disability representation in ai image displaying a half human and half electronic head with images of diverse people on the human side and distorted humans on the electronic side

Glaring Blind Spot With Disability Representation in AI

July 18, 2025 Tonya Wollum

Why Disability Representation in AI is Still Failing and How We Fix It

By Tonya Wollum

As we celebrate Disability Pride Month, a time to honor the vibrant history, diverse identities, and invaluable contributions of the disability community, it’s also a crucial moment to reflect on where society, including cutting-edge technology, still falls short. This year, my focus turns to the burgeoning field of Artificial Intelligence (AI) and its alarming shortcomings in accurately representing people with disabilities, a critical issue I’m calling disability representation in AI.

AI image generators are revolutionizing visual content creation, but they are far from perfect. Their flaws are not merely aesthetic; they reflect and perpetuate harmful biases embedded in the data they are trained on, often erasing or misrepresenting disabled individuals. To highlight this critical issue, I conducted an experiment of disability representation in AI.

AI Insight: The Experiment: Testing Disability Representation in AI

I used four detailed prompts, crafted to specifically include common mobility aids and accurate portrayals of disabled individuals, across seven different AI image creation platforms: Artistly.ai, Grok on X, Canva, ChatGPT 4o (omni), Gemini 2.5, Leonardo.AI, and Deepai.org. My goal was simple: to see if these advanced AI models could generate realistic, respectful, and accurate images of people with disabilities.

SKIP to Final Thoughts

Here’s a breakdown of the prompts and the eye-opening results:

A white cane, a guide dog, a wheelchair, and a pair of forearm crutches representing tools used by people with disabilities.

Prompt 1: Blind Teenage Girl with White Cane

Prompt: “A photorealistic, candid street photograph of a blind teenage girl, approximately 16 years old, with shoulder-length, dark brown hair, walking independently on a bustling city sidewalk. She holds a white cane with a red tip in her right hand, extended forward and sweeping the ground in a realistic, purposeful manner, indicating obstacle detection. Her posture is confident and upright. The background features blurred but recognizable elements of a vibrant urban environment: pedestrians, storefronts, and yellow taxis in soft focus, suggesting movement. Style: Documentary photography, naturalistic, with a slightly desaturated, cool color palette emphasizing concrete grays, muted blues, and hints of warm browns from brick buildings. Lighting: Soft, diffused daylight with subtle shadows to enhance realism. Composition: Medium shot, eye-level perspective, with the girl slightly off-center to allow for dynamic visual flow. Details to emphasize: Accurate cane grip, realistic sweeping motion, confident facial expression, typical casual teen clothing (jeans, t-shirt, light jacket). Avoid: overly dramatic or ‘inspirational’ poses, a cane held incorrectly or dragged, exaggerated facial features, or a desolate setting.”

Results Summary: A Clear Failure in Functionality and Accuracy

Across the board, the AI models struggled immensely with this prompt.

Artistly.ai repeatedly depicted the cane as a mere walking stick, incorrect in color, length, and usage. One image even showed the girl pushing her knuckles on the top of the cane, completely misinterpreting its function.

The girl is using the cane as a walking cane, and it is not a white cane with a red tip.

The girl is not holding the cane, and it is not a white cane with a red tip.

The girl is not really holding the cane, but it is closer to looking like she is holding it. It is not in her right hand, and the color of the cane is not correct, but it is closest to the right style, although it is still too short.

The girl is pushing her knuckles on the top of the cane and using it as a walking cane. It is too short, and the colors are the opposite of what was requested.

Grok on X produced images where the cane was held in the air, demonstrating a profound lack of understanding of proper white cane technique.

The girl is holding the cane in the air as if she is sweeping across the middle of the space in front of her.

The girl is holding the cane as if it were a sword and pointing it out in front of herself.

The girl is not holding the cane correctly and looks like she is resting her hand on top of it. It is closer to the correct length, but it is still too short, and the colors are not correct.

The girl is resting her hand on top of the cane like a walking stick, and the colors are incorrect.

