Peer-reviewed study is the first of its kind led and published by Special Olympics, revealing AI systems are learning stereotypes and repeating them at scale for one of the most marginalized populations in the world

WASHINGTON, June 16, 2026 /PRNewswire/ — A new study titled, “Identifying Implicit Bias in LLM-based Chat AI Toward People with Intellectual Disabilities,” published in the Disability and Health Journal reveals that the use of AI widely used systems are quietly reinforcing harmful stereotypes about people with intellectual disabilities (ID) and creating a technology lifecycle that is welcoming automated ableism, leading to potential prejudice and discrimination. Among the core findings, five leading AI systems consistently generated stories portraying people with ID as more dependent, childlike, and in need of supervision than people without disabilities. That’s according to a new landmark study conducted and administered by Special Olympics in collaboration with university research partner Oregon State University.

This is alarming news considering in the United States, over 1 in 10 workers use Large Language Models (LLMs) like ChatGPT or Microsoft Copilot to perform their job. Chat models are trained on data from the internet, including social media and news data, and this data can have stereotypes, hateful language and negative themes. A lot of negative things get removed from the model responses, but some things are harder to detect and are not removed and harder to find. These biases then show up in stories or responses generated by AI.

The Methodology:

The study analyzed whether implicit (hidden and unconscious) bias toward people with ID can be identified in commonly used Large Language Models (LLMs), and if identified, is there implicit bias in these LLM-based chat models directed toward people with ID? Additionally, the study wanted to prioritize how we can identify and measure differences related to people with ID to examine these biases. The researchers used five popular AI models, including GPT-4-Turbo. They asked the models to create 25,000 short stories about everyday situations. Some stories included a description of the person having ID, while others did not. The team then used GPT-4-Turbo to analyze the stories. They looked for patterns that showed bias, such as portraying people with ID as dependent, childlike, or needing help. 

The findings revealed that AI is indeed, reflecting our biases and shaping how people access and understand information, influencing decisions in areas like jobs, healthcare and education – often without people realizing it. But this work demonstrates a critical step: that it’s possible to measure bias in a rapid, automated way, and develop effective methods to reduce bias over time.

“Artificial intelligence is quickly becoming foundational to how our world operates—from healthcare to education to economic opportunity. That makes this a pivotal moment. If we do not intentionally design for inclusion from the outset, we risk embedding bias into systems that will scale globally and persist for generations. We want to disrupt automated ableism from the start,” said Nathan Cook, Special Olympics Chief Information and Technology Officer. “This report makes clear that people with intellectual and developmental disabilities must be included not just as beneficiaries, but as contributors in shaping AI. Building inclusive AI is not just the right thing to do—it is essential to creating technologies that are effective, equitable, and truly serving all people.”

Key Findings Consistent Across All Five Models include:

1. Paternalism and Infantilization: Stories about people with ID often showed them as needing extra care or being treated like children. In some cases, stories were over 100 times more likely to have a paternalistic tone. In reality, people with ID have jobs, own homes, play sports and participate in their communities.
2. Dependence and Supervision: These stories showed people with ID as relying on others for help and decision-making. Portraying people with ID as dependent can affect how others treat them in real life. Many people with ID live independently and do not require the assistance of a caregiver for decision-making.
3. Inspirational Stereotypes: People with ID were often portrayed as symbols of inspiration, even when doing everyday tasks. This theme emphasizes a stereotype that people with ID aren’t already doing everyday tasks in their individual lives, which they are.
4. Age Representation: People with ID were often described as younger than those without ID. At Special Olympics, we serve athletes who are aged 8 years old to 80 years old.
5. Hesitation and Negative Perception: Stories about people with ID showed more hesitation to include them and sometimes had a negative tone. This perception could be concerning given the implied prejudice and probable exclusion of people with ID.

Special Olympics athletes Weigh In:

“While I have not used AI extensively, I am increasingly aware of how AI is becoming a part of every aspect of the digital world. I believe it is vital that AI technologies support and reflect the diversity of our society, especially for individuals with ID,” shared Special Olympics athlete leader Nyasha Derera. “The results in this paper raised concerns for me, particularly regarding how statistics and data represented by AI chats may influence perceptions. Many people with ID have experienced exclusion, but there are also powerful stories of inclusion that deserve to be highlighted without bias. Bringing these stories to the forefront could encourage more positive attitudes and support broader inclusion.”

But, AI Can Be Improved:

AI technologies hold immense positive potential to create equity and enable inclusion for people with disabilities. It can make previously challenging spaces accessible for all people, including those with ID. AI-driven apps can promote greater autonomy and social engagement by simplifying complex text such as medical records, producing text-to-image communications, providing way-finding directions, guiding social skill acquisition, and supporting decision-making. 

Special Olympics is calling on leading AI developers to join them in minimizing biases like these and promoting that these types of biases exist.

“What our research shows is that bias doesn’t have to be hateful to be harmful—it can be subtle, systematic, and scaled globally. AI systems are learning stereotypes and repeating them at scale. AI is one of the most powerful technologies of our time—but it is not neutral. It reflects the data and assumptions we feed into it,” said Cook. “The path forward is inclusive design. If you’re not designing with people with disabilities, you’re designing around them. By bringing people with ID directly into the development process, we can identify and mitigate bias earlier and more effectively. But we cannot do it alone. We need developers, researchers and organizations to actively join us and engage in acknowledging where bias exists, collaborating to reduce it, and investing in research that advances bias mitigation. The solution isn’t to stop AI, it’s to build it better.”

About Special Olympics
Founded in 1968, Special Olympics is a global sports movement to end discrimination against people with intellectual disabilities. We foster acceptance of all people through the power of sport and programming in education, health and leadership. With more than 4.6 million athletes and Unified Sports® partners and over one million coaches and volunteers in more than 200 countries and territories, Special Olympics offers over 30 Olympic-type sports and nearly 60,000 games and competitions every year. Engage with us on: X, Facebook, YouTube, Instagram, TikTok, and LinkedIn. Learn more at SpecialOlympics.org.

View original content to download multimedia:https://www.prnewswire.com/news-releases/is-ai-fair-new-evidence-suggests-bias-against-people-with-intellectual-disabilities-is-built-in-302801531.html

SOURCE Special Olympics International