Tips for using AI to create image alt text

Why use AI to create image alt text?

ALT text plays an important role in making your images accessible to everyone. It also features strongly in how search engines understand and rank your posts, whether galleries, blog posts, or social media.

I struggled for a while when writing ALT text for my photos because it’s challenging talking about my work and myself. Also, I didn’t have a good grasp about how long, what to focus on, and what to avoid in ALT text.

Then I realised I can use AI to generate a first draft that I can quickly edit if necessry. Of course, with all things AI, what you receive depends heavily on the specific tool you use, and the design of the prompt you give it.

So here’s some tips I’ve gathered over the past while using various AI tools to generate ALT text for photos on my website and social media.

What is “alt text”?

Alt text, or alternative text, is a short description added to an image on a web page. It’s important for several reasons:

Accessibility for people using screen readers

Screen readers are software that read out the content of a web page. They help people who are blind or have low vision understand what a web page contains, both text and images. When a screen reader encounters an image, it speaks the alt text out loud, so the user can understand what the image is about.

Improved visibilty to search engines (search engine optimization or SEO)

Search engines can’t see images, but they can read alt text. Including descriptive alt text helps search engines understand image content, which can improve the site’s ranking in search results. Plus, search engines generally boost rankings for pages that include accessibility enhancements, like alt text.

Backup plan for images that fail to load

If an image doesn’t load due to a broken link or slow internet connection, the alt text appears in its place, explaining to users what the image was supposed to convey.

Enhanced user experience

By providing descriptions for images, you ensure that all users, regardless of their abilities or circumstances, can fully engage with your content. And you provide a fail-safe method for including the text explanations for missing images if something goes wrong when loading a page.

Alt text is a simple yet powerful tool that promotes an inclusive and positive experience for everyone.

Best practices when writing alt text

Be descriptive and specific

Clearly describe what’s in the image.

Examples:

A simple image: “A red apple with a single green leaf on a white background.”

A complex representational photo: “A grassy field, rolling hills, and distant blue mountains, all lit with golden hour late afternoon sun. The sky is a deep blue with scattered, fluffy white clouds stretching across, providing a stark contrast to the darker landscape.”

An abstract image: “A symmetrical architectural interior with an abstract, futuristic design. The scene is dominated by an intense golden glow. Two large rectangular panels extend from both sides towards the center, leaving a dark, elongated opening in the middle that adds depth to the image. The walls are geometrically patterned with converging lines and angular cutouts, enhancing the sense of symmetry and perspective. The entire structure appears to be illuminated from within, casting soft shadows and emphasizing the sleek lines and smooth textures of the surfaces.”

Keep it short

Focus on the most important aspects of an image, and avoid going into extreme detail unless the detail helps someone understand the image properly.

Of course, “short” is subjective. In the olden days, guidelines said alt text should be 125 characters or less because back then some screen readers stopped reading at that point. But nowadays, screen readers read all the alt text, although some screen readers might break up the text into paragraphs. Many social media platforms allow alt text up to 1,000 characters or more.

The essential point here isn’t character count, but rather how your visitors experience your image, even if they can’t see it. When in doubt, keep it as succinct as possible without losing important points.

Avoid obvious phrases

It’s not necessary to start with “image of” or “picture of” since screen readers already announce that it’s an image.

Example: “A dog playing with a toy…”, not “Image of a dog playing…”).

Include context when appropriate

Explain the image’s purpose or context if that helps the visitor to understand how the image relates to an article or blog/social media post.

Example: “Carrie Fisher presents a Lifetime achievement award to her mom Debbie Reynolds at the 1997 American Comedy Awards.”

Skip decorative images

For purely decorative images, don’t bother with alt text since it doesn’t add anything to the visitor’s experience.

Crafting AI prompts for alt text

As with anything AI-related, the better the prompt, the better the result. Just saying “please generate alt text for this image” is unlikely to result in helpful output.

Here’s an example from one of my images that I ran through Microsoft’s Copilot, without giving it a useful prompt.

Unfortunately, when I gave Copilot a more specific prompt for the same image, here’s how it responded:

Obviously, that picture of mountain and clouds is completely innocuous, and the AI just couldn’t compute it. So instead of saying “sorry, computer says no”, it blames the victim (me).

That glitch illustrates another problem with AI tools: automated moderation that restricts what the AI can do. In an effort to avoid generating dangerous, illegal, or offensive output, some AI vendors have moderation algorithms in place that might ban completely innocuous images.

Fortunately other AI alt text platforms are more sophisticated (see below). The Arizona State University Image Accessibility tool doesn’t seem to have any problem with any of my images (none of which are sketchy in any way). Here’s how the ASU tool responded to the same image and same prompt that Copilot found offensive (click the image to enlarge):

So even if you have a great prompt that usually works well for you, you may encounter AIs that randomly refuse to engage with it. Always have at least a couple AI alt text generator apps at your disposal to rebound from situations like this.

Challenges when using AI to generate alt text for images

Using an AI tool to create image alt text can make life a lot easier. But you need to be aware of and solve some challenges you might face. These considerations actually apply any time you use an AI tool to generate any content, whether text or images.

Here are some of the main challenges and how to mitigate them.

Challenges

Intellectual property

Your images are yours. An AI tool shouldn’t steal them to use them as training data. The terms and conditions will specify what an AI vendor promises to do with any images you upload to their AI alt text tool. Be careful.