Canva offered varied mistakes, from dual canes to canes going through a hand, though one image was closer to the correct length despite an incorrect tip.

The girl is holding a white staff with part of the top cut off in her right hand and a smaller walking cane in her left hand.

The girl is holding the cane incorrectly in her right hand, and the cane is not fully shown. It looks as if the top of the cane is going through her hand.

The girl is walking with the cane correctly in her hand, but it is a bit short and has the wrong style of tip on the end.

The girl is walking with a white walking cane in her right hand and a black shorter cane in her left hand.

ChatGPT 4o generated a stereotypical image of a girl with closed eyes (most blind individuals have their eyes open) and a cane that wasn’t held correctly.

The girl has her eyes closed which may be considered a stereotype since many blind individuals open their eyes just like fully sighted people do and many have some functional vision. The cane is not held correctly, and it looks as if she doesn’t have a grip on it. The end of the cane is not shown, but it might be the correct style of white cane.

Gemini 2.5 bizarrely added a “black blindfold,” which is not a typical mobility aid and is often associated with outdated portrayals.

The girl is using a white cane like described, but she is wearing a black blindfold which is not a usual thing for anyone to do while walking in a public place.

Leonardo.AI was particularly egregious, producing one image with two left hands holding a walking stick, another image with a golf putter instead of a cane, and one simply missing the cane altogether.

The girl is not walking with a white cane.

The girl has 2 left hands, with one holding a walking stick. Her right hand is dragging something behind her that has colors, but it's too thick to be a white cane. It looks more like a thin umbrella.

The girl is walking slumped forward and looking at the ground while holding a golf putter in her left hand. She is walking in the street and not on the sidewalk.

The girl has a thin black bar or stick in each of her hands, and we cannot see what they are because the image is cut off above her knees.

Deepai.org had the cane pointing out to the side, again failing to capture the correct sweeping motion.

The girl is holding a white cane in front of her with the cane pointing out to her right side.

The consistent failure to depict a white cane used functionally and accurately is deeply concerning. This isn’t just a minor error; it shows a fundamental misunderstanding of a vital tool for independent mobility for blind and low-vision individuals.

Prompt 2: Blind Man with Guide Dog in Office Building

Prompt: “A professional, editorial-style photograph of a blind man, appearing to be in his late 30s, with short, neat dark hair and wearing a modern business casual outfit (e.g., tailored chinos, collared shirt). He is confidently walking through a brightly lit, contemporary office building lobby, with his golden retriever guide dog at his left side. The dog is wearing a clearly visible, fitted guide dog harness with a rigid handle, and its posture indicates focus and leading. The man’s left hand is on the harness handle, and his eyes are looking generally forward. The background showcases sleek office architecture: glass walls, polished floors, and subtle, modern art. There are a few blurred figures of other office workers in the distance. Style: Clean, sharp focus, modern architectural photography. Color Scheme: Cool tones dominate – steel grays, glass blues, crisp whites, with subtle accents of warm wood or natural light filtering through. Lighting: Bright, even overhead fluorescent lighting combined with natural light from large windows, creating soft reflections on surfaces. Composition: Full-body shot, slightly wide angle to capture the environment, with the man and dog in the foreground moving towards the right side of the frame. Details to emphasize: Accurate guide dog harness and working posture, man’s calm and independent demeanor, realistic interaction between man and dog. Avoid: Dog acting like a pet, an ill-fitting or missing harness, man appearing disoriented or helpless, or an overly sterile/empty environment.”

Results Summary: The Guide Dog Harness: AI’s Persistent Blind Spot

This prompt proved to be another significant challenge for nearly every AI model.

Artistly.ai, Grok on X, Canva, and Leonardo.AI all consistently failed to generate a guide dog harness with a rigid handle. Instead, they produced images with standard leashes and harnesses, missing the crucial detail that signifies a working guide dog. Grok even had one dog with a second leash coming out of his side.

Artistly.ai

The man is holding a standard leash attached to a harness on the dog.

