Can AI Smell Disease? Vasant Dhar on AI in Healthcare

False Fact Fixed

Can AI Smell Disease Before Symptoms Appear? Vasant Dhar on AI, Smell, and the Future of Medicine

00:00 / 1:01:58

Can AI Smell Disease Before Symptoms Appear?

What if disease leaves a biological signature before a patient develops obvious symptoms—and artificial intelligence could learn to recognize it?

That question sits at the intersection of artificial intelligence, machine learning, biology, and medical diagnosis.

In his conversation with Randy Bock, NYU professor and artificial intelligence researcher Vasant Dhar describes a remarkable journey that began with an early medical AI system and eventually led him to investigate whether machines can learn to recognize disease through smell.

The idea sounds almost futuristic: Can we literally sniff out disease with AI?

Dhar’s answer is not that medicine has already solved the problem. Instead, he describes an experimental approach in which biological responses to odors become data that machine-learning systems can analyze.

The concept is important because it illustrates something much bigger about AI: machines may discover patterns that humans struggle to perceive, measure, or remember.

Research outside the interview also shows that AI-based olfactory systems are being investigated for Parkinson’s disease detection, including systems that analyze volatile organic compounds and use machine-learning models to distinguish patients from controls.

Vasant Dhar’s Journey Into Artificial Intelligence

Dhar’s path into AI began almost accidentally.

In 1979, while he was a young doctoral student at the University of Pittsburgh, another student invited him to visit a medical laboratory. There, Dhar witnessed a physician interacting with a computer system called Internist.

The machine was being used to reason through a real medical case.

At one point, the computer asked the physician a question. When the physician asked why that particular question mattered, the system explained that the evidence provided so far was consistent with a particular hypothesis and that the question could discriminate between the leading possibilities.

For Dhar, this was an intellectual shock.

He had previously thought of computers primarily as calculating machines. Suddenly, he was watching a computer participate in a form of medical reasoning.

That experience became his entry point into AI.

Dhar later joined NYU and continued working in artificial intelligence. In 1994, he brought machine learning into Wall Street and subsequently developed machine-learning-based systematic trading.

His career eventually returned to medicine—but through a very different sensory pathway: smell.

From Machine Learning to the Sense of Smell

Smell presents an unusual challenge for artificial intelligence.

A photograph can be represented by pixels. A sound can be represented as a waveform. But how do you objectively describe what an orange, coffee, cheese, or lemon actually smells like?

Human language is imperfect at describing smell.

As Dhar explains, we can say an orange smells more like a lemon than a pizza, but those descriptions don’t capture the intrinsic experience of the smell itself.

This creates a problem for conventional machine learning.

A machine-learning system generally needs some form of data and a target or label. If the target is subjective or poorly defined, the quality of the training process can suffer.

Dhar’s approach attempts to get around that problem by turning biology itself into the source of truth.

How Could AI Learn What Something Smells Like?

Dhar describes an approach involving the olfactory system of mice.

The basic idea is fascinating.

When a mouse smells a substance, researchers can observe activity in its olfactory bulb. The resulting biological response produces a time-series pattern.

That pattern becomes the biological reference—the “ground truth”—for what the organism detected.

Machine learning can then be trained on those biological signals.

In simplified form, the process looks like this:

Molecule or mixture → biological olfactory response → measurable signal → machine learning → predicted odor

The interview explains that the system can learn from these biological responses and predict what the mouse is smelling, such as particular chemical compounds.

The next question is much more consequential:

Could the same principle help identify disease?

Could AI Detect Disease Through Smell?

This is where the research becomes particularly intriguing.

Dhar proposes that disease could potentially produce a different olfactory response from that of a healthy control.

If that difference can be measured consistently, machine learning could potentially learn the pattern.

In the interview, he describes the proposed sequence:

Observe biological responses to smells.
Convert those responses into measurable data.
Train machine-learning models on the biological signals.
Determine whether disease produces a distinguishable pattern.
Eventually investigate whether those patterns can be used for disease identification.

The transcript is careful to describe this as an ongoing research direction rather than a finished diagnostic technology. Dhar says the next step is determining whether the approach can actually identify disease.

That distinction matters.

The possibility of AI-based disease detection is not the same thing as having a clinically validated diagnostic test.

