Researchers at Duke are building artificial intelligence (AI) tools that look for patterns in children’s electronic health records (EHR)—the digital files doctors keep from routine visits—to help spot children who may be autistic earlier in their life. In the broader Duke project, funded by the National Institutes of Health, these models have predicted a later autism diagnosis from data collected in infancy, with the predictions growing more accurate as children got older.
That promise raises hard questions about fairness, trust, consent, the role of parents and caregivers, and how results should be shared and used, so the team ran an ethics study alongside the technical work. Where the larger project asks, “Can EHR-based AI models identify autism earlier?”, the ethics study asks, “How should we use these models in a way that is trustworthy, equitable, transparent, and helpful to families?”
Who Took Part
The researchers talked with 26 adults from across the United States: eight caregivers of young autistic children, eight autistic adults, and 10 clinicians involved in autism screening, diagnosis, or intervention. They joined online interviews and small group discussions between March and July 2025.
Most were cautiously optimistic, seeing a chance for AI to catch patterns that traditional screening may miss, especially for groups whose autism is more often overlooked. But they were just as clear about the risks: bias, gaps in health records, “black box” systems that are hard to explain, and the stress of very early screening.
Most participants preferred using AI predictions together with what caregivers report rather than relying on either source of information alone.
By involving the people who would be the most impacted by the potential use of this technology, this study provides valuable information about key considerations for the ethical implementation of AI-based predictive models. Sam Brandsen, a researcher on the project based at the University of Alberta, emphasized discussions about caregiver consent, the accuracy and fairness of predictive models, and the protection of patient privacy and prevention of potential misuse.
| Investigators |
|---|
Jessica Sperling, Duke Social Science Research Institute Erin Haseley, Duke Social Science Research Institute Stella Quenstedt, Duke Clinical and Translational Science Institute Lauren Driggers Jones, Duke Center for Autism and Brain Development Sam Brandsen, Department of Education, University of Alberta Adrian Brown, Duke Social Science Research Institute Perusi B. Muhigaba, Duke Clinical and Translational Science Institute Geraldine Dawson, Duke Center for Autism and Brain Development Matthew Engelhard, Duke Biostatistics and Bioinformatics Gary Maslow, Duke Psychiatry and Behavioral Sciences, Duke Pediatrics Benjamin A. Goldstein, Duke Biostatistics and Bioinformatics Lauren Franz, Duke Center for Autism and Brain Development |
The Promise of Earlier Answers
For families, the biggest draw is time, and Jessica Sperling, director of applied research, evaluation, and engagement at Duke’s Social Science Research Institute, said earlier identification means “families would have the potential for reduced lag time between initial identification and diagnosis, and thus reduced lag time to valuable supports for the child and their caregivers.”
An early autism diagnosis can have a profound lasting impact. “It can offer opportunities for greater self-understanding, self-acceptance, and access to appropriate supports and services,” Brandsen said.
Because identification today relies largely on what parents report, an AI prediction would change how the process begins. Sperling said “it could be a bit jarring to shift to a context where there is a more passive identification—in other words, where parents may not be actively involved in sharing information and experience that leads to this identification.” Most participants preferred using AI predictions together with what caregivers report rather than relying on either source of information alone.
Brandsen noted that “it’s especially critical to carefully consider both the opportunities and risks of emerging tools, like AI-based predictive models, that may allow autism to be identified or predicted at increasingly early ages.”
When the Data Fall Short
Like any screening method, an AI tool can err in either direction, sending families to unnecessary appointments or delaying support for a child who is missed. Sperling noted that “these same concerns could be noted for other current ‘non-AI’ forms of autism screening."
Bias is a deeper worry, because a model can only learn from the information it is given. As Sperling put it, “if there is missing information in current data, or biases in prior diagnostic processes, there is potential for that to be baked into the model.”
One autistic adult in the study summed up the risk: “Bias in, bias out. If the training data under‑represent girls, rural families, or minoritized groups, the model may under‑identify them.”
Matthew Engelhard, assistant professor of biostatistics & bioinformatics at Duke, said the team has watched for exactly that problem: “We’ve been very careful not to conflate autism being present with the health system recognizing it, which are two very different things.”
Participants felt screening is most valuable when it happens early, but not so early that families get results with no clear next step. A clinician in the study described what that could feel like: “If I was a parent of an infant, knowing that my child had indicators of autism and I couldn’t do anything about it right now, like there’s not even an evaluation tool, I just imagine the impact on caregiver stress would be enormous.”
“Our systems aren’t trying to make a diagnosis,” Engelhard explained. The models estimate how likely a child is to be diagnosed in the future, and a diagnosis still requires a detailed clinical assessment.
Participants also wanted clear explanations, and Sperling said it could be relevant to tell families that “machine learning models were used to generate a prediction,” though she sees “a real challenge for health systems in asking clinicians, in the limited time they have, to also become AI experts and educators themselves.”
Who Decides, and What Comes Next
Health systems will decide whether and how these tools are used, but participants also debated whether parents should get to choose. “There was definitely support for parental decision to have an AI tool used for their child’s care, but there was also some concern that this could limit the value of this tool for some children,” Sperling said.
Sperling wants to see close attention paid to how clinicians use the information, how families understand it, and what happens as a result. “Taking a lens on ethics and safeguards, we need to be sure that we are prepared for each of these to happen appropriately—and we need to be prepared to have a way to determine when this is not the case,” she said.
Even once the tools are in place, that work won’t end. “Electronic healthcare documentation is always changing, so we’ll need to keep updating our models over time to make sure they remain effective and unbiased,” said Engelhard.
Read More
This article is based on the following open-access publication in the journal BMC Medical Ethics in October 2026:
Sperling, J., Haseley, E., Quenstedt, S. et al. AI–based clinical prediction models for early autism detection: ethical considerations and use implications across caregivers, autistic adults, and clinical providers. BMC Med Ethics (2026). https://doi.org/10.1186/s12910-026-01634-w