Artificial Intelligence (AI) struggles to create true-to-life images of people with amputations. Now amputees are teaching AI to be responsible and inclusive – with help from prosthetics manufacturer Ottobock and Microsoft. 

Surrounded by books, dumbbells and children’s toys, Andreas Onea is working from his living room to fill a gap in AI knowledge. Many of the most widely used AI tools simply don’t know what a person looks like, who had their arm amputated right at the shoulder.

A person like Andreas Onea. “I’m being digitally disabled,” says the 34-year-old, who lost his left arm in a car crash when he was six.

Onea lives in Mödling, a town of about 20,000 people on the southern outskirts of Vienna. Remember the recent viral trend, when people were using AI to create images of themselves as an action figure toy? Onea opens his laptop to show the disappointing results, when he tried to use AI to generate images of himself as a toy action figure. Time and again, the images were just wrong.  

“That just isn’t me,” he says. Some of the AI images show him with half an arm, or two arms, or one arm reduced to a shapeless stump. Occasionally, the AI even refused to generate any image at all, claiming that depicting a person with an amputation would breach its guidelines.  

Whether it’s hopping on the viral action figure trend or generating any AI picture of himself, Onea knows that he can’t take part.

But accepting such barriers has never been Onea’s style. He successfully fought for the right to compete in mainstream events and even won the B final of the Austrian national championships in the 200m breaststroke, racing against swimmers without disabilities. 

He went on to win medals at world championships and the Paralympic Games, became a professional athlete with the Austrian Armed Forces, and built a successful career as Austria’s “one-armed TV presenter.”  

Now Andreas Onea is taking on AI’s failure to be inclusive. 

In Mödling, his sons, aged one and three, race happily around the family home.. When they are old enough to use AI tools, Onea wants them to be able to generate images of the whole family – and that means also showing a person with an amputation or a prostheses. After all, that’s what normal life looks like.  

Of course, generating AI images of action figures may seem little more than harmless fun. But the consequences of digital visibility are anything but. 

According to the World Health Organization, 1.3 billion people worldwide live with a significant disability. That’s roughly 16 percent of the global population. 

And yet, when AI models are asked to generate images, people with disabilities are usually absent. “If you’re not visible, you don’t exist,” Onea says.

And the more important AI becomes, the more this becomes a societal problem: representation is participation, shaping both public debate and public awareness. Responsible AI, he argues, must be inclusive AI. And that’s what he is working to achieve. 

First annotate, then train: Building more inclusive AI 

At his computer, Oneas works his way through a long gallery of images. People in high-vis vests on a construction site; a civil engineer holding a tablet with a prosthetic hand; next, a man is gardening, using his prosthetic hand; then comes a close-up of a prosthetic forearm. 

Onea moderates a community of people with amputations who joined a new project: they are curating a “community library” designed to make AI more inclusive., Here, it’s the amputees who decide how they want AI to portray them.

Community members review, select and then annotate images, describing precisely what each photo shows. That helps AI models to understand images – and how to generate them.  

Onea leads the “upper limbs” team, which includes people who have had a hand, forearm or entire arm amputated. Another group covers lower limb amputations, from the foot and lower leg to the thigh. The annotated images will then be used to train AI models, so that they become more inclusive.

The project is backed by Ottobock, a world-leading prosthetics manufacturer based in Duderstadt, Germany, and by Microsoft.

The first batch of annotated images was uploaded as open-source training data to Hugging Face in July 2026 – a platform popular with AI developers to share training data, models and tools.

Any AI company can tap into this community library free of charge and use it as a training dataset. “Now we can be represented properly, accurately and inclusively – in AI models everywhere,” says Onea.

Generative AI models are normally trained using tens of thousands of images. But only a few hundred may be needed to teach them that a diverse portrayal of humanity should also include people with amputations and prosthetic limbs. 

At launch, the two datasets – for upper and lower limbs – each contain around 400 training images and up to 100 videos. More than 60 people from around the world submitted material, which was then reviewed, selected and annotated. “We have found that AI models can be trained with considerably less data, as long as the data is of high quality,” says Anja Thieme, Principal Researcher at Microsoft Research in Cambridge, UK, who oversees this AI project. “That’s why the material for the community library, including the metadata, was selected and prepared with particular care.”

Dear AI: Greetings from the podium, the red carpet and the top of the world 

Ottobock is promoting the inclusive AI initiative with its “Dear AI” campaign. Images show the world what people with amputations can achieve: climbing Mount Everest, competing at the highest levels in sport, and appearing on the red carpet as fashion influencers.

At Ottobock’s Vienna site on Brehmstraße, Chief Experience Officer Martin Böhm discusses the campaign in a room fitted with wall bars, an exercise ball and a model skeleton. On most days, this is a training room for people with amputations; they practice using their prostheses, supported by physiotherapists and specialists in orthopedic technology. Böhm has come here to illustrate the AI training program. He may look relaxed, but it quickly becomes clear that images which portray people with prostheses as cyborgs irritate him very much. “It’s simply not acceptable!” says Böhm, who’s based in Hamburg but rearranged a business trip to join this meeting. 

What matters most, he says, is that neither Ottobock nor Microsoft decide how AI represents people with amputations. “This are decisions made by the people who are affected,” he says – and insists that this cannot be a short-term project that runs for just a few months. “It’s about changing AI models in a genuinely sustainable and permanent way.”  