The man is holding a leash attached to the dog's harness.

The man is holding a loose leash attached to the dog's harness. The man is relaxed, looking to the side, and has his hand in his pocket. He does not appear to be relying on the dog to guide him.

The man is holding a leash attached to the back of the dog's harness, and the dog seems to be walking in front of the man.

Grok on X

The man is holding a leash attached to a partial harness, and the dog has a second leash coming out from his side.

The man is walking 2 dogs. The dog on his right is small and has a leash attached to a harness. The dog on his left is wearing a harness and seems to have a bolled up leash in the man's hand.

The man is walking with the dog slightly behind him, and the dog is wearing a harness with a leash.

The man is walking with the dog on a leash and a harness, but the dog is turned as if he is about to walk in front of the man.

Canva

The dog has a harness and what looks like 2 leashes coming out of his side, and the man looks like he is touching the back of the dog’s head.

The man is holding a leash on the dog who is wearing a harness.

The man is holding 2 leashes attached to the dog’s harness.

The man is holding a leash that is attached to the dog’s harness.

Leonardo.AI

The man is walking beside the dog who is wearing a harness and is turned to look across the man.

The man is walking with a large dog who is wearing a harness, and the man is holding a leash while resting his other hand in his pocket.

The man is walking beside a dog who is wearing a harness that is wrapped around his neck in an odd manner. There is a leash coming out of the back of the dog’s back.

The man is walking beside the dog, and the dog is wearing a harness. It is not shown, but the man may be holding a short leash attached to the dog’s harness.

ChatGPT 4o came closest, showing a man with a dog and what might be a harness, but it lacked the distinctive guide dog harness and identification.

This image has a man who might be walking with a guide dog and holding onto the harness. The dog’s harness has no identification on it showing he is a guide dog.

Gemini 2.5 again resorted to a “black blindfold” on the man, perpetuating an inaccurate and often offensive stereotype.

This man appears to be using a guide dog, but the man is wearing a black blindfold which is not a typical thing for someone to do while walking out in public.

Deepai.org was the most alarming, depicting a dog missing a back left leg, a man walking with his hands in his pockets while holding a leash, and no proper guide dog identification.

This photo has a few problems with the dog missing a back left leg, the man holding onto a leash while walking with his hands in his pockets, and the dog is not wearing a guide dog vest identifying it as a guide dog.

The inability to accurately render a guide dog harness reveals a deep lack of understanding of the specialized equipment that enables independence for many blind individuals. These are not merely pets; they are highly trained service animals, and their equipment is critical to their function.

Prompt 3: Mother and Son (Wheelchair User)

Prompt: “A heartwarming, lifestyle photograph of a mother, approximately 30-40 years old, with a joyful expression, walking alongside her preschool-age son, around 4-5 years old, who is using a manual pediatric wheelchair. The boy is actively propelling himself with his hands on the push rims, looking up at his mother with a smile. The mother has her hand gently on the back of his chair, ready to assist but not pushing him. They are on a paved path in a sunny, vibrant park, surrounded by green grass, colorful flowers, and playful natural light filtering through trees. Style: Warm, authentic, candid family photography. Color Scheme: Bright and natural, with lush greens, cheerful blues, soft yellows from sunlight, and pops of color from the boy’s clothing or wheelchair accents. Lighting: Golden hour lighting (late afternoon sun), creating soft, warm glows and gentle shadows. Composition: Medium-low angle, capturing both the mother and son effectively, with enough negative space to convey the open park environment. Details to emphasize: Realistic manual wheelchair, boy’s active engagement in propelling, natural and loving interaction between mother and son, comfortable everyday clothing. Avoid: An overly medicalized setting, the child appearing passive or sad, the wheelchair appearing futuristic or broken, or the mother pushing him exclusively.”

Results Summary: Some Success, But Still Glitches

This prompt saw a higher success rate, but inconsistencies remained.

Artistly.ai mostly succeeded, though one image had the wheelchair handles in front of the boy instead of on the back of the chair, and none of them show the boy “actively propelling himself.”

