Cyberpediaby Brain

AI for accessibility at work

Live captions, image descriptions, read-aloud: what Microsoft and Google say about these tools, and what an employee should check.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4How it works
  5. 5What you do
  6. 6Pitfalls
  7. 7Sources
  8. 8Cite this article

In 30 seconds

Where it shows up
Narrator[1] and live captions[3] in Windows 11, captions in Teams meetings[4], then Lookout[5], Expressive Captions[6] and TalkBack[7] on Android.
What it does
These tools turn a sound or an image into displayed text or into speech. For image descriptions, Microsoft writes that the feature may provide an inaccurate description[2].
The reflex
Reread the description or caption before relying on it, especially for a number, or a legal or financial document[2].
If it is already done
Narrator's image descriptions can be turned off in the Accessibility settings of Windows[2], and Teams captions can be turned off from the meeting controls[4].

What it is

Narrator, the screen reader built into Windows 11, can provide detailed descriptions of images and charts using AI on Copilot+ PCs[1]. Microsoft's page presenting this feature acknowledges that the description may be inaccurate[2].

This article gathers what Microsoft and Google say about their AI-based accessibility features: live captions, image description, read-aloud. It explains why an employee reviews what these tools produce, what data goes to an online service and which framework applies.

Straight answers

What does Narrator describe when asked about an image?

On a Copilot+ PC, Narrator gives a contextual description that details the people, objects, colors, text and numbers in the image, using the Narrator key+Ctrl+D shortcut[1]. On PCs that are not Copilot+, the experience remains a basic description[1]. In its March 2026 version, the guide says Narrator can work with Copilot on all Windows 11 devices: Copilot opens with the image ready and the user types their own request[1]. The image is shared only after the user chooses to describe it[1].

Can you rely on a generated description or caption?

Microsoft writes that the rich description feature may provide an inaccurate description, inaccurate data for a chart, or emotional inferences, which can lead to false assumptions about the image[2]. For Teams captions, Microsoft notes that captions may become less accurate as you move away from the mic[4]. The text produced is therefore an aid to be reread, not a source that needs no checking.

For which content does Microsoft advise against image description?

Microsoft writes that the feature should not be used to describe:

  • medical or health-related images, which could be mistaken for medical advice[2];
  • images in legal or financial documents where accuracy is critical[2];
  • images containing cultural or religious symbols without appropriate context[2];
  • images containing maps, flags or globes[2].

Microsoft adds that the feature should not be used to infer the emotions of people at work or in an educational setting, for example employees or students[2].

What data goes to an online service?

It depends on the tool. For Teams captions, Microsoft writes that data is processed in the geographical region where the organization is based[4]. The automatic speech recognition service (Microsoft Automatic Speech Recognition) may use the meeting's subject, the invitation, participant names, attachments and participants' recent emails to improve accuracy[4]. According to Microsoft, no one has access to it, and caption data is permanently deleted after the meeting[4].

For Narrator, Microsoft's FAQ (updated in June 2025) says the standard image description experience on other Windows devices relies solely on online services[2]. For the February 2024 version, the guide specifies that an active internet connection is required and that the option to get image descriptions must be turned on[1]. By contrast, Narrator's natural voices, once downloaded, work on the device without an internet connection[1], and Windows 11 live captions can appear offline[3].

Which framework applies to these tools?

The European Commission[8] presents the European Accessibility Act as a directive that aims to improve the functioning of the internal market for accessible products and services. It covers computers and operating systems, smartphones and e-commerce, among others, and Member States were to transpose it into national law in June 2022[8].

For personal data, the CNIL[9] points out that an AI system that uses personal data must always be developed, trained and deployed with a well-defined purpose. It also points out that data must be adequate, relevant and limited to what is necessary, and that people who exercise their rights are, in principle, to receive an answer within one month[9].

The NIST[10] published the AI RMF, intended for voluntary use, to better manage the risks that AI poses to individuals, organizations and society. It was published on January 26, 2023, and NIST states that it is under revision as part of the White House AI action plan[10].

How it works

For image description, Microsoft explains that the feature uses AI models that produce detailed text descriptions of images, charts and diagrams[2]. To measure quality, the company says it built a set of varied images (photographs, charts, screenshots, application interfaces) and assessed the descriptions on accuracy, completeness, relevance and usefulness, with human expert judgments and LLM-assisted scoring[2]. The user can regenerate the description and copy it[2].

Windows 11 live captions are turned on with Windows logo key + Ctrl + L or from the accessibility settings[3]. They support several languages, including French, and the caption window also lets you caption your own speech by connecting a microphone[3].

