app builderUsable with gapsStrong, polished experience; minor gaps in generation cues and prompt/response separation.
8BGrade BBe My EyesAssistive (accessibility)Usable with gapsStrong accessibility as expected for an assistive tool; minor settings-confirmation lag.
9BGrade BGoogle AI StudioChat assistantUsable with gapsGood announcements, but no copy button and only a cancel (no retry).
10BGrade BGrokChat assistantUsable with gapsGood experience marred by an unlabelled prompt field and no regenerate option.
11BGrade BPoeChat assistantUsable with gapsSolid multi-model chat; copy lacks confirmation and stop placement is awkward.
12BGrade BYou.comAI searchUsable with gapsSolid overall; verification fields and missing status announcements add friction.
13CGrade CJanitor AICompanion chatUsable with gapsUsable, but there is no copy button and no send or response announcements.
14CGrade CMeta AIChat assistantUsable with gapsReadable, but several unlabelled controls and no prompt/response separation.
15CGrade Cv0 (Vercel)Code / app builderUsable with gapsInterface is navigable, but generated code reads as “blank” to screen readers.
16CGrade CMistral Le ChatChat assistantUsable with gapsWorks, but key controls are unlabelled and settings/history are hidden in the search box.
17CGrade CNotion AIProductivityUsable with gapsUsable, but unlabelled login and prompt controls and no generation announcements.
18CGrade CPerplexityAI searchUsable with gapsFunctional but noisy: auto-announcements and no prompt/response distinction hurt clarity.
19CGrade CRemove.bgImage editingUsable with gapsBackground removal works and download is accessible, but no result or processing announcements.
20DGrade DAdobe FireflyImage generationLargely inaccessibleSign-in focus trap, missing image alt text and no stop or regenerate weigh it down.
21DGrade DLeonardo AIImage generationLargely inaccessibleCore generation works, but copy and download are hover-only and there is no stop control.
22DGrade DSunoMusic generationLargely inaccessibleSongs generate and play, but no status cues, no copy (share link only) and no stop or retry.
23DGrade DDeepSeekChat assistantLargely inaccessibleUsable only via Google login and arrow-key workarounds; unlabelled buttons, no status cues.
24DGrade DIdeogramImage generationLargely inaccessibleGeneration works, but results appear in a popup that loses focus and copy is hover-based.
25DGrade DManusAI agentLargely inaccessibleWorks via arrow-key workarounds; no copy, and New Chat, settings and sign-out aren’t focusable.
26FGrade FFreepik AIImage generationLargely inaccessibleImage output is inaccessible: unhelpful alt text, hover-only copy and download, no stop control.
27FGrade FGensparkAI agentLargely inaccessibleCritically inaccessible by keyboard: almost every control works only in browse mode.
28FGrade FCharacter.aiCompanion chatLargely inaccessibleBlocked for keyboard users: Settings and Logout can’t be activated, and there is no copy option.
29FGrade FHailuo AIVideo generationLargely inaccessibleSeverely inaccessible: settings and sign-out are hover-only with no keyboard path.
30FGrade FKling AIVideo generationLargely inaccessibleLargely inaccessible: controls lack roles and focus, and image actions are mouse-hover only.

Showing all 30 tools.

What good looks like

These tools prove accessible AI is a design choice, not a trade-off.

Gemini, grade A

Gemini offers a strong screen reader experience for the core journey. Login uses the standard Google flow, response progress is announced, copy actions are confirmed, and settings and sign-out are fully operable.

The main concerns are non-descriptive and partially redundant accordion labels, content that is exposed while accordions are collapsed, a mislabelled logo, and keyboard shortcuts that are not announced.

WCAG compliance findings

  • Status messages. Clear, non-disruptive feedback for response generation and copy actions.
  • Focus order. Focus order supports task completion.
  • Name, role, value. Controls are operable, but some names and state cues are misleading or redundant.
  • Headings and labels. Most labels support navigation, but some are not self-describing.
  • Info and relationships. Page structure is largely conveyed, but collapsed regions are not hidden from assistive technology.
Watch the Gemini test on YouTube, opens in a new tab

ChatGPT, grade A

ChatGPT demonstrates strong accessibility across the core AI workflow, particularly in status messaging, screen reader compatibility, focus management, and content structure.

Most accessibility concerns are minor usability issues rather than barriers to task completion. The primary WCAG issue relates to the “More” control, which does not adequately expose its role and state to assistive technologies.

