AI speech screening tools falsely flag multilingual speakers as cognitively impaired at 2.5x the rate of monolingual users

Artificial intelligence models trained to detect early cognitive decline through speech analysis disproportionately misidentify multilingual speakers as impaired, raising urgent concerns about health inequities in diverse populations, according to research submitted to arXiv.

The study#

Researchers recruited 1,395 participants across Sheffield and Bradford—a mix of monolingual English speakers and multilingual individuals who also spoke Somali, Chinese, Hindi, Urdu, Punjabi, Mirpuri, or Arabic. Over 263 hours of conversational speech were collected through the CognoMemory agent. The team then evaluated three automatic speech recognition systems (Whisper, Wav2Vec 2.0, NeMo) and downstream AI classifiers designed to flag cognitive impairment and estimate cognitive scores based on speech patterns.

The disparity#

While the speech recognition systems themselves worked equally well across language groups, the downstream AI models showed stark differences. Multilingual participants were substantially more likely to be wrongly classified as cognitively impaired—with false-positive rates between 28 and 37% for multilingual speakers compared to 12 to 16% for monolingual peers. In practical terms, multilingual individuals faced roughly 2.5 times higher risk of receiving an incorrect impairment diagnosis, particularly on memory, fluency, and reading tasks.

The problem intensified when models were trained on DementiaBank, a widely-used research dataset, suggesting the training data itself may embed these biases.

Why it matters#

The findings arrive at a critical moment. Dementia is projected to grow fastest among Black and Asian communities in the UK—populations where multilingualism is prevalent. Current AI screening tools, despite showing strong overall performance, could inadvertently harm the very groups most in need of equitable healthcare. Deploying these models without addressing the bias risks unnecessary worry, further medical testing, or delayed diagnosis for people already navigating complex healthcare landscapes.

This marks the first large-scale examination of false-positive bias in speech-based cognitive screening for UK multilingual ethnic minorities, highlighting that technical excellence in one dimension (speech recognition) does not guarantee fairness downstream.