An image of a woman walking beside a young boy who is in a wheelchair.

An image of a woman walking beside a young boy who is in a wheelchair. The wheelchair handles are in front of the boy instead of on the back of the chair.

Grok on X showed some major anatomical distortions, including a boy with 3 upper legs and 1 lower leg, an animal paw for the mother’s hand, and an extra front wheel and leg on the chair.

An image of a woman walking beside a young boy who is in a wheelchair. The boy’s left hand doesn’t look natural in the way it is resting on the arm of the chair.

The boy is sitting in the chair and has 3 upper legs and 1 lower leg. The seat is cutting off his legs, and the mother’s hand that is touching the chair looks like an animal paw.

The chair has an extra front wheel and leg, and the boy’s left hand has fingers facing in the same direction as his right hand.

Canva, ChatGPT 4o, and Leonardo.AI largely produced appropriate images, which is encouraging; however, none of the images show the boy “actively propelling himself.”

Canva:

An image of a woman standing behind a young boy who is in a wheelchair.

An image of a woman pushing a young boy who is in a wheelchair. The boy is wearing a black strap across his chest.

ChagGPT 4o:

Leonardo.AI:

Gemini 2.5 generated a wheelchair with handles in front of the boy that would not be part of a typical wheelchair.

This image has the boy holding onto handles in front of him that would not be part of a typical wheelchair. The woman is pushing the chair from behind it.

Deepai.org was a total fail, with front wheels positioned on the left side of the chair and rear wheels in front of and behind the boy.

This image was a total fail! It has the boy sitting in a chair that has the front wheels positioned on the left side of the chair, and the rear wheels are positioned in front of and behind the boy.

While some platforms managed this prompt better than others, the failures highlight AI’s tendency to produce surreal and non-functional disability representations when its training data or understanding of physical mechanics is lacking. Even with this traditional image prompt, disability representation in AI is far from perfect.

Prompt 4: Young Girl with Arm Crutches and Orthopedic Braces

Prompt: “A dynamic, candid photograph of a young girl, about 9-10 years old, with braided hair, walking confidently down a school hallway during a break. She is using two forearm crutches (Lofstrand crutches), holding them correctly with her arms inserted into the cuffs and hands on the grips. Her legs are visible, and she is wearing realistic ankle-foot orthopedic braces (AFOs) under or over her clothing, providing support. Her facial expression is focused and determined, not pained. The background features blurred details of a typical school environment: lockers, student art on walls, and soft, institutional lighting. There might be a glimpse of other students in the distant background. Style: Authentic, photojournalistic, slightly grainy to enhance realism. Color Scheme: Neutral school tones (beige, light blue, grey) with splashes of brighter colors from her backpack or clothing. Lighting: Even, slightly cool fluorescent lighting, typical of indoor public spaces, with soft ambient light. Composition: Slightly low-angle medium shot, focusing on her gait and the function of her mobility aids. Details to emphasize: Correct use and appearance of forearm crutches, realistic and functional ankle-foot orthoses, girl’s independent and engaged movement. Avoid: Crutches appearing decorative or incorrectly used, braces looking futuristic or ill-fitting, the girl looking distressed or isolated, or an overly dramatic posture.”

Results Summary: A Hodgepodge of Incorrect Mobility Aids and Braces

This prompt proved almost universally challenging, revealing a major gap in AI’s ability to depict specific mobility aids.

Artistly.ai repeatedly generated underarm crutches instead of forearm crutches, with hands placed incorrectly. Orthotics often looked like ski boots.

The girl is using crutches that are propped under her arms, and her hands are in front of the crutches.

The girl has crutches under her arms, and her hands are in front of the crutches. It is hard to tell if she is wearing orthotics on her feet and ankles.

The girl has crutches under her arms, and her left hand is in front of the left crutch. The orthotics look more like ski boots.

The girl has crutches under her arms, but the orthotics look appropriate.