Current research nevertheless demonstrates that artificial-intelligence olfactory systems are being investigated for Parkinson’s disease. One published study used volatile organic compounds from skin sebum and machine-learning models to distinguish Parkinson’s patients from healthy controls.

Another study developed an artificial intelligent olfactory system using gas chromatography, sensors, and machine-learning algorithms to investigate Parkinson’s diagnosis through odor profiles.

These studies illustrate why Dhar’s idea deserves attention without requiring us to assume that the technology is already ready for routine clinical diagnosis.

Parkinson’s Disease and the Promise of Early Detection

Parkinson’s disease is particularly interesting in this context because biological changes may occur before conventional clinical diagnosis.

Researchers have therefore investigated whether chemical signatures associated with Parkinson’s disease can be detected through biological samples.

One 2024 study of an AI olfactory-like system analyzed volatile organic compounds from skin sebum. The researchers reported that their system could distinguish Parkinson’s patients from healthy controls, although it did not successfully predict disease progression by clinical stage.

This is an important lesson about AI diagnostics.

A model can perform well at one task without necessarily being capable of answering every related medical question.

Detection is not the same as prognosis.

Association is not necessarily causation.

And statistical accuracy in a research setting does not automatically mean clinical usefulness in the real world.

Those distinctions should remain central as AI moves deeper into medicine.

Why Biology May Be Better Than Language

One of the most interesting ideas in Dhar’s discussion is that biology may provide a better description of smell than language does.

Human beings experience smells directly, but we describe them using words.

Words such as:

fruity
woody
pungent
cheesy
citrusy

are useful approximations, but they aren’t the smell itself.

Dhar’s research instead attempts to observe what happens inside a biological olfactory system.

That changes the machine-learning problem.

Rather than asking:

“How would a human describe this smell?”

the system asks:

“What biological pattern does this stimulus produce?”

That is a fundamentally different approach.

The transcript describes the biological response as a time-series signal involving activity in the olfactory bulb. Machine learning can then operate on that observed biological pattern.

This approach also illustrates a broader principle in AI research:

Sometimes the best way to understand a complex phenomenon is not to translate it into human language first.

Instead, researchers can measure the phenomenon directly.

AI’s Real Superpower: Finding Patterns

The smell research connects to a theme that runs through Dhar’s entire career: machines can discover patterns before humans understand why those patterns exist.

Dhar describes an early machine-learning project involving consumer data from tens of thousands of households.

The algorithm identified unusual consumption patterns. One example was that older women in the Northeast appeared to shop more heavily on Thursdays.

The explanation turned out to be simple: Thursday was coupon day.

The machine found the pattern first. The human supplied the explanation afterward.

The same phenomenon appeared when Dhar analyzed trading data at Morgan Stanley.

His machine-learning analysis found that trading performance was dramatically better when 30-day volatility was in its lowest quartile.

The traders knew that volatility caused them problems. What the algorithm did was quantify the relationship.

Dhar’s larger realization was profound:

Patterns can emerge before the reasons for those patterns become apparent.

He describes the machine as becoming a generator and tester of hypotheses—something that could potentially contribute to scientific discovery.

From Wall Street to Medicine

At first glance, financial markets and medical diagnosis appear to have little in common.

But Dhar sees a common structure.

Both involve enormous quantities of information.

Both contain patterns that humans may struggle to recognize.

And both involve prediction under uncertainty.

In finance, Dhar discovered that he did not need a perfect model. He argued that a relatively small statistical edge could become powerful when consistently applied and scaled.

His experience reinforced another principle:

The machine does not necessarily need to understand everything to find something useful.

That idea becomes particularly interesting in medicine.

A physician may see one patient at a time. An AI system could potentially analyze patterns across huge numbers of medical records.

The challenge is turning those patterns into reliable evidence rather than merely plausible predictions.

AI Could Help Medicine With “Scorekeeping”

One of the most compelling sections of the interview concerns something much less glamorous than futuristic diagnostics: medical scorekeeping.

Dhar argues that medicine often lacks systematic feedback about what happens to patients after clinical decisions.

He gives the example of medical records and a hypothetical patient whose PSA levels remain elevated.

A physician may offer a clinical hypothesis, but the system does not necessarily maintain an accessible database showing what happened to thousands of patients with similar characteristics.