That’s why Ottobock needed a parter who can provide the right technical platform, says Böhm and explains that this is “where Microsoft came in; they have given us tremendous support.”

AI already has made a profound effect on the lives of amputees – maybe not in terms of digital representation, but when it comes to improving their everyday lives. “AI is a strategic pillar for us,” says Böhm. “We already use AI to support the entire patient journey.” It starts with registering new patients, continues with creating 3D body scan for a perfect fit of the prosthetic socket, supports the rehabilitation journey, and even manages the reimbursement of treatment costs.

Several of Ottobock’s digital technologies are built on Microsoft AI solutions. “Where it says AI at Ottobock, Microsoft is often part of it,” says Böhm.

A perfect fit, and technology that learns 

Andreas Onea benefits from AI advances every single day. In June, he started to wear a new arm prosthesis fitted with the latest version of Ottobock’s Michelangelo hand.  

“The system learns with me,” says a clearly enthusiastic Onea. Before, he did not wear his prosthesis while presenting on television. He worried that the prosthetic might misinterpret a muscle signal and open unexpectedly – scattering his cue cards on the floor. Now, with a broad smile, he demonstrates how he can use the new prosthesis to hold his mobile phone or read a book to his children.

He calls the new high-tech device a “game changer,” praising its carbon construction (“very sturdy, super light”) and the intuitive control system that now uses six electrodes instead of the previous two. The electrodes – silvery and the size of a coins – are fitted into the black lining of the prosthesis and are the interface between Onea’s body and the artificial limb. They detect muscle impulses and translate them into commands for the prosthesis. The more electrodes a prosthesis has, the greater the range of commands available. For this so-called myoelectric control to work, it first has to be trained. A special cuff records the wearer’s muscle impulses, while AI learns from the patterns which movements the person is trying to make. 

Using Ottobock’s connectgrip app on his smartphone – which is continuously connected to the Microsoft Azure cloud – Onea can refine and personalize how his prosthesis works. With this feedback, the system learns to distinguish between intentional and unintentional signals. A movement that means “open the hand” while sitting down could otherwise be triggered accidentally simply by walking. On his way to the television studio, Onea can use the app to switch off any functions he does not need in front of the camera – for example to make sure his cue cards remain firmly in the Michelangelo hand. 

Perfect Socket: Using AI to create the best possible fit 

Paul Peyer, who heads Ottobock’s global product management for prosthetics, sees digital technology as the key driver for improving prosthetic limbs. AI and connected prostheses – where specific components communicate with each other and get support from systems in the cloud – are pushing the boundaries of what is possible.

One example is a futuristic projected called Perfect Socket. The aim is to use AI to generate prosthetic sockets that are individually designed to precisely fit each patient.  “The socket is an extremely important and very personal component, because it connects the person to the prosthesis and is worn on the body all day long,” Peyer explains. Producing a socket today is still a complex process and prone to error, as it combines analogue craftsmanship with digital technology.

In future, AI-crafted digital models should make it possible to achieve successful fittings faster and more reliably. “Even today, we can create an individual, precisely fitting 3D model of a prosthetic socket based on our vast volume of historical fitting data. We see great potential in it for better care worldwide, and a milestone for the future of prosthetics,” Peyer says. A perfect fit means more stability, fewer wounds and pressure sores, and more precise control of the prosthesis.

Future improvements to could also reduce the number and duration of in-person visits to orthopedic workshop to make adjustments. In fact, Ottobock is exploring to deliver over-the-air software updates from the Microsoft cloud throughout the lifetime of a prothesis – similar to the updates sent to connected cars and smartphones.

Cooking at last: How AI supports inclusion in daily life

The combination of software updates, AI and the wearer’s own training of the system means the prosthesis will not remain the same product it was when first fitted. Instead, it keeps getting better.

Medical-device regulation place strict limits on what can be changed. That’s vital to ensure leg prostheses, for example, don’t learn incorrect movement patterns and cause a patient to fall.

Machine learning may only be used on the smartphone app to recognize movement patterns, explains Christoph Schlotterbach, who leads development of Ottobock’s myosmart prosthetic control system and the connectgrip app.  

The prosthesis’ small central processor itself runs on fixed and pre-programmed algorithms. For Onea, the ability to personalize and fine-tune his device represents tangible progress: 

“Every amputation is different, so every fitting is very personal,” he says. “After just a few weeks, my prosthesis is already different – and better adapted to me.”  

The new arm is now making tasks possible that many people will take for granted, but which he could not perform with his previous prosthesis. “I can now hold a bottle with the prosthesis and unscrew the cap with my other hand,” he says, “that wasn’t possible before.” He also can now be of much more help when he’s out shopping with his wife, who is pregnant with their third child. Holding the shopping bag in one hand while putting things in with the other was impossible before. 

What he’s looking forward to the most, though, is that soon he will finally be able to prepare a family meal all by himself. “I’ve never properly cooked my entire life,” Onea says. “Now it’s actually becoming possible.” Every the competitive athlete, he has already set himself a culinary goal: “A really good Bolognese – that’s what I absolutely want to make all by myself.” 

Further information

• Community Library on Hugging Face 
• Ottobock press release 
• “Dear AI” campaign website for inclusive AI 
• A new approach to AI data puts communities in charge | Microsoft Signal Blog | Microsoft