A patient misses an appointment.
In the medical record, the story may end there.
No-show. Appointment not completed. Follow-up recommended.
What the record may not show is everything that happened before that empty appointment slot.
Perhaps the patient spent an hour searching for a specialist who accepted their insurance. Maybe the first available appointment was two months away. Perhaps the office required a referral the patient did not know they needed. Maybe they couldn’t get away from work, couldn’t arrange childcare, couldn’t afford the deductible or simply became overwhelmed by a system filled with terms such as prior authorization, coinsurance, formulary, in-network and out-of-pocket maximum.
And perhaps, later that evening, they opened Reddit.
There, the story might sound very different.
Instead of a checkbox or missed-appointment code, there may be several paragraphs describing what they tried, where the process failed, what confused them, what it cost and how exhausted they became trying to get care.
That distinction is becoming increasingly important as researchers explore an unconventional source of healthcare information: the conversations people have online when they are not sitting in an examination room.
At the University at Albany, doctoral student Arissa Spitalny is examining more than one million posts from Reddit’s AskDocs community to better understand the barriers people encounter while navigating healthcare. Her current work focuses particularly on posts mentioning health insurance. With assistance from artificial intelligence, roughly 15,000 insurance-related posts have been classified for analysis. Before AI was used across that larger collection, Spitalny manually reviewed 300 posts to help establish how the material should be labeled. The study is now in its analysis and writing phase, so its final findings have not yet been published.
At first glance, this may sound like another example of researchers finding a new use for artificial intelligence.
But there is a much larger story here.
The research points toward an emerging role for AI in healthcare that has little to do with diagnosing disease, interpreting an X-ray or generating a clinical note.
AI may also help healthcare understand something it has historically struggled to capture:
what it actually feels like to navigate the system.
The medical record tells one version of the story
Healthcare generates enormous quantities of data.
There are diagnoses, medications, laboratory results, imaging studies, claims, appointments, clinical notes, procedure codes, insurance transactions and patient portal messages.
All of these sources tell us something.
But they do not necessarily tell us everything.
Consider a patient who stops taking a medication.
The record might eventually document:
Patient nonadherent with prescribed therapy.
But that description alone cannot tell us whether the medication cost $400 at the pharmacy, whether the patient’s insurer denied coverage, whether a prior authorization was never completed, whether the patient experienced side effects, whether transportation prevented them from getting to the pharmacy, or whether they simply did not understand why the medication was necessary.
The difference matters.
“Patient did not take medication” describes an outcome.
“I couldn’t afford the medication after my insurance stopped covering it” begins to explain the barrier behind that outcome.
Healthcare organizations are increasingly interested in social determinants of health and the nonclinical circumstances that affect outcomes. Yet many of these circumstances are difficult to measure consistently.
That is where online communities become interesting to researchers.
People often describe their experiences online in language that is less structured, more emotional and sometimes more detailed than what eventually appears in their health record.
They ask questions.
They complain.
They compare experiences.
They admit confusion.
They describe mistakes.
They explain what they were afraid to ask.
They recount phone calls with insurers, arguments over bills, unsuccessful attempts to find specialists, concerns about symptoms and uncertainty about whether something is serious enough to justify an emergency department visit.
None of this makes Reddit a substitute for clinical data.
It makes it a different kind of data.
Why people are talking about health outside the doctor’s office
Healthcare professionals remain an important and highly trusted source of medical information.
But they are no longer the only source people consult.
A 2026 Pew Research Center study of 5,111 U.S. adults found that 85% get health information from healthcare providers at least sometimes. Yet nearly as many people are participating in a much broader information ecosystem.
About 66% say they get health information from people dealing with health issues similar to their own, 36% get health information from social media at least sometimes, and 22% say they use AI chatbots for health information at least sometimes.
The appeal of peer-to-peer information is not difficult to understand.
A physician can explain what a colonoscopy involves.
Another patient may tell you what the preparation actually felt like.
A cardiologist can explain why a medication is needed.