WCAG compliance findings

  • Name, role, value. Users could identify and operate controls independently throughout the core user journey.
  • Status messages. Provides clear, non-disruptive feedback about system status and user actions.
  • Focus order. Focus generally follows a logical sequence that supports task completion.
  • Headings and labels. Content structure supports navigation and understanding.
  • Info and relationships. Information and page structure are conveyed effectively to assistive technologies.
Watch the ChatGPT test on YouTube, opens in a new tab

Claude, grade A

Claude provides a strong screen reader experience across the core journey. Generation progress is announced (“Claude is responding”, “Claude finished response”), prompts and responses sit under their own headings, and onboarding, Retry, settings and sign-out are fully operable.

The main concerns are a sign-in step that asks for a verification code but emails a link, an email field without an accessible name, no confirmation when a response is copied, and a Stop button that sits third in the focus order. The sidebar toggle and “Recent” button also announce their state twice, and the main landmark is read twice.

WCAG compliance findings

  • Status messages. Generation start and finish are announced clearly, but copying a response gives no feedback.
  • Info and relationships. Prompts and responses are separated by headings, so answers are easy to find.
  • Focus order. Focus lands in the prompt field after onboarding and new chats, but Stop is the third focusable element after the input.
  • Name, role, value. Controls are operable, but the email field has no accessible name and two controls convey their state twice.
  • Headings and labels. Most labels support navigation, but “Recent” is not self-describing.
Watch the Claude test on YouTube, opens in a new tab

Canva (Magic Studio), grade A

Canva (Magic Studio) provides a strong screen reader and keyboard experience. Generation status is announced (“Loading”, “Loading finished”, “Canva AI responded with”), copy and stop are confirmed, and settings, history and sign-out are fully operable.

The main WCAG concerns are the lack of structural separation between prompts and responses, and duplicate Copy labels. Users must also locate Canva AI within the left navigation after login.

WCAG compliance findings

  • Status messages. Generation, completion, copy and stop are all clearly announced.
  • Focus order. Focus is managed well across the flow.
  • Info and relationships. Prompts and responses are not structurally separated.
  • Headings and labels. Most labels are clear, but Copy is duplicated.
Watch the Canva (Magic Studio) test on YouTube, opens in a new tab

Works, but not without friction

Small fixes stand between these tools and a genuinely accessible experience.

Lovable, grade B

Lovable offers a usable screen reader experience. Login and verification are announced, focus moves straight to the prompt field after login and new chats, the Stop button is properly announced, and copy actions and settings changes are confirmed.

The main concerns are silent response generation, prompts and responses that are not separated by headings or landmarks, several buttons sharing the same “Copy” or “Project Action Menu” label, and no Regenerate or Continue option. The “message was cancelled” notice is not announced, and drag-and-drop in the history is not fully keyboard accessible.

WCAG compliance findings

  • Focus order. Focus moves directly to the prompt field after login, new chats and settings changes.
  • Status messages. Copy and settings changes are confirmed, but generation progress and cancellation are not announced.
  • Info and relationships. Prompts and responses are not separated by headings or landmarks.
  • Headings and labels. Several Copy and action menu buttons share identical labels.
  • Keyboard. Drag-and-drop in the conversation history is not fully keyboard accessible.
Watch the Lovable test on YouTube, opens in a new tab

Grok, grade B

Grok supports the core journey with a screen reader. All login options are accessible, focus moves straight to the prompt field, copying is confirmed with “Copied”, and New Chat, History, settings and sign-out are operable.

The main concerns are loading indicators that are not announced during login or response generation, a prompt field without an accessible name, and no separation between prompts and responses. Sidebar controls have conflicting labels and states, the Stop button is third in the focus order, there is no Regenerate option, and signing out is not confirmed.

WCAG compliance findings

  • Status messages. Copy is confirmed, but login, generation and sign-out give no spoken feedback.
  • Name, role, value. The prompt field has no accessible name, and sidebar controls have conflicting names and states.
  • Info and relationships. Prompts and responses are not structurally separated.
  • Focus order. Focus lands in the prompt field after login and New Chat, but Stop is the third focusable element.
  • Headings and labels. The “History” label could be more descriptive.
Watch the Grok test on YouTube, opens in a new tab

Perplexity, grade C

Perplexity is usable with a screen reader, but its feedback is inconsistent. Onboarding, the Copy button, conversation history, theme settings and sign-out are accessible, prompts are marked up as headings, and focus moves to the prompt field after a new chat.

The main concerns are continuous automatic announcements that talk over the screen reader in the prompt field, and no feedback when a response starts, finishes, stops or is copied. The verification field does not explain what to enter, messages do not identify their sender, the Stop button is hard to reach, and sidebar sections can only be expanded with a mouse.