Dhar asks why a physician could not instead see comparable cases, their trajectories, and their outcomes.

His answer is that medicine has historically not done enough systematic “scorekeeping.”

AI, he argues, could help organize medical records so that this information becomes usable.

This is potentially one of the most practical applications of AI in healthcare.

Modern medicine produces enormous amounts of data, but data alone is not knowledge.

The information has to be organized, compared, evaluated, and connected to outcomes.

AI could potentially help with that process.

Current reviews of AI in healthcare similarly identify applications involving electronic health records, disease diagnosis, patient monitoring, medical imaging, and clinical decision support.

The Problem: AI Can Sound Right Without Being Right

But Dhar’s optimism about AI is accompanied by an important warning.

A system can generate an answer without necessarily possessing the kind of responsibility humans associate with knowledge.

This is especially important in medicine.

An AI model may produce a coherent explanation. That does not automatically mean the explanation is true.

The distinction between coherence and truth becomes critical when an AI system is used for medical decisions.

AI can identify correlations.

AI can rank possibilities.

AI can search enormous datasets.

AI can recognize patterns.

But these capabilities do not eliminate the need for validation, clinical judgment, accountability, and evidence.

Research reviews of AI in medical diagnosis continue to identify challenges involving reliability, interpretability, generalizability, ethics, privacy, and clinical adoption.

The World Health Organization likewise emphasizes that AI for health requires governance, safety, equity, and ethical practices.

The New AI Divide: Amplifier or Crutch?

Dhar believes AI may create an important distinction between people who use it as an amplifier and people who use it as a crutch.

The person who knows how to ask good questions, evaluate an answer, identify errors, and ask the next question can potentially become much more capable.

But someone who accepts every AI response uncritically may become dependent on the machine.

Dhar describes this as a potential divide between people who amplify themselves through AI and those who fail to evaluate what the machine produces.

This may be particularly consequential in healthcare.

A doctor using AI as a sophisticated analytical assistant is different from a doctor simply accepting whatever an AI system says.

The difference is judgment.

AI Is Moving Beyond Narrow Expertise

Dhar argues that one of the biggest changes in artificial intelligence is the weakening of the traditional boundary between specialized expertise and common sense.

For decades, AI systems were often designed around narrowly defined domains.

Medical diagnosis was one domain.

Engineering was another.

Tax planning was another.

These systems could be highly capable inside their boundaries.

Dhar believes newer AI systems increasingly blend information across domains, creating machines that can know something about many different subjects rather than being confined to one narrow area.

That development has enormous implications for medicine.

A medical AI system could potentially combine information from medical literature, patient records, laboratory data, imaging, genetics, and other sources.

But greater breadth also creates greater responsibility.

The more capable the system becomes, the more important it is to understand what it actually knows—and what it merely predicts.

“An Alien of Our Own Making”

Dhar uses a striking metaphor for modern AI: an alien of our own creation.

The phrase captures the strange position humans now occupy.

We built the machines.

We trained them on human-created information.

Yet increasingly sophisticated systems can produce combinations, associations, and answers that their creators did not explicitly program.

The result is not simply another tool.

It is a new kind of technological intelligence whose behavior requires careful evaluation.

Dhar’s description of AI as something that has learned broadly about the world appears in the context of his argument that the boundary between expertise and common sense is dissolving.

That is why the question is no longer simply:

“Can AI do this?”

The more important question may be:

“Where should AI stop?”

Can AI Really Sniff Out Disease?

The answer today should be nuanced.

There is legitimate scientific research investigating AI-assisted odor analysis and disease detection. Parkinson’s disease is one example, with researchers studying volatile organic compounds and artificial olfactory systems.

But Dhar’s interview describes a research direction, not a universally validated medical test.

The proposed idea is powerful:

Disease → biological change → altered signal → measurable pattern → machine-learning detection

If researchers can establish reliable, reproducible relationships between disease and biological olfactory signatures, the approach could eventually contribute to earlier and less invasive forms of diagnosis.

But moving from an intriguing research signal to a clinical diagnostic requires extensive validation.

That means larger datasets, independent testing, reproducibility, clinical trials, appropriate controls, and evidence that the technology works outside the original research environment.