Someone living with the same condition can explain how they manage taking it every morning while getting children ready for school.
An insurance representative can explain the technical definition of a deductible.
Someone who has just received a $2,700 bill can explain what discovering that deductible meant for their household.
These sources perform different functions.
And patients appear to recognize that distinction.
Pew found that Americans still rate healthcare providers considerably higher than social media for accuracy. Among people who obtain health information from healthcare providers, 65% described the information as extremely or very accurate. Among people obtaining health information through social media, only 7% gave it similarly high marks for accuracy, while 47% considered it not too accurate or not accurate at all.
So the rise of online health discussions should not automatically be interpreted as patients rejecting physicians.
Sometimes people are looking for something medicine does not traditionally provide:
the experience of someone who has already lived through the problem.
The healthcare system is difficult to navigate—even when someone has insurance
The University at Albany project is particularly relevant because Spitalny is initially examining posts related to health insurance.
Insurance is one of the places where the difference between having access to healthcare on paper and being able to obtain healthcare in practice becomes especially visible.
Recent polling illustrates the scale of the problem.
A February 2026 KFF survey found that nearly half of insured adults—47%—reported that access to a healthcare service, treatment or medication had been denied, delayed or altered by their insurer during the previous two years. Among insured people with chronic conditions, that figure rose to 57%.
Prior authorization emerged as a particularly significant frustration. About one-third of insured adults identified prior authorization as their single biggest burden among several challenges associated with navigating care.
Cost creates another layer of difficulty.
Updated KFF data published in 2026 found that 36% of U.S. adults reported skipping or postponing needed healthcare during the previous year because of cost. About 18% said their health became worse because they delayed or skipped care.
Those statistics are important.
But statistics tell us how often something happens.
They do not always tell us how it unfolds in someone’s life.
That is where thousands of individual narratives may add something traditional datasets cannot.
A survey can identify that patients experience prior-authorization problems.
Thousands of patient conversations might reveal where the confusion starts, which instructions are misunderstood, what happens after a denial, how many calls patients make, whether they abandon treatment and what kinds of help they seek when the formal system fails them.
For health services researchers, those details can be extremely valuable.
AI changes what researchers can realistically examine
There is a practical problem with using online conversations as research data.
There is far too much of it for one person to read.
That is precisely where artificial intelligence becomes important.
Imagine asking a researcher to manually review one million Reddit posts, determine which discuss insurance, classify the type of problem described and identify recurring themes.
It would be extraordinarily labor-intensive.
AI can dramatically reduce that burden by helping researchers sort, categorize and organize large volumes of text.
That does not mean the AI “understands patients” in the human sense.
It means AI can perform a task at a scale that would previously have been impractical.
In Spitalny’s work, human involvement came first. She manually reviewed hundreds of posts, and those examples were then used to guide the AI classification process. The technology helped make it feasible to review the much larger dataset.
This human-plus-machine model may ultimately prove more important than either one alone.
The machine brings scale.
The researcher brings context.
AI can help identify that thousands of comments contain language associated with insurance denials.
A human researcher must still ask:
Why?
Are people misunderstanding coverage?
Are provider directories inaccurate?
Are medications being denied?
Are specialists unavailable?
Are prior authorizations causing delays?
Are people abandoning care?
Are certain populations describing different barriers?
Those are not merely classification problems.
They are healthcare questions.
This is AI being used to listen, not diagnose
Much of the public conversation surrounding artificial intelligence in medicine concentrates on high-profile clinical applications.
Can an algorithm identify cancer on an image?
Can AI predict sepsis?
Can a chatbot answer a medical question?
Can an ambient documentation system write a physician’s note?
Can generative AI assist with diagnosis?
Those developments deserve attention.
But they can create an unnecessarily narrow picture of what healthcare AI may eventually become.
Some of the most useful applications may be far less dramatic.
AI can analyze language.
It can identify patterns across large collections of documents.
It can organize previously unstructured information.
It can help researchers see patterns that would be difficult to detect manually.