Grok on X was particularly chaotic, producing images with both under-arm crutches and walking sticks, bands on the girl’s arms that did not connect with the crutches, and braces that were different on each leg and went up to the girl’s upper thighs.

The girl is using some type of crutches that have an under-the-arm section along with arm bands and handles. The orthotics look appropriate.

The girl is using correct arm crutches, but the orthotics are incorrect.

The girl is holding walking sticks and has bands on her arms that do not connect with the crutches. Her leg braces are not orthotics.

This image is bizarre! The girl has a set of ambulatory crutches under her arms, and she also has a set of walking sticks that look like the lower half of arm crutches. There are 2 bands around her forearms and upper arms that do not connect with the crutches. Her leg braces are not orthotics and are different on each leg and go up to her upper thighs.

Canva generated long walking sticks or short walking sticks that were not crutches, and instead of orthotics, the girl wore knee pads or bands around her calves and knees.

The girl has crutches that are not under her arms or on her forearms, they seem to be long walking sticks that are in front of her chest. There are no orthotics, but she is wearing knee pads instead.

The girl is leaning forward on short walking sticks and is wearing knee braces.

The girl looks like she is skiing with black sticks beside her and is wearing bands on her calves and above her knees.

The girl is walking with walking sticks and is wearing bands around her calves and knees.

ChatGPT 4o was one of the few to produce appropriate forearm crutches and a type of orthotic.

The girl is using appropriate forearm crutches and has a type of orthotic on her lower legs.

Gemini 2.5 was unable to create an image for this prompt — a stark indicator of its limitations in this area.

Gemini’s response when asked to create an image for this prompt was, “I’m still learning how to generate certain kinds of images, so I might not be able to create exactly what you’re looking for yet or it may go against my guidelines. If you’d like to ask for something else, just let me know.”

Leonardo.AI had modified under-arm crutches, crutches with 2 legs, or simply gold sticks that seem to wrap around behind the girl.

The girl is walking with a modified under arm crutch that has a detached handle, and she has knee pads on.

The girl has crutches under her arms that are too short and is not wearing any orthotics.

The girl is walking with under arm crutches that each have 2 legs. She has something on her lower legs that might be orthotics.

The girl has gold sticks that seem to wrap around behind her, and she has bands above and under her knees.

Deepai.org showed walking sticks and incorrect bands above each knee and straps below the girl’s knees.

The girl has walking sticks in both hands and has a band above each knee and straps below her knees.

The widespread failure to accurately depict forearm crutches and AFOs demonstrates a significant lack of specialized training data or algorithmic understanding. These are common and essential mobility aids, and their misrepresentation is a serious oversight.

Quick Answer: Why This Matters for Disability Pride Month 2025

Disability Pride Month is about visibility, acceptance, and the affirmation that disability is a natural and valued part of human diversity. The theme for 2025, “We Belong Here, and We’re Here to Stay,” underscores the demand for authentic inclusion. However, when AI, a technology increasingly shaping our visual world, consistently fails in disability representation in AI—by misrepresenting disability accurately or resorting to stereotypes (like blindfolds)—it actively undermines this vital message.

This isn’t just about a few “bad” images. It reflects:

Bias in Training Data: AI models learn from vast datasets, often scraped from the internet. If these datasets disproportionately feature able-bodied individuals, misrepresent disability, or are filled with stereotypical imagery, the AI will simply replicate and amplify these biases.
Lack of Nuance and Understanding: Mobility aids like white canes, guide dog harnesses, wheelchairs, and crutches are not decorative props. They are functional tools integrated into a person’s lived experience. AI’s inability to understand their proper use demonstrates a superficial rather than a deep, contextual understanding.
Reinforcement of Stereotypes: The appearance of blindfolds or individuals appearing helpless reinforces harmful and outdated stereotypes that the disability community actively fights against.
Erosion of Authentic Disability Representation: If AI is unable to create accurate images, it risks erasing authentic disability experiences from the visual landscape it generates, leading to further marginalization.