WCAG compliance findings

  • Status messages. Announcements repeat continuously in the prompt field, while generation, stop and copy are silent.
  • Keyboard. Collapsed sidebar sections can only be expanded on mouse hover.
  • Focus order. Stop is the fifth button after the prompt field, and activating it moves focus back to the input.
  • Name, role, value. The History control mixes an action with its state, and “New” and “Collapse sidebar” are unclear.
  • Labels or instructions. The code field announces “Digit 1 of 6” without saying a verification code is needed.
  • Info and relationships. Prompts have heading semantics, but the sender of each message is not identified.
Watch the Perplexity test on YouTube, opens in a new tab

Meta AI, grade C

Meta AI supports the core journey with a screen reader. Responses are readable, the Copy button is clearly labelled and confirms with “Copied to clipboard”, and New Chat, conversation history, settings and sign-out are all operable.

The main concerns are loading indicators that aren’t announced during login or response generation, and a prompt field and profile button without clear accessible names. There’s also no separation between prompts and responses. The Stop button is third in the focus order, there’s no Regenerate option, and the sidebar doesn’t clearly show whether it’s expanded or collapsed.

WCAG compliance findings

  • Status messages. Copy is confirmed, but loading during login and response generation is not announced.
  • Name, role, value. The “Ask Meta AI” field has no accessible name, and the profile button is not clearly labelled.
  • Info and relationships. Prompts and responses are not structurally separated, and the Meta logo link sits inside the sidebar visually but not in its structure.
  • Focus order. The Stop button is accessible, but it is the third focusable element after the prompt field.
  • Headings and labels. The “History” label could be more descriptive.
Watch the Meta AI test on YouTube, opens in a new tab

Where AI fails blind users

These aren’t edge cases. They’re the core product experience, broken.

Adobe Firefly, grade D

Adobe Firefly lets screen reader users complete most tasks. Sign-in and verification are announced, focus moves straight to the prompt field, “Image copied to clipboard” is announced, and Preferences and sign-out are operable.

The main concerns are a keyboard focus trap in the sign-in modal, silent image generation, generated images without meaningful alt text, and history items announced only as “grouping”. Drop-downs, checkboxes and Favourite buttons lack accessible names, action controls appear only on hover, and there is no Stop or Regenerate option.

WCAG compliance findings

  • Non-text content. Generated images have no meaningful alternative text.
  • Keyboard. A focus trap in the sign-in modal and hover-only action controls block keyboard users.
  • Status messages. Copy is confirmed, but generation start and completion are not announced.
  • Name, role, value. History items, drop-downs, checkboxes and Favourite buttons lack meaningful names.
  • Label in name. The prompt field’s accessible name does not match its visible label.
Watch the Adobe Firefly test on YouTube, opens in a new tab

DeepSeek, grade D

DeepSeek’s Google sign-in works, focus moves straight to the prompt field, and previous conversations can be reopened and read with a screen reader.

The main concerns are unlabelled buttons across the page, including Copy, Stop, Send, Search, the sidebar toggle and a button inside Settings, and missing roles and states on New Chat, the profile menu and Sign Out. Profile menu options only receive focus at the end of the page, generation, completion and copy are not announced, and email verification codes never arrive.

WCAG compliance findings

  • Name, role, value. Many buttons are announced only as “button”, and key controls lack roles and states.
  • Status messages. Generation, completion and copy actions are not announced.
  • Focus order. Profile menu options receive focus only at the end of the page, and Stop is third after the prompt field.
  • Info and relationships. Prompts and responses are not separated, and history labels such as “Today” are not marked as headings.
  • Non-text content. The DeepSeek logo has no text alternative.
Watch the DeepSeek test on YouTube, opens in a new tab

Character.ai, grade F

Character.ai lets screen reader users sign in, create a character, read responses and regenerate, but several core actions are blocked by keyboard-inaccessible controls. The Stop and Logout controls and the Settings button cannot be activated by keyboard, and there is no Copy control or generation status messaging.

Key WCAG concerns are keyboard operability of essential controls, missing status messages, missing form error messaging, and a non-descriptive history label.

WCAG compliance findings

  • Keyboard. Several essential controls cannot be operated by keyboard.
  • Status messages. Generation and form-validation states are not announced.
  • Headings and labels. History is reachable but not clearly labelled.
Watch the Character.ai test on YouTube, opens in a new tab

Kling AI, grade F

Kling AI presents severe accessibility barriers throughout the workflow. The reading order is illogical, many controls do not receive Tab focus and lack roles, names and states, and core actions on generated output are available only on mouse hover.