Applied responsibly, that means AI could potentially help healthcare organizations examine not only what clinicians document about patients but what patients say about healthcare.
That concept is worth paying attention to.
Healthcare has historically been very good at collecting information after someone enters the system.
It has been much less successful at understanding the experiences of the person struggling to enter it.
The equity implications may be significant
There is another reason this kind of research matters.
Not everyone interacts with the healthcare system in the same way, and not everyone uses the same sources when seeking health information.
A Pew analysis published in August 2026 found substantial differences in the use of social media for health information across racial and ethnic groups.
Among U.S. adults, the percentage who reported getting health information from social media at least sometimes was:
- 49% among Hispanic adults
- 47% among Asian adults
- 45% among Black adults
- 30% among White adults
Pew notes that its estimates for Asian adults are representative of English-speaking Asian adults.
That does not mean social media provides a representative sample of these populations. It does not.
But it does reinforce why healthcare researchers should take seriously the fact that some conversations about health are taking place outside traditional healthcare institutions.
If researchers can responsibly identify recurring barriers in those conversations, the information could help point toward problems in care navigation, communication, insurance literacy and access.
It might show, for example, that patients repeatedly misunderstand the same insurance language.
Or that certain administrative processes generate disproportionate confusion.
Or that patients repeatedly seek online help because they do not know which healthcare professional or organization to contact.
Those findings could eventually inform better patient education, navigation programs and communication strategies.
The goal should not be to replace the patient conversation.
It should be to notice where the healthcare system repeatedly fails to create one.
But Reddit is not a waiting room filled with representative patients
There is an important limitation that should not be overlooked.
One million Reddit posts may constitute an enormous dataset.
That does not make it a representative dataset.
Reddit users differ from the general population. People who choose to post in an online health community differ from people who remain silent. Someone experiencing an especially frustrating insurance problem may be more motivated to post than someone whose care went smoothly.
Some posters may be patients.
Others may be relatives, caregivers or people asking hypothetical questions.
Demographic details may be absent, incomplete, self-reported or difficult to verify.
The loudest problem online is not necessarily the most common problem offline.
This distinction becomes particularly important when AI enters the process.
A model can efficiently detect patterns within the information it receives.
It cannot make an unrepresentative dataset representative.
If certain communities are missing from the source data, AI cannot magically restore their experiences.
If people who have limited internet access are underrepresented, their absence matters.
If the way people write online causes an algorithm to misunderstand sarcasm, cultural language, slang or context, that matters too.
Large datasets can create an illusion of certainty.
Healthcare researchers still have to ask who is represented, who is missing and whether the conclusions being drawn actually match the evidence.
There is also a privacy question healthcare cannot ignore
Online health conversations create another complicated issue.
A Reddit post may be publicly accessible.
But the person who wrote it may not have imagined that it would eventually become part of a healthcare research dataset analyzed by artificial intelligence.
Those two things can both be true.
Researchers have been debating this issue for years.
Reviews of social-media research ethics have identified concerns involving privacy, informed consent, anonymity, confidentiality, authenticity, context and potential harm. Researchers have also warned that simply because information is publicly accessible does not mean users necessarily understand every future way their information could be collected and analyzed.
Health information makes those questions particularly sensitive.
Someone might anonymously describe infertility, cancer, addiction, a psychiatric condition, sexual health concerns, financial hardship or a frightening medical symptom because they feel protected by a username.
That post may be public in a technical sense.
It can still feel deeply private to the person who wrote it.
Responsible research therefore requires much more than stripping away usernames.
Researchers must think carefully about whether quotations could be searchable, whether individuals could be reidentified, how data are stored, how findings are reported and whether the potential benefit of the research justifies the use of the information.
AI makes these questions more urgent because it dramatically increases the amount of information researchers can analyze.
The technical ability to analyze something is not, by itself, an ethical justification for doing so.
AI can find a pattern. Humans still have to understand it.
There is another limitation that deserves equal attention.
Language is messy.
A patient may write:
“My insurance company is killing me.”