The Bigger Lesson From Vasant Dhar

The most important lesson from Dhar’s work may not ultimately be about smell.

It is about patterns.

His career moved from medical expert systems to machine learning, from consumer data to Wall Street trading, and eventually back toward medicine.

Across those different fields, the same question keeps appearing:

What can machines see that humans cannot easily see?

Sometimes the answer is hidden in consumer behavior.

Sometimes it is hidden in financial data.

And perhaps someday, it could be hidden in biological signals produced when a patient encounters an odor.

That is the promise of AI in healthcare—not necessarily replacing physicians, but giving medicine new ways to observe, compare, measure, and learn.

The Future of AI in Healthcare

Artificial intelligence is already being explored across medical diagnosis, patient monitoring, medical imaging, drug development, and clinical decision support.

Dhar’s work adds another possibility: using biology itself as a sensor and AI as the pattern-recognition engine.

That is an intriguing reversal of the traditional relationship between humans and machines.

Instead of forcing biology into crude human descriptions, researchers can measure biological responses directly and allow machine learning to discover the patterns.

The promise is enormous.

But so is the responsibility.

AI can process more information than any individual physician.

It can uncover relationships that humans miss.

It can potentially help medicine learn from millions of cases rather than isolated encounters.

Yet it cannot automatically turn prediction into truth.

The future of medicine may therefore depend less on choosing between humans and machines and more on learning how to combine them intelligently.

As Dhar argues, thinking with machines may be inevitable.

The question is whether we will use that intelligence to amplify human judgment—or surrender our judgment to it.

Key Takeaways

AI in healthcare could help identify patterns across enormous medical datasets.
AI-assisted smell analysis is being investigated as a possible approach to disease detection.
Parkinson’s disease is one area where researchers have studied artificial olfactory systems and machine learning.
Vasant Dhar’s approach uses biological olfactory responses as a form of “ground truth” for machine learning.
Machine learning can identify patterns before humans understand the reasons behind them.
AI could potentially improve medical “scorekeeping” by organizing patient records and outcomes.
AI-generated answers can be coherent without necessarily being true.
Human judgment, validation, accountability, and evidence remain essential.
The future may belong to people who use AI as an amplifier rather than a substitute for thinking.

Frequently Asked Questions

Can AI detect disease through smell?

Researchers are investigating AI-based olfactory systems that analyze chemical or biological odor signatures to identify disease-related patterns. Parkinson’s disease is one area of active research. However, these approaches should not be confused with universally validated clinical diagnostic tests.

What is AI in healthcare?

AI in healthcare refers to the use of artificial intelligence and machine learning to analyze medical information and support tasks such as diagnosis, patient monitoring, clinical decision-making, and healthcare management.

How could smell help diagnose disease?

Disease may alter biological processes that affect chemical signals or odors. Researchers can attempt to measure those signals and use machine learning to determine whether patterns distinguish disease from healthy states.

What is Vasant Dhar known for?

Vasant Dhar is an NYU professor and longtime AI researcher whose career has included artificial intelligence, machine learning, data analysis, financial prediction, and research into digitizing smell for potential disease prediction. The interview describes nearly four decades of work across these areas.

Can AI replace doctors?

The interview does not establish that AI should replace doctors. Instead, Dhar emphasizes AI’s ability to analyze patterns, improve information processing, and potentially amplify human capabilities. The clinical use of AI still requires evidence, validation, oversight, and accountability.

What is the biggest challenge with AI medical diagnosis?

One major challenge is distinguishing a model’s ability to produce a plausible or coherent answer from its ability to produce a clinically reliable and truthful one. Current research also identifies concerns involving validation, interpretability, bias, privacy, and implementation.

Final Thought

The most fascinating possibility in Vasant Dhar’s research is not simply that a machine might someday “smell” Parkinson’s disease.

It is that biology may contain information that humans have never learned how to read properly—and artificial intelligence may give us a new way to read it.

If that happens, diagnosis could become less about waiting for obvious symptoms and more about recognizing subtle biological patterns early.

But the machine should remain a tool for discovering and testing those patterns, not an unquestioned authority.

The future of medicine may depend on a partnership between human judgment, biological signals, massive datasets, and machine intelligence.

That is the real promise—and the real challenge—of thinking with machines.