Microsoft 365 apps, including Word, and PowerPoint, now offer AI-generated alternative text (alt text) to support accessibility. Alt text provides written descriptions of images, helping people who use screen readers better understand visual content.

You can get AI-generated alt text in two ways: automatically or on demand, depending on your device.

  • On Copilot+PCs, alt text is generated automatically and locally when you insert an image. The built-in AI model runs entirely on the device, and no data leaves it.

  • On all other devices, automatic generation does not occur at image insertion. Instead, you can request alt text manually from the Alt Text pane or the Accessibility Assistant. When you do, the image is securely sent to a Microsoft cloud model, and a description is returned for your review. To learn more, please refer - Richer alt text in Word and PowerPoint, powered by generative AI.

In every case, AI-generated descriptions are clearly labeled with a disclaimer indicating they may be inaccurate.

Copilot+PCs are Windows devices that contain a Neural Processing Unit (NPU) with a capacity > 40 TOPS. NPUs are a specialized hardware component designed to run AI models efficiently and privately on-device. This allows features like image captioning to work without requiring internet connectivity or sending data to the cloud. Learn more about Copilot+ PCs and Windows 11 PCs from Surface

No. A Microsoft 365 subscription is required, but a M365 Copilot license is not necessary.

Privacy and data protection are core to this feature’s design:

  • When using Copilot+PCs, all processing is handled locally on your device.

  • When using cloud-based generation, image data is securely transmitted and processed only in response to your explicit request.

  • Alt text is generated solely from the image’s visual content — Microsoft 365 does not use organizational data (e.g., Microsoft Graph) to influence the description.

  • No image data is retained or used to train future models.

AI-generated alt text is designed to support accessibility and may not always be fully accurate. Descriptions can be incomplete, overly general, or imprecise.

The feature is intended only for accessibility and should not be used for:

  • Emotion detection or interpretation

  • Identification of individuals or sensitive attributes

  • Legal, medical, financial, or employment-related decisions

  • Surveillance or profiling

Users are strongly encouraged to review and revise generated alt text as appropriate, especially for sensitive or specialized content. Where suitable, images can be marked as decorative, or manually authored descriptions can be provided.

If you are using an Copilot+PC device and wish to disable automatic alt text generation, go to:File > Options > Accessibility > Uncheck Enable the option to generate alt text using AI.

On other devices, automatic generation is not enabled by default, and alt text must be requested manually.

The underlying models were evaluated using a variety of real-world images, such as photos, diagrams, screenshots, and infographics. Testing focused on accuracy, relevance, and accessibility effectiveness. Evaluation involved a combination of human reviewers and automated quality checks, along with Responsible AI reviews and safety testing.

Yes. When alt text is generated using online services, the following languages are not supported: Quechua, Konkani, Amharic, Scottish Gaelic, and Māori. All other previously supported languages remain available.

When alt text is generated directly on your Copilot+PC device (without using online services), English is the only supported language currently. Support for additional languages may be added in future updates.

Language Code

Language Name​​​​​​​

af-ZA

Afrikaans (South Africa)

sq-AL

Albanian (Albania)

am-ET

Amharic (Ethiopia)​​​​​​​

ar-SA

Arabic (Saudi Arabia)

hy-AM

Armenian (Armenia)

as-IN

Assamese (India)

az-Latn-AZ

Azerbaijani, Latin script (Azerbaijan)

eu-ES

Basque (Spain)

be-BY​​​​​​​

Belarusian (Belarus)

bn-BD

Bengali (Bangladesh)

bn-IN

Bengali (India)

bs-Latn-BA​​​​​​​

Bosnian, Latin script (Bosnia and Herzegovina)

bg-BG

Bulgarian (Bulgaria)

ca-ES

Catalan (Spain)

zh-CN

Chinese (Simplified, China)

zh-TW

Chinese (Traditional, Taiwan)

hr-HR

Croatian (Croatia)

cs-CZ

Czech (Czech Republic)

da-DK

Danish (Denmark)

prs-AF

Dari (Afghanistan)

nl-NL

Dutch (Netherlands)

et-EE

Estonian (Estonia)

fil-PH

Filipino (Philippines)

fi-FI

Finnish (Finland)

fr-FR

French (France)

gl-ES

Galician (Spain)

ka-GE

Georgian (Georgia)

de-DE

German (Germany)

el-GR

Greek (Greece)

gu-IN

Gujarati (India)