A human reader understands that this might be frustration rather than a literal allegation of attempted homicide.
Someone might say a physician “never listens,” even though they have seen that physician only once.
A person might misunderstand why a claim was denied.
Another poster might confidently provide completely incorrect information.
A single Reddit thread could contain useful lived experience, misinformation, sarcasm, anger, medical advice, jokes and speculation—all within a few dozen comments.
AI classification systems can make mistakes in those environments.
That is one reason human oversight remains essential.
AI may be extraordinarily useful for narrowing a million posts into thousands of potentially relevant conversations.
But identifying a pattern is not the same as establishing its meaning.
Researchers must still validate the findings, consider alternative explanations and resist the temptation to assign clinical or policy significance simply because a computational model detected a recurring theme.
There is an intriguing lesson here for healthcare organizations
This research also raises a question that extends beyond academia.
Healthcare organizations already collect enormous amounts of feedback.
Patient satisfaction surveys.
Complaints.
Call-center recordings.
Portal messages.
Appeals.
Grievances.
Appointment requests.
Billing questions.
Denial information.
Patient navigation notes.
Many organizations hold these sources in separate systems, managed by different departments.
Individually, they may look like operational noise.
Collectively, they could contain patterns.
Imagine discovering that hundreds of patients are asking essentially the same question about a discharge instruction.
Or that patients repeatedly misunderstand one portion of an explanation of benefits.
Or that people referred to a particular service consistently describe difficulty getting an appointment.
Or that certain types of authorization denials repeatedly result in treatment delays.
This is where the broader concept behind the Albany research becomes especially interesting.
AI’s value may not simply come from generating something new.
Sometimes its value may come from helping humans see what was already there.
The rise of a new kind of healthcare data
For decades, healthcare data largely meant information created inside healthcare institutions.
The electronic health record expanded that dramatically.
Claims databases added another view.
Wearables introduced continuous physiological information.
Patient portals added direct patient communication.
Now social platforms, online communities, search behavior and AI conversations are producing another enormous layer of health-related information.
We should be careful about calling all of it “patient data.”
And healthcare organizations should certainly not assume they have a right to collect or use every digital trace a person creates.
But researchers can no longer ignore the fact that meaningful conversations about illness, access, affordability and healthcare experiences increasingly occur outside the traditional healthcare environment.
Pew’s findings make that particularly clear.
Half of Americans report at least some difficulty determining whether health information they encounter is accurate, and 76% say they come across conflicting health information at least sometimes.
People are not simply receiving information anymore.
They are moving between doctors, websites, friends, social communities and AI systems—comparing one source against another as they try to make decisions.
Understanding that behavior may become an important part of understanding modern healthcare itself.
Perhaps the most important information is sometimes outside the chart
A medical record may tell us that a patient did not obtain a medication.
It may not tell us about the three hours spent on the telephone trying to get it covered.
A claims database may show that a recommended procedure never occurred.
It may not reveal that the patient called four specialists and none accepted the insurance plan listed on the insurer’s website.
An appointment system may record a no-show.
It may never capture the parent who could not afford to lose another day’s wages.
And a diagnosis code may accurately describe someone’s disease while telling us almost nothing about what living with that disease actually requires.
This does not diminish the value of clinical documentation or traditional healthcare data.
It demonstrates their boundaries.
The University at Albany project is still underway, and it will be important to see what patterns ultimately emerge from the research. But the study is already pointing toward something larger than Reddit.
Artificial intelligence is making it possible to examine enormous collections of human language in ways that were previously impractical.
Healthcare now has a choice about what it does with that ability.
It can use AI primarily to automate the information it already collects.
Or it can also use these tools, carefully and ethically, to ask what its existing information systems have been failing to capture.
Because sometimes the most revealing part of a patient’s healthcare story is not buried somewhere inside the medical record.
It is the part that never made it into the record at all.
Researchers are analyzing more than one million Reddit posts to uncover barriers to healthcare. Their work raises a larger question: What could healthcare learn by paying attention to the conversations that never make it into the medical record?
