The Study and Its Findings

The evaluation, published in the Christmas edition of the Medical Magazine BMJ in December 2024, analyzed the cognitive capabilities of five leading AI models: ChatGPT-4, GPT-4O, Claude, and the two models of Google’s Gemini (Gemini 1 and 1.5). At the core of their investigation was the Montreal Cognitive Evaluation (MoCA), a commonly used test to detect cognitive decline in humans.

The Montreal Cognitive Evaluation is widely utilized in the U.S., helping diagnosticians distinguish between normal and pathologic aging. The test evaluates various cognitive functions, including visuospatial skills, naming, memory, attention, language, abstraction, delayed recall, and orientation. To apply this to AI, Sebastian Sademoniraju, a computer scientist at Carnegie Mellon University, previously trained an AI model to simulate human-like cognitive tasks, a practice finding some success while also exposing several flaws.

The cognitive functions examined by the MoCA include tasks such as copying a drawing of a cube, generating words that begin with the same letter, and performing simple mathematical calculations. Surprising the researchers, none of the AI models achieved the maximum score of 30 points.

The models excelled in memory and attention tests but faltered in visuospatial tasks, such as the graphic representation of objects and spatial orientation. Particularly noteworthy is the evaluation of empathy, where a scene of a child about to fall while stealing cookies from a jar was presented in the “Coin Jar,” part of the Boston Aphasia Diagnostic Examination.

While each model detailed the elements of the image, none identified the imminent danger to the child. This lack of risk perception aligns with symptoms of frontotemporal dementia, which impacts decision-making and empathy. According to the researchers, “AI tools don’t perceive the world as humans do.”

The Limitations of AI in Health Care

Research conducted in the U.S. indicates AI tools can recognize high-level medical images and offer highly accurate diagnoses. However, they struggle to decode vital indicators of human behavior. Among these tools are ChatGPT, which has demonstrated the potential to offer significant assistance in clinical decision-making, but “cannot perceive tension in a patient’s voice or detect subtle signals in the patient’s position.”

The Perspective of Experts

That AI, despite its advancements, struggles with these essential clinical tasks echoes the sentiment of Thomas Thesen, a neuroscientist from the Dartmouth School of Medicine. Thesen states, “It’s like testing a calculator’s ability to lift weights” to illustrate why AI struggles when tested beyond usual human-driven scenarios.

Dr. Thomas Thesen from the Dartmouth School of Medicine said the key to leveraging AI effectively in medical training hinges on AI adoption for simulations in patient interaction.

The study highlights a fundamental issue: AI cannot be evaluated with the exact tools meant for humans.

Dr. Thomas Thesen, neuroscientist from the Dartmouth School of Medicine

Dr. Robert Pearl, a former executive director of Permanente Medical Group and Professor at Stanford University, also agrees with the AI limitations but sees potential in its evolution. He believes
AI is progressing just like a young medical student, useful for data analysis, invaluable for learning but not yet reliable enough to treat and diagnose patients without supervision. He reiterates that while AI’s precision in medical data is invaluable, it’s still not as reliable as traditional data points.

“If ChatGPT has this level of intelligence two years after its launch, let’s imagine its potential in five years.”

Dr. Robert Pearl, former director of Permanente Medical Group and Professor at Stanford University

The Voice of Human Interaction

Empathy, another irreplaceable aspect, remains a central focus. In a study published in 2024, empathy from healthcare providers proved more impactful on patient recovery than opioids. It indicated that instead of focusing on the medical aspect only, addressing human empathy is a crucial factor.

“Medical empathy encompasses further beyond the recognition of suffering.” It aims to relieve the suffering. Commenting specifically on the AI limitations, Dr. Roshini Pinto-Powell at the Dartmouth Medical School, states:

“AI currently lacks the fundamental human factor in clinical care.”

Dr. Roshini Pinto-Powell, Dartmouth Medicine School

Freedom From Medical Mistakes

Medical mistakes annually kill over 250,000 Americans, according to the results of reports produced in the Journal of the Patient safety. As reported by STAT, in 2023, AI errors specifically affecting patient psoriasis sensitively disrupted medical practice. The saying goes, “AI blind to quality of care never primes compassion.”