In Teams, the feature detects what is said in a meeting and shows real-time captions[4]. Setting the meeting's spoken language to the language actually used gives more accurate captions[4]. Teams does not save them: to get a transcript, you need to turn on transcription[4]. An alternative exists, a CART (Communication Access Real-time Translation) captioner, whose captions are produced by a human[4].

On Google's side, Lookout on Android helps people who are blind or have low vision learn about their surroundings through their phone's camera, and AI-generated image captions are available in Lookout globally in English, according to Google in May 2024[5]. In December 2024, Google says that Image Q&A in Lookout uses Gemini 1.5 Pro for richer descriptions, read aloud, with follow-up questions[6]. In May 2025, Google states that TalkBack, Android's screen reader, offers Gemini-generated image descriptions even without alt text, and that you can ask questions about an image or about the whole screen[7]. Expressive Captions uses AI to convey what a person says and how they say it[7].

What you do

  • Reread before using. A description or caption is an aid; for a figure, a name or a legal or financial document, check at the source[2].
  • Improve the audio input. Microsoft recommends speaking clearly, slowly and directly into the mic, avoiding noisy places and not having several people speak at once[4]. For the Teams desktop version, it adds to use the highest-quality Teams-certified equipment available[4].
  • Set the meeting's spoken language. If the detected language differs from the set one, Teams prompts the organizer, the co-organizer or the person who started the transcription to update it[4].
  • Keep control of sharing. In Narrator, the image is shared only after the user chooses to describe it[1].
  • Report errors. The thumbs up and thumbs down in the Narrator interface let you give feedback on a description[2].

For a manager or colleague who turns these tools on:

  • Tell the participants. In Teams, people can choose not to be identified in captions[4].
  • Ask the administrator if needed. According to Microsoft, the administrator can disable the captions feature[4].
  • For translated captions, check the license: they are available with Teams Premium and Microsoft Copilot[4].
  • If human captioning is preferable, the CART captioner exists[4].

Pitfalls

  • A plausible description is not an accurate description. Microsoft notes that the model may also associate an image with a person or describe emotions, based on context clues such as a jersey and its number, without using biometric data[2]. Answers about emotions are based on appearance and do not necessarily reflect a person's inner state[2].
  • Captions hide some words. In Teams, obscenities are obscured by asterisks[4].
  • Availability varies. Microsoft specifies that its articles are written for the United States market and that features and availability may vary by region[3]. The new version of Expressive Captions is rolling out in English in the U.S., U.K., Canada and Australia, on Android 15 and above[7].
  • Quality in testing does not tell you quality in real conditions. The CNIL writes that the performance of the learning phase must be kept separate from that of the system in production: the quality of the former does not predict that of the latter[9].

Sources

  1. Microsoft, “Complete guide to Narrator”. https://support.microsoft.com/en-us/accessibility/windows/narrator/complete-guide-to-narrator (accessed 14 November 2026)
  2. Microsoft, “Frequently asked questions about rich image descriptions”, 11 February 2025. https://support.microsoft.com/en-us/accessibility/windows/copilot/frequently-asked-questions-about-rich-image-descriptions (accessed 14 November 2026)
  3. Microsoft, “How to customize your live caption settings for any purpose”, 5 July 2023. https://www.microsoft.com/en-us/windows/learning-center/choose-best-caption-settings (accessed 14 November 2026)
  4. Microsoft, “Use live captions in Microsoft Teams meetings”. https://support.microsoft.com/en-us/teams/meetings/use-live-captions-in-microsoft-teams-meetings (accessed 14 November 2026)
  5. Google, “8 new accessibility updates across Lookout, Google Maps and more”, 16 May 2024. https://blog.google/company-news/outreach-and-initiatives/accessibility/ai-accessibility-update-gaad-2024/ (accessed 14 November 2026)
  6. Google, “Connect and share in more ways with new Android features”, 5 December 2024. https://blog.google/products-and-platforms/platforms/android/new-android-features-december-2024/ (accessed 14 November 2026)
  7. Google, “New AI and accessibility updates across Android, Chrome and more”, 15 May 2025. https://blog.google/company-news/outreach-and-initiatives/accessibility/android-gemini-ai-gaad-2025/ (accessed 14 November 2026)
  8. European Commission, “European Accessibility Act (EAA)”, 3 February 2025. https://commission.europa.eu/strategy-and-policy/policies/justice-and-fundamental-rights/disability/european-accessibility-act-eaa_en (accessed 14 November 2026)
  9. CNIL, “IA : comment être en conformité avec le RGPD ?”, 5 April 2022. https://www.cnil.fr/fr/intelligence-artificielle/ia-comment-etre-en-conformite-avec-le-rgpd (accessed 14 November 2026)
  10. NIST (United States), “AI Risk Management Framework”, 12 July 2021. https://www.nist.gov/itl/ai-risk-management-framework (accessed 14 November 2026)