A patient misses an appointment.
In the medical record, the story may end there.
No-show. Appointment not completed. Follow-up recommended.
What the record may not show is everything that happened before that empty appointment slot.
Perhaps the patient spent an hour searching for a specialist who accepted their insurance. Maybe the first available appointment was two months away. Perhaps the office required a referral the patient did not know they needed. Maybe they couldn’t get away from work, couldn’t arrange childcare, couldn’t afford the deductible or simply became overwhelmed by a system filled with terms such as prior authorization, coinsurance, formulary, in-network and out-of-pocket maximum.
And perhaps, later that evening, they opened Reddit.
There, the story might sound very different.
Instead of a checkbox or missed-appointment code, there may be several paragraphs describing what they tried, where the process failed, what confused them, what it cost and how exhausted they became trying to get care.
That distinction is becoming increasingly important as researchers explore an unconventional source of healthcare information: the conversations people have online when they are not sitting in an examination room.
At the University at Albany, doctoral student Arissa Spitalny is examining more than one million posts from Reddit’s AskDocs community to better understand the barriers people encounter while navigating healthcare. Her current work focuses particularly on posts mentioning health insurance. With assistance from artificial intelligence, roughly 15,000 insurance-related posts have been classified for analysis. Before AI was used across that larger collection, Spitalny manually reviewed 300 posts to help establish how the material should be labeled. The study is now in its analysis and writing phase, so its final findings have not yet been published.
At first glance, this may sound like another example of researchers finding a new use for artificial intelligence.
But there is a much larger story here.
The research points toward an emerging role for AI in healthcare that has little to do with diagnosing disease, interpreting an X-ray or generating a clinical note.
AI may also help healthcare understand something it has historically struggled to capture:
what it actually feels like to navigate the system.
The medical record tells one version of the story
Healthcare generates enormous quantities of data.
There are diagnoses, medications, laboratory results, imaging studies, claims, appointments, clinical notes, procedure codes, insurance transactions and patient portal messages.
All of these sources tell us something.
But they do not necessarily tell us everything.
Consider a patient who stops taking a medication.
The record might eventually document:
Patient nonadherent with prescribed therapy.
But that description alone cannot tell us whether the medication cost $400 at the pharmacy, whether the patient’s insurer denied coverage, whether a prior authorization was never completed, whether the patient experienced side effects, whether transportation prevented them from getting to the pharmacy, or whether they simply did not understand why the medication was necessary.
The difference matters.
“Patient did not take medication” describes an outcome.
“I couldn’t afford the medication after my insurance stopped covering it” begins to explain the barrier behind that outcome.
Healthcare organizations are increasingly interested in social determinants of health and the nonclinical circumstances that affect outcomes. Yet many of these circumstances are difficult to measure consistently.
That is where online communities become interesting to researchers.
People often describe their experiences online in language that is less structured, more emotional and sometimes more detailed than what eventually appears in their health record.
They ask questions.
They complain.
They compare experiences.
They admit confusion.
They describe mistakes.
They explain what they were afraid to ask.
They recount phone calls with insurers, arguments over bills, unsuccessful attempts to find specialists, concerns about symptoms and uncertainty about whether something is serious enough to justify an emergency department visit.
None of this makes Reddit a substitute for clinical data.
It makes it a different kind of data.
Why people are talking about health outside the doctor’s office
Healthcare professionals remain an important and highly trusted source of medical information.
But they are no longer the only source people consult.
A 2026 Pew Research Center study of 5,111 U.S. adults found that 85% get health information from healthcare providers at least sometimes. Yet nearly as many people are participating in a much broader information ecosystem.
About 66% say they get health information from people dealing with health issues similar to their own, 36% get health information from social media at least sometimes, and 22% say they use AI chatbots for health information at least sometimes.
The appeal of peer-to-peer information is not difficult to understand.
A physician can explain what a colonoscopy involves.
Another patient may tell you what the preparation actually felt like.
