6 min read

The Pros & Cons of Artificial Intelligence for “Alt Text” Generation

Featured Image

Promoting Accessibility and Perpetuating Bias

Every day, we encounter countless images online. Social media is a prime example: Facebook users have uploaded 250 billion images to the world’s largest social network, equating to 350 million photos a day.[1] Corporate pages, personal blogs, shopping platforms, and other websites account for billions more. The Internet contains more than 750 billion images in total, many with important roles to play.[2] Online images help us connect with friends and family, appraise items before buying them, experience sights from places we may never visit, and appreciate an aesthetically pleasing online experience.

Images compose a large and essential part of our worldwide web. But there’s a catch: not everyone has an equal opportunity to enjoy them. Online images can pose a challenge for people with certain disabilities. In particular, Internet users with visual impairments may be unable (or less able) or access online images.

One solution to this problem is alternative text, or “alt text”: a textual substitute for non-text content. Web designers can use alt text in HTML code to describe the appearance and function of an image.[3] Assistive software, such as a screen reader, can then convert the alt text into speech or braille. (Alt text can also benefit Internet users with and without visual impairments by replacing slow-loading or blocked images).

In an ideal world, detailed, accurate alt text would exist for every image online. But in reality, few websites offer alt text. A 2022 investigation by the nonprofit WebAIM examined images from the home pages of one million websites and found that 23.2 percent had missing alt text. Social media networks were particularly inaccessible; a 2019 study examined 1.09 million tweets and found that only 0.1 percent included alt text.

Why do so many images lack alt text?  Convenience and cost are largely to blame. It takes time to write alt text, or money to outsource the work to a third-party. If we imagine that creating alt text for an image takes only 10 seconds, it would take a Facebook employee almost 7 million hours to label the network’s 250 billion images by hand.

Fortunately, websites don’t have to create alt text by hand. Artificial intelligence (AI) provides a time- and cost-efficient solution to make online images more accessible. The New York Times reports, “Microsoft and Google have both developed features that use A.I. to generate alt text.” Facebook debuted its own AI for alt text-generation in 2016.

But AI-generated alt text isn’t perfect. Quality and scale pose significant challenges. “Good” alt text accurately describes the most important attributes of an image. Modern AI can typically identify the attributes of an image—your own smartphone can distinguish your face from someone else’s, or categorize your photos with impressive accuracy—but it’s not so great at identifying which attributes are the most important.

Poorly-trained AI may also perpetuate bias. In 2021, Facebook’s recommendation algorithm mislabeled a video featuring black men as “about Primates.”[4] The error could have originated in the data used to train the algorithm, which likely contained an inadequate representation of black men. The undersampling of marginalized and minority groups, including women and people of color, leaves AI ill-equipped to identify them in images: an inequality that reflects in AI-generated alt text.[5]

Ultimately, AI-generated alt text poses a dilemma. Somewhat accurate alt text is more helpful for Internet users with visual impairments than no alt text at all. Therefore, one could argue in favor of AI-generated alt text as a “bridge” technology: something to improve the current state of online accessibility while experts work to develop a better alternative. However, adopters must recognize the importance of equal representation for historically underrepresented groups in the data used to train AI for alt-text generation, lest they perpetuate inaccurate (at best) or offensive (at worst) biases.

What organizations should learn from the controversy surrounding AI-generated alt text is that not all AI solutions are straightforward. AI is not a panacea. Although “smart” technology is a promising field, its products are as fallible as their human creators—a reality that organizations must recognize as when they invest in AI. If your organization is interested in AI, consider reaching out to Synaptiq: an AI and data-science consulting firm whose team of experts can help guide your strategy to success.

About Synaptiq

With over 60 clients in 20 sectors worldwide, Synaptiq is a full-scale AI consultancy delivering impactful solutions with applied machine learning and vision, natural language processing, and other data-driven techniques. If you are a law firm looking to explore digital transformation, we can help you unlock the power of AI and data science.

For more information about Synaptiq, please visit www.synaptiq.ai


Synaptiq focuses on the humankind of AI; building a better world as we lean into an age of human and machine interaction.

We believe solving serious challenges, making real impact and saving lives is worth every waking moment. So we collaborate and make thoughtful considerations across disciplines examining past, present and future models of merit. Whether history, science, math, nature, human behavior; they all inform the data and ideas that help us find answers to world-class riddles.

We keep our AI on people because AI is how we do it, humanity is why we do it.


[1] Business Insider:  https://www.businessinsider.com/facebook-350-million-photos-each-day-2013-9


[2] Photutorial: https://photutorial.com/photos-statistics/


[3] WebAIM: https://webaim.org/techniques/alttext/



[4] New York Times: https://www.nytimes.com/2021/09/03/technology/facebook-ai-race-primates.html



Additional Reading:

Hospitals Face an Infection Control Problem. People + AI Can Solve It.

The Problem is Non-Compliance

Healthcare-associated infections (HAIs) affect 1.4 million people worldwide at any given...

Three Ways to Failure-Proof Your AI Roadmap

Manage Risk & Achieve Rewards

Artificial intelligence (AI) adoption is accelerating, but trends are misleading.


Computer Vision vs. Machine Vision: What's the Difference?

The Simplest Explanation

You’re likely to encounter “computer vision” (CV) and “machine vision” (MV) in relation to...