Overall, the screen reader experience is very poor and independent use is extremely difficult. The findings span keyboard operability, name, role and value, focus management, and non-text content.

WCAG compliance findings

  • Keyboard. Many interactive elements cannot be reached or operated with the keyboard.
  • Name, role, value. Controls widely fail to expose appropriate names, roles and states.
  • Focus order. Focus is not moved to dialogs or newly revealed content.
  • Non-text content. Generated images carry no meaningful description.
  • Status messages. Submission is announced but completion is not.
Watch the Kling AI test on YouTube, opens in a new tab

Three failures we saw everywhere

The same problems appeared across tools, categories, and company sizes.

Hover-only controls

Buttons, menus, and settings that appear only when a mouse hovers over them. If you can’t use a mouse, these controls don’t exist. The single most common hard blocker, it appeared in Kling AI, Hailuo AI, Leonardo, Mistral, and others.

Unlabelled or mis-coded controls

A screen reader announces “button” with no name, or a control coded without a role. This forces users to guess what controls do. Ideogram, DeepSeek, Genspark and Kling AI were among the worst offenders.

Silent streaming responses

AI responses stream token by token. Without proper ARIA live regions, a screen reader either re-reads the entire response constantly or says nothing at all. ChatGPT, Gemini and Canva handled this well. Most others didn’t.

How we tested

One blind tester. One screen reader. Ten tasks. Thirty AI tools. No mouse.

The ten tasks

  1. Sign in
    Create or access an account; note CAPTCHA, OAuth, or verification steps.
  2. Start a conversation
    Locate and activate the prompt or input field.
  3. Send a prompt
    Submit a fixed text or media generation prompt.
  4. Read the response
    Find and read the full output; confirm whether it is announced.
  5. Copy the response
    Locate and use a copy control; note any confirmation.
  6. Stop or regenerate
    Reach and operate stop and regenerate controls.
  7. New chat
    Begin a second conversation while preserving the first.
  8. Find history
    Locate history and reopen a prior conversation.
  9. Change a setting
    Open settings, change one option, confirm it is saved.
  10. Sign out
    Locate and activate the sign-out control.

Scoring

Each task was assessed for control discoverability, accessible names and roles, keyboard focusability and operability, focus management, reading order, and status announcements, then given a single completion-effort score.

The completion-effort score given to each task.
Score Result Meaning
2PassDone within the time limit, no workarounds needed.
1Pass with difficultyDone, but took more time or required workarounds.
0FailTester took much time, or the task was impossible to finish efficiently.

Per-task effort was aggregated into an overall time and effort score, then mapped to a letter grade. All testing was performed with a screen reader and keyboard only, on free or trial accounts, the same way any new user would.

How the time and effort score maps to each letter grade.
Grade Time / effort score Interpretation
A≤ 2.0Largely accessible; minimal friction
B2.5Good, with minor gaps
C3.0Usable but with recurring friction
D3.5Significant barriers; workarounds needed
F≥ 4.0Severe barriers; often not independently usable
Download the full dataset

CSV · 30 tools × 10 core tasks

Disclosures and limitations

What this study can support, and what it can’t.

  • What the score measures, and what it doesn’t

    The completion-effort score is a friction proxy, not a conformance audit. It captures how much time and how many workarounds a task requires, which blends genuine accessibility barriers with ordinary task complexity and tester familiarity. A strong score means a task was easier to complete, not that the tool conforms to WCAG. No success criteria were formally tested, and no pass or fail conformance claim should be inferred.

  • Scores and written impressions can diverge

    Tier placement and scores are derived solely from recorded time-and-effort data. The qualitative notes reflect the tester’s lived experience, which sometimes points the other way. Where the two disagree, treat them as two different lenses on the same product rather than a contradiction to resolve.

  • Single tester, single configuration

    All findings come from one evaluator using one screen reader and one browser on one operating system. Screen reader behaviour varies meaningfully across JAWS, NVDA, VoiceOver, and TalkBack. A barrier seen here may not appear elsewhere. Results are illustrative of real friction, not statistically generalisable.

  • Point-in-time snapshot, June 2026

    AI tools ship interface changes frequently, often weekly. Specific findings such as unlabelled buttons, broken focus order, or hover-only menus may already be fixed or newly introduced by the time this is read. Treat every tool-specific observation as accurate only as of the testing window.

  • Free and trial access

    Tools were generally evaluated on free tiers or trial accounts. Paid plans may expose different sign-in flows, settings, and features that were not assessed.