A cardiologist can explain why a medication is needed.
Someone living with the same condition can explain how they manage taking it every morning while getting children ready for school.
An insurance representative can explain the technical definition of a deductible.
Someone who has just received a $2,700 bill can explain what discovering that deductible meant for their household.
These sources perform different functions.
And patients appear to recognize that distinction.
Pew found that Americans still rate healthcare providers considerably higher than social media for accuracy. Among people who obtain health information from healthcare providers, 65% described the information as extremely or very accurate. Among people obtaining health information through social media, only 7% gave it similarly high marks for accuracy, while 47% considered it not too accurate or not accurate at all.
So the rise of online health discussions should not automatically be interpreted as patients rejecting physicians.
Sometimes people are looking for something medicine does not traditionally provide:
the experience of someone who has already lived through the problem.
The healthcare system is difficult to navigate—even when someone has insurance
The University at Albany project is particularly relevant because Spitalny is initially examining posts related to health insurance.
Insurance is one of the places where the difference between having access to healthcare on paper and being able to obtain healthcare in practice becomes especially visible.
Recent polling illustrates the scale of the problem.
A February 2026 KFF survey found that nearly half of insured adults—47%—reported that access to a healthcare service, treatment or medication had been denied, delayed or altered by their insurer during the previous two years. Among insured people with chronic conditions, that figure rose to 57%.
Prior authorization emerged as a particularly significant frustration. About one-third of insured adults identified prior authorization as their single biggest burden among several challenges associated with navigating care.
Cost creates another layer of difficulty.
Updated KFF data published in 2026 found that 36% of U.S. adults reported skipping or postponing needed healthcare during the previous year because of cost. About 18% said their health became worse because they delayed or skipped care.
Those statistics are important.
But statistics tell us how often something happens.
They do not always tell us how it unfolds in someone’s life.
That is where thousands of individual narratives may add something traditional datasets cannot.
A survey can identify that patients experience prior-authorization problems.
Thousands of patient conversations might reveal where the confusion starts, which instructions are misunderstood, what happens after a denial, how many calls patients make, whether they abandon treatment and what kinds of help they seek when the formal system fails them.
For health services researchers, those details can be extremely valuable.
AI changes what researchers can realistically examine
There is a practical problem with using online conversations as research data.
There is far too much of it for one person to read.
That is precisely where artificial intelligence becomes important.
Imagine asking a researcher to manually review one million Reddit posts, determine which discuss insurance, classify the type of problem described and identify recurring themes.
It would be extraordinarily labor-intensive.
AI can dramatically reduce that burden by helping researchers sort, categorize and organize large volumes of text.
That does not mean the AI “understands patients” in the human sense.
It means AI can perform a task at a scale that would previously have been impractical.
In Spitalny’s work, human involvement came first. She manually reviewed hundreds of posts, and those examples were then used to guide the AI classification process. The technology helped make it feasible to review the much larger dataset.
This human-plus-machine model may ultimately prove more important than either one alone.
The machine brings scale.
The researcher brings context.
AI can help identify that thousands of comments contain language associated with insurance denials.
A human researcher must still ask:
Why?
Are people misunderstanding coverage?
Are provider directories inaccurate?
Are medications being denied?
Are specialists unavailable?
Are prior authorizations causing delays?
Are people abandoning care?
Are certain populations describing different barriers?
Those are not merely classification problems.
They are healthcare questions.
This is AI being used to listen, not diagnose
Much of the public conversation surrounding artificial intelligence in medicine concentrates on high-profile clinical applications.
Can an algorithm identify cancer on an image?
Can AI predict sepsis?
Can a chatbot answer a medical question?
Can an ambient documentation system write a physician’s note?
Can generative AI assist with diagnosis?
Those developments deserve attention.
But they can create an unnecessarily narrow picture of what healthcare AI may eventually become.
Some of the most useful applications may be far less dramatic.
AI can analyze language.
It can identify patterns across large collections of documents.
It can organize previously unstructured information.
It can help researchers see patterns that would be difficult to detect manually.