Accuracy and Relevance

Understanding context correctly

AI might misinterpret content of the image or the context in which it’s placed. That will create inaccurate text.

Over-generalization

AI may generate alt text that’s too generic and not meaningful enough.

Insufficient details

Missing details

An AI might miss visual nuances or details in the image might be overlooked, resulting in incomplete descriptions.

Complex images

Images that include complex multiple elements can be challenging for AI to understand and describe accurately.

Cultural sensitivity

Cultural insensitivity

An AI might not (probably won’t!) understand cultural nuances or references. Of course, that will result in inappropriate or insensitive output.

Relevance to surrounding content

No AI will be able to include those aspects in its results unless you provide it with the context of the image and the intended message. Remember, “artificial intelligence” is no such thing. There’s nothing intelligent about AI. Instead, AI tools are essentially robots that are trained with data sets—text and images—that they use as references to match patterns and then generate results. AIs are computer programs, very sophisticated of course, but still the old computing adage applies whenever you work with any AI: “garbage in, garbage out”.

Tone and style

Consistency

Maintaining a consistent tone and style across all alt text entries can be challenging, especially if you use different AI tools or versions.

Tone

Without specific guidance, an AI may not generate alt text that reflects the voice and tone you want. The default is almost universally a professional and approachable tone. But you can’t assume the voice and tone will align with your preferences.

Ethical considerations

Bias

AI tools will reflect inherent biases in their training data, leading to biased or unfair descriptions. Those biases arise because the people who select the training materials for AIs are human, prone to consciously or unconsciously select training data they consider “appropriate”. Beware of small skews in wording, for example an AI might describe a White male as “stern” but a Black male as “menacing”.

I’ve written about this in an earlier article, “Is Stable Diffusion Racist?”

Accessibility

AI-generated may not always meet accessibility standards. Accessibility is a huge topic with a lot of nuance, and AI sometimes gets nuances wrong. For example, here’s the guidelines specified by the Web Accessibility Initiative: “Essential Components of Web Accessibility”.

Technical limitations

Training data

As I mentioned above, an AI’s performance is limited by the quality and scope of its training data.

Complex images

Specialized images like technical diagrams or infographics, may need more specific descriptions that AIs might struggle to provide. And AIs are universally notorious for their inability to generate text in images, and the reverse is also true—understanding text in images, especially when the text is decorative, distorted by perspective, or otherwise not crystal clear.

How to avoid and overcome challenges

Choose the right AI alt text generation tool

You’ll find a lot of AI alt text generation tools online. Which one(s) you use depend on your personal experience using them.

Even so, only use AI tools from ethical vendors who specifically say they won’t save and re-use your valuable intellectual property (your images).

Here’s what Microsoft’s privacy statement says about Copilot:

In order to provide a relevant response, Copilot will use this prompt, along with the user’s location, language, and similar settings, to formulate a helpful response. In some markets, authenticated users can choose to allow Copilot to have access to prior prompt history to better personalize the product.

https://www.microsoft.com/en-us/privacy/privacystatement#mainartificialintelligenceandmicrosoftcopilotmodule

That’s not really reassuring, because when Microsoft, Apple, or other vendors talk about “personalising” your experience, what they really mean is “take your stuff and use it however we want, even selling it to others”.

In contrast, the Arizona State University Image Accessibility tool explicitly states:

This tool currently utilizes the GPT-4o model provided by OpenAI. We are using their API platform, which ensures the data sent is not retained or stored to improve their training or models.

https://asuo-ai-labs.streamlit.app/

Read carefully and edit

As with any AI output, always (always!) closely review and edit AI-generated alt text to ensure accuracy and relevance. Beware of and correct any implicit or explicit biases. Human oversight is critical whenever using an AI tool. Remember, they aren’t actually intelligent. They’re just a tool to help you get your job done faster.

Also, editing AI-generated alt text can help you understand your images from a different perspective. That might give you insights that grow your practice.

Use good quality, appropriate prompts

In your prompts for AIs, always provide necessary context (if needed) to the AI to help it generate accurate alt text.

Using AI to create alt text for images will save time and effort. Nevertheless, it’s essential to beware of and address challenges to ensure the alt text is accurate, effective, and accessible.

What do you think?

Please let me know if you have any questions or thoughts you want to share in the comments below.

FAQ

What constitutes a good alt text description for an online image?

Clearly describe what’s in the image. Focus on the most important aspects of an image, and avoid going into extreme detail unless the detail helps someone understand the image properly. There’s no need to limit yourself to 125 characters (that’s an obsolete technical limitation of old screen readers). The essential point here isn’t character count, but rather how your visitors experience your image, even if they can’t see it. When in doubt, keep it as succinct as possible without losing important points.

How should I prompt an AI alt text generation tool?

Be specific about what you want the AI to include in the alt text. If necessary, explain the image’s purpose or context so the AI will understand how the image relates to an article or blog/social media post. Include how the alt text should be structured, eg no headings or bullet points, just a single paragraph. Keep it simple by limiting the AI to just a couple of sentences. Ask for a one sentence summary in addition to a longer alt text.

An AI thinks my image is too sensitive to process—why?

Some AI vendors have moderation algorithms in place to avoid processing images that include content that’s obscene, illegal, or offensive. If your image isn’t any of those things, then try a different AI alt text generation tool. Different vendors have different definitions of what they consider “sensitive”.