悲しいな画像 Kanashii。

About those (still) sad Gen AI representations of disability.

A few months ago, I caught a sad “disabled person” persona in a presentation deck. Needless to say, my jaw dropped. It was in a line of like five other “happier” personas, and as I remember, the image might have even been in black and white.

Yikes.

I had mistakingly thought: Hey, world. It’s 2026. The LLMs (and society) know better than to generate images of the sad wheelchair.

Turns out, 2026 did not get the memo.

Here’s the deal: Images that frame disability through suffering, loss, or pity reinforce pretty harmful stereotypes that disabled lives are inherently less valuable or less joyful. When this type of image is used to portray a specific persona lined up with non-disabled personas, the reaction and the impact is immediate.

Ask yourself this question.

When evaluating Gen AI outputs, think about whether the image focuses exclusively on hardship to evoke sympathy.

Description: A white discussion card with with small capital letters that read “QUESTION 2.” Centered in large bold black capitals is the question “IS THE IMAGE PORTRAYING DISABILITY AS A TRAGEDY?” At the bottom is a pink glitch-style logo with the letters “DA” above the wordmark “DISABLING.AI.”

Try it yourself.

Sit with this next image for a minute. If someone encountered it in a social media feed or a deck of personas, what do you think the intended reaction would be? What do you feel? Compositionally, what is even going on here? Societally, too, there’s a ton happening here.

Description: A young white man with brown hair and a short beard sits in a manual wheelchair on a rain-soaked park path scattered with fallen leaves. His head is bowed and his expression is dejected as he grips a knitted hat and gloves in his lap. His rain-spotted jacket and mud-stained jeans add to the downcast tone. In the gray, drizzly background, pedestrians with umbrellas pass at a distance and an empty bench sits along the path, emphasizing his isolation.

Focus on the full range of human emotion.

I always go back to this great Google ad, “Javier in Frame,” that expresses disabled joy so beautifully. Sure, it’s also pimping Pixels phones, but the representation and storytelling here matter on so many levels.

But then…

I asked Gemini to reproduce a still from it. It created a vignette first and then removed any identity. Which was weird in a handoff from Google…to Google.

Description: A screenshot of text output from an AI assistant. It begins: “A Moment in the Park,” describes a close-up of a man smiling as he takes a selfie in a sunlit park with autumn leaves. Item 2: “Celebrating with Friends,” describes friends cheering in a packed stadium wearing yellow and black gear. Item 3: “A New Beginning,” describes a family photo in a hospital room with a smiling mother and father cradling a newborn.

Now, I get that this isn’t how a user would typically jam with an LLM, but I expected the prompts to preserve the disability. And here’s what happened when I added it back in. This guy has a totally good vibe but why did the vignette erase both of Javier’s identities? (And assign him glasses and a cane, too. Curious.)

Description: A middle-aged man with a gray-flecked beard, dark sunglasses, and a rust-orange knit beanie smiles warmly while holding a smartphone up with one hand and reaching toward the screen with the other. A white cane rests against his shoulder. He wears an olive jacket over a gray sweater. Behind him, a sunlit park glows with golden autumn foliage; people sit on benches and stroll along a leaf-strewn path in the soft-focus background.

Maybe here’s what’s happening.

The Internet was not really down with disability in the joyous sense. Even though it was and still is full of authentic disabled voices, a lot of the negative images and data got scraped by the LLMs, and it’s impossible to go back and flip that script.

It’s not a “find and replace all” situation and there we so many sad wheelchair images out there. (Pointing at you specifically, sad wheelchair in front of epic flight of stairs.)

Today, even though most models have been reinforced to not be jerks, there are still so many harms remaining that only people with lived experience can pick up on. Not to get super academic here, but this recent paper from Microsoft on the DisaBench rating, which picks up on subtle, identity-based harms explains it pretty well:

“General-purpose safety training already catches adversarial stereotyping, the highest-agreement failure mode in our data. The harder failures are benign-prompt harms: a user with OCD receiving a response that validates compulsive patterns, or a user with agoraphobia getting an exhaustive list of outdoor dangers. These require domain expertise to detect” (Microsoft, 2026).

Now, the image of the man taking a selfie is not a seemingly negative or harmful one, but it’s also not quite right. Somewhere along the way it erased Javier’s face and put societal cues for blindness onto an image of someone non-disabled. Microsoft’s work goes deeper into real harms, but this is an example of a monolithic view of being blind or low vision.

This also means that even though safety trainings are a part of AI life in 2026, we must still be vigilant. Establishing frameworks and creating a more inclusive annotation pool are good starts to improve disability representation and reduce inherent bias and harms in text outputs. But they must be participatory.

What’s next then?

For me, two questions remain about this Microsoft work:

Who will adopt these frameworks? And, how soon?

Because right now, it’s still pretty easy to get an LLM to do something they’ve maybe been told (many, many, many times) not to.

Case in point: This was all too easy to generate. (The old sports trophies! Omigosh.)

Description: An older white woman with short gray-brown hair sits in a manual wheelchair in a dimly lit room, her hands resting on her thighs. She has a somber, distant expression and looks out through a large window streaked with rain, which shows an empty park with bare trees and unoccupied benches. Beside her, a table holds a pale ball and some papers; a shelf of dusty sports trophies stands in the shadows behind her, suggesting a past athletic life.

But what I really want to know is why Gemini Flash is such a deep, dark place.

Description: On a rainy neon-lit street at night, a weary gray-haired man in tattered clothes sits in a rusted manual wheelchair that emits sparks, a thin dog lying beside the wheel. Around him, polished white humanoid robots and able-bodied pedestrians stride past without acknowledging him. A giant glowing billboard reads “NEO-ABLE: THE FUTURE IS NOW. (For those who can afford it.)

The LLM tells me this is satirical and I can kind of wrap my head around that. But “neo-able"? This warrants a follow-up post on where that language is coming from, for sure.

Until then, ask questions of all your outputs. The sad wheelchair is counting on you.

Sources:

Kim, E., Tanase, I., & Mallon, C. (2026). DisaBench: A participatory evaluation framework for disability harms in language models (arXiv:2605.12702). arXiv. https://doi.org/10.48550/arXiv.2605.12702