Applied responsibly, that means AI could potentially help healthcare organizations examine not only what clinicians document about patients but what patients say about healthcare.
That concept is worth paying attention to.
Healthcare has historically been very good at collecting information after someone enters the system.
It has been much less successful at understanding the experiences of the person struggling to enter it.
The equity implications may be significant
There is another reason this kind of research matters.
Not everyone interacts with the healthcare system in the same way, and not everyone uses the same sources when seeking health information.
A Pew analysis published in August 2026 found substantial differences in the use of social media for health information across racial and ethnic groups.
Among U.S. adults, the percentage who reported getting health information from social media at least sometimes was:
- 49% among Hispanic adults
- 47% among Asian adults
- 45% among Black adults
- 30% among White adults
Pew notes that its estimates for Asian adults are representative of English-speaking Asian adults.
That does not mean social media provides a representative sample of these populations. It does not.
But it does reinforce why healthcare researchers should take seriously the fact that some conversations about health are taking place outside traditional healthcare institutions.
If researchers can responsibly identify recurring barriers in those conversations, the information could help point toward problems in care navigation, communication, insurance literacy and access.
It might show, for example, that patients repeatedly misunderstand the same insurance language.
Or that certain administrative processes generate disproportionate confusion.
Or that patients repeatedly seek online help because they do not know which healthcare professional or organization to contact.
Those findings could eventually inform better patient education, navigation programs and communication strategies.
The goal should not be to replace the patient conversation.
It should be to notice where the healthcare system repeatedly fails to create one.
But Reddit is not a waiting room filled with representative patients
There is an important limitation that should not be overlooked.
One million Reddit posts may constitute an enormous dataset.
That does not make it a representative dataset.
Reddit users differ from the general population. People who choose to post in an online health community differ from people who remain silent. Someone experiencing an especially frustrating insurance problem may be more motivated to post than someone whose care went smoothly.
Some posters may be patients.
Others may be relatives, caregivers or people asking hypothetical questions.
Demographic details may be absent, incomplete, self-reported or difficult to verify.
The loudest problem online is not necessarily the most common problem offline.
This distinction becomes particularly important when AI enters the process.
A model can efficiently detect patterns within the information it receives.
It cannot make an unrepresentative dataset representative.
If certain communities are missing from the source data, AI cannot magically restore their experiences.
If people who have limited internet access are underrepresented, their absence matters.
If the way people write online causes an algorithm to misunderstand sarcasm, cultural language, slang or context, that matters too.
Large datasets can create an illusion of certainty.
Healthcare researchers still have to ask who is represented, who is missing and whether the conclusions being drawn actually match the evidence.
There is also a privacy question healthcare cannot ignore
Online health conversations create another complicated issue.
A Reddit post may be publicly accessible.
But the person who wrote it may not have imagined that it would eventually become part of a healthcare research dataset analyzed by artificial intelligence.
Those two things can both be true.
Researchers have been debating this issue for years.
Reviews of social-media research ethics have identified concerns involving privacy, informed consent, anonymity, confidentiality, authenticity, context and potential harm. Researchers have also warned that simply because information is publicly accessible does not mean users necessarily understand every future way their information could be collected and analyzed.
Health information makes those questions particularly sensitive.
Someone might anonymously describe infertility, cancer, addiction, a psychiatric condition, sexual health concerns, financial hardship or a frightening medical symptom because they feel protected by a username.
That post may be public in a technical sense.
It can still feel deeply private to the person who wrote it.
Responsible research therefore requires much more than stripping away usernames.
Researchers must think carefully about whether quotations could be searchable, whether individuals could be reidentified, how data are stored, how findings are reported and whether the potential benefit of the research justifies the use of the information.
AI makes these questions more urgent because it dramatically increases the amount of information researchers can analyze.
The technical ability to analyze something is not, by itself, an ethical justification for doing so.
AI can find a pattern. Humans still have to understand it.
There is another limitation that deserves equal attention.
Language is messy.