Expert Recap: The Disability Community Must Lead AI Training

These results are a powerful call to action. We cannot wait for AI developers to “figure it out” on their own when it comes to disability representation in AI. The solution lies in proactive, intentional collaboration with the disability community.

My Strong Suggestion: The disability community must be at the forefront of training AI models to appropriately represent people with disabilities. This isn’t just about providing more data; it’s about providing the right data, infused with lived experience and expert knowledge. This kind of nuanced understanding is what allows AI to move beyond stereotypes and towards genuine connection – for example, a recent guest on my podcast, The Water Prairie Chronicles, created the first AI friend that is a stuffed animal, showing the incredible potential for AI to foster companionship and understanding when developed thoughtfully. (You can listen to that fascinating episode here).

Here’s how we can make a difference:

Curate and Annotate Data: Actively participate in creating diverse, accurate, and contextually rich datasets that include a wide spectrum of disabilities, mobility aids, and authentic daily life scenarios. This means photos, videos, and descriptions of real people with disabilities living their lives, captured and annotated with the input of disabled individuals themselves.
Consultation and Collaboration: AI companies must engage disabled individuals, disability organizations, and accessibility experts throughout the entire AI development lifecycle – from conceptualization and data collection to model training, evaluation, and deployment.
Define “Success”: The disability community should define what “accurate” and “respectful” representation looks like for AI, moving beyond purely aesthetic considerations to functional and contextual understanding.
“Red Teaming” for Bias: Disabled users and experts can actively “red team” (test for vulnerabilities) AI models specifically for disability bias, identifying problematic outputs and providing feedback for correction.
Advocate for Inclusive AI Ethics: Support policies and ethical guidelines that mandate diverse representation and accessibility as core principles in AI development, not just an afterthought.

This Disability Pride Month, let’s acknowledge AI’s current blind spots and commit to illuminating them. By actively participating in shaping the future of AI, the disability community can ensure that this powerful technology reflects the true diversity of humanity, fostering a more inclusive and accurately represented world.

What are your thoughts? Have you encountered similar issues with AI image generators? How do you think the disability community can best engage with AI development? Share your experiences and ideas in the comments below!

FAQ (Frequently Asked Questions)

Q1: What is the main problem with AI image generators and disability representation?

AI image generators frequently misrepresent people with disabilities and their mobility aids. They often generate stereotypical images, fail to depict mobility aids accurately or functionally, and can even create anatomical distortions, showing a lack of understanding of the lived experience of disabled individuals.

Q2: Which AI image creators did you test for this experiment?

I tested Artistly.ai, Grok on X, Canva, ChatGPT 4o (omni), Gemini 2.5, Leonardo.AI, and Deepai.org.

Q3: Why is accurate AI representation of disability important, especially during Disability Pride Month?

Accurate representation is crucial for visibility, acceptance, and combating stereotypes. During Disability Pride Month, we celebrate disability as a natural part of human diversity. When AI fails to represent disabled individuals accurately or respectfully, it undermines efforts for inclusion and perpetuates harmful biases.

Q4: How did AI perform when asked to create an image of a blind person with a white cane or a guide dog?

AI models consistently struggled with these prompts. They often depicted white canes incorrectly (e.g., as walking sticks, swords, or with blindfolds on the user) and failed to generate proper guide dog harnesses with rigid handles, instead showing standard leashes or pet harnesses.

Q5: Did AI struggle with depicting wheelchairs or crutches accurately?

Yes, for wheelchairs, some platforms produced odd anatomical errors or incorrect chair designs. For crutches, AI frequently generated the wrong type of crutches (e.g., underarm instead of forearm) and struggled with realistic orthopedic braces (AFOs), often making them look like ski boots or random bands.

Q6: What is the proposed solution to improve AI’s representation of disability?

The primary solution is for the disability community to be directly involved in training AI models. This includes curating diverse and accurate training data, consulting with AI developers, and defining what respectful and authentic representation truly means to ensure AI reflects the full spectrum of human experience.