he-IL

Hebrew (Israel)

hi-IN

Hindi (India)

hu-HU

Hungarian (Hungary)

is-IS

Icelandic (Iceland)

id-ID

Indonesian (Indonesia)

ga-IE

Irish (Ireland)

it-IT

Italian (Italy)

ja-JP

Japanese (Japan)

kn-IN

Kannada (India)

kk-KZ

Kazakh (Kazakhstan)

km-KH

Khmer (Cambodia)

kok-IN

Konkani (India)

ko-KR

Korean (South Korea)

ky-KG

Kyrgyz (Kyrgyzstan)

lv-LV

Latvian (Latvia)

lt-LT

Lithuanian (Lithuania)

lb-LU

Luxembourgish (Luxembourg)

mk-MK

Macedonian (North Macedonia)

ms-MY

Malay (Malaysia)

ml-IN

Malayalam (India)

mt-MT

Maltese (Malta)

mi-NZ

Māori (New Zealand)

mr-IN

Marathi (India)

mn-MN

Mongolian (Mongolia)

ne-NP

Nepali (Nepal)

nb-NO

Norwegian Bokmål (Norway)

nn-NO

Norwegian Nynorsk (Norway)

or-IN

Odia / Oriya (India)

fa-IR

Persian / Farsi (Iran)

pl-PL

Polish (Poland)

pt-BR

Portuguese (Brazil)

pt-PT

Portuguese (Portugal)

pa-IN

Punjabi (India)

quz-PE

Quechua (Peru)

ro-RO

Romanian (Romania)

ru-RU

Russian (Russia)

gd-GB

Scottish Gaelic (United Kingdom)

sr-Cyrl-BA

Serbian, Cyrillic (Bosnia and Herzegovina)

sr-Cyrl-RS

Serbian, Cyrillic (Serbia)

sr-Latn-RS

Serbian, Latin (Serbia)

sd-Arab-PK

Sindhi, Arabic script (Pakistan)

si-LK

Sinhala (Sri Lanka)

sk-SK

Slovak (Slovakia)

sl-SI

Slovenian (Slovenia)

es-ES

Spanish (Spain)

sw-KE

Swahili (Kenya)

sv-SE

Swedish (Sweden)

ta-IN

Tamil (India)

tt-RU

Tatar (Russia)

te-IN

Telugu (India)

th-TH

Thai (Thailand)

tr-TR

Turkish (Turkey)

tk-TM

Turkmen (Turkmenistan)

uk-UA

Ukrainian (Ukraine)

ur-PK

Urdu (Pakistan)

ug-CN

Uyghur (China)

uz-Latn-UZ

Uzbek, Latin script (Uzbekistan)

ca-ES-valencia

Valencian (Catalan, Spain)

vi-VN

Vietnamese (Vietnam)

cy-GB

Welsh (United Kingdom)

Microsoft’s commitment to responsible AI and Privacy

Microsoft has been working to advance AI responsibly since 2017, when we first defined our AI principles and later operationalized our approach through our Responsible AI Standard. Privacy and security are core principles as we develop and deploy AI systems. We work to help our customers use our AI products responsibly, sharing our learnings, and building trust-based partnerships. For more about our responsible AI efforts, the principles that guide us, and the tools and capabilities we've created to assure that we develop AI technology responsibly, see Responsible AI.

Disclaimer: Bias, Stereotyping, and Ungrounded Content

Despite robust training by OpenAI and the application of Microsoft’s responsible AI safeguards, AI-generated content is not perfect. These systems are probabilistic and may sometimes produce biased, stereotypical, or ungrounded descriptions. While Microsoft applies filtering, red-teaming, and content moderation techniques to mitigate risks, it is not always possible to prevent all inappropriate or incorrect output.

The system may generate less accurate or less useful descriptions for images that are heavily blurred, abstract, low contrast (e.g., black-and-white), or contain complex visual scenes with minimal distinguishable elements. Facial blurring, applied for privacy, can sometimes reduce the detail available in the output. The system currently supports only static image files and performs best with clear, well-lit visuals.

Users should be aware that alt text suggestions may reflect systemic biases present in training data or model behavior. For critical use cases, human-written descriptions are recommended.

For more on the known limitations of AI-generated content, see the Azure OpenAI Transparency Note, which includes references to the LLMs behind the Alt Text service.

Need more help?

Want more options?

Explore subscription benefits, browse training courses, learn how to secure your device, and more.