A patient may write:
“My insurance company is killing me.”
A human reader understands that this might be frustration rather than a literal allegation of attempted homicide.
Someone might say a physician “never listens,” even though they have seen that physician only once.
A person might misunderstand why a claim was denied.
Another poster might confidently provide completely incorrect information.
A single Reddit thread could contain useful lived experience, misinformation, sarcasm, anger, medical advice, jokes and speculation—all within a few dozen comments.
AI classification systems can make mistakes in those environments.
That is one reason human oversight remains essential.
AI may be extraordinarily useful for narrowing a million posts into thousands of potentially relevant conversations.
But identifying a pattern is not the same as establishing its meaning.
Researchers must still validate the findings, consider alternative explanations and resist the temptation to assign clinical or policy significance simply because a computational model detected a recurring theme.
There is an intriguing lesson here for healthcare organizations
This research also raises a question that extends beyond academia.
Healthcare organizations already collect enormous amounts of feedback.
Patient satisfaction surveys.
Complaints.
Call-center recordings.
Portal messages.
Appeals.
Grievances.
Appointment requests.
Billing questions.
Denial information.
Patient navigation notes.
Many organizations hold these sources in separate systems, managed by different departments.
Individually, they may look like operational noise.
Collectively, they could contain patterns.
Imagine discovering that hundreds of patients are asking essentially the same question about a discharge instruction.
Or that patients repeatedly misunderstand one portion of an explanation of benefits.
Or that people referred to a particular service consistently describe difficulty getting an appointment.
Or that certain types of authorization denials repeatedly result in treatment delays.
This is where the broader concept behind the Albany research becomes especially interesting.
AI’s value may not simply come from generating something new.
Sometimes its value may come from helping humans see what was already there.
The rise of a new kind of healthcare data
For decades, healthcare data largely meant information created inside healthcare institutions.
The electronic health record expanded that dramatically.
Claims databases added another view.
Wearables introduced continuous physiological information.
Patient portals added direct patient communication.
Now social platforms, online communities, search behavior and AI conversations are producing another enormous layer of health-related information.
We should be careful about calling all of it “patient data.”
And healthcare organizations should certainly not assume they have a right to collect or use every digital trace a person creates.
But researchers can no longer ignore the fact that meaningful conversations about illness, access, affordability and healthcare experiences increasingly occur outside the traditional healthcare environment.
Pew’s findings make that particularly clear.
Half of Americans report at least some difficulty determining whether health information they encounter is accurate, and 76% say they come across conflicting health information at least sometimes.
People are not simply receiving information anymore.
They are moving between doctors, websites, friends, social communities and AI systems—comparing one source against another as they try to make decisions.
Understanding that behavior may become an important part of understanding modern healthcare itself.
Perhaps the most important information is sometimes outside the chart
A medical record may tell us that a patient did not obtain a medication.
It may not tell us about the three hours spent on the telephone trying to get it covered.
A claims database may show that a recommended procedure never occurred.
It may not reveal that the patient called four specialists and none accepted the insurance plan listed on the insurer’s website.
An appointment system may record a no-show.
It may never capture the parent who could not afford to lose another day’s wages.
And a diagnosis code may accurately describe someone’s disease while telling us almost nothing about what living with that disease actually requires.
This does not diminish the value of clinical documentation or traditional healthcare data.
It demonstrates their boundaries.
The University at Albany project is still underway, and it will be important to see what patterns ultimately emerge from the research. But the study is already pointing toward something larger than Reddit.
Artificial intelligence is making it possible to examine enormous collections of human language in ways that were previously impractical.
Healthcare now has a choice about what it does with that ability.
It can use AI primarily to automate the information it already collects.
Or it can also use these tools, carefully and ethically, to ask what its existing information systems have been failing to capture.
Because sometimes the most revealing part of a patient’s healthcare story is not buried somewhere inside the medical record.
It is the part that never made it into the record at all.
Discover more from Zed Health Intelligence
Subscribe to get the latest posts sent to your email.
