In the day’s roundup, among reports of sanctions and missile launches, a mundane little item flashed by, easy to scroll past: people search for their symptoms online. Nothing new—almost everyone does it. But behind that gesture stands a machine that answers a different question from the one it is asked, and does so in good faith.
The scenario is familiar to anyone: you wake up with a headache that has not gone away for a second day, and instead of calling a doctor, you type a line into a search bar. Twenty minutes later, you are reading about a brain tumour, though you only came to find out whether it might be lack of sleep. The usual explanation—“the internet scares people”—describes the symptom and says nothing about the mechanism. I wanted to understand exactly where the malfunction happens: whether it is a lie, an algorithm, or the question itself.
The answer turned out to be more interesting than the horror stories. Search does not lie about the connection between headaches and tumours—it honestly finds it in the texts. The problem is that “what could this be?” and “what has been written about this?” are two different questions, and they have different denominators.
When a person types the name of a symptom into a search engine, they want to know the probability: what they most likely have. A search engine does something else—it looks for matches. It cannot calculate what proportion of people with this symptom actually have a rare disease; it can show pages where those words appear together. Those are not the same thing, and the difference has been measured quantitatively—with numbers, not impressions.
The clearest example is headache and brain tumour. In a random sample of forty million web pages, the two topics appear together in roughly 3% of cases. In the top search results for “headache,” it is already 26%. The actual incidence of brain tumours in the population is around one case per ten thousand people a year, about 0.0116%. In other words, search puts the odds at more than two thousand times their real level.
The same pattern recurs with other symptoms. Muscle twitching on ordinary web pages appears alongside the term “ALS” (amyotrophic lateral sclerosis) in 7% of cases; in the top search results, it is already 50%, even though ALS incidence is roughly one case per 55 thousand people a year. Chest pain and “heart attack”: 15% on the web versus 37% in search results, even though the most common cause of such pain is heartburn. Curiously, specialised medical search produces almost none of these distortions: the data there are checked, and it honestly puts the boring answer first.
The key study of the subject was conducted in 2008–2009 by two Microsoft researchers, Ryen White and Eric Horvitz. They had what nobody else did: anonymised logs from hundreds of thousands of volunteers who had agreed to have their web activity recorded, spanning eleven months. Medical queries made up about 2% of all searches; roughly a quarter of users looked up something health-related at least once during that time.
Separately, they sent a survey to five thousand Microsoft employees and received 515 responses. The picture was not one of hypochondriacs. Only 3.5% of respondents considered themselves hypochondriacs, and the average health-anxiety score was 2.78 out of 10—people were generally calm. But nine out of ten recalled at least one occasion when a search for an ordinary symptom had led them to texts about serious illnesses, and one in five said it happened often. Around 40% admitted they had worried about a serious disease they did not have.
Then comes the most valuable part. The logs show that escalation from a harmless symptom to a frightening diagnosis is not all that common: of more than eleven thousand sessions involving any concern at all, only around 5% arrived at a “scary” formulation. But the researchers took the trouble to manually annotate a sample of sessions their automated system had classified as “unchanged,” and found that almost a third of them actually contained escalation—simply expressed in words that had not made it into the dictionary. With that correction, the share of sessions drifting toward serious illnesses rose to around a third. The automated count substantially understated reality—and the study leaves that seam honestly visible.
And one more detail explains why it all happens like an attack. Nearly 80% of searches about a specific symptom are made within two weeks of the first one: anxiety comes in bursts. A session that moves into escalation is, on average, noticeably longer: it takes more time, more queries, and more pages viewed to get from a symptom to a frightening diagnosis than to reach a harmless explanation. The difference between the two kinds of people is stable, too: if escalation happens to someone once, it happens again, while the overlap between the “frightened” and “unfrightened” groups is just twenty-one people out of seven hundred.
The word “cyberchondria” itself is “cyber” plus “hypochondria,” and it was coined by journalists. Vlado Starcevic, the psychiatrist behind one of the major overviews of the subject, traced its history: the leading contenders for the first use are a 1999 note in the Wall Street Journal and a 2001 article in Britain’s The Independent, though the exact author remains impossible to establish. The term was born in the years when newspapers eagerly wrote about “internet addiction” and other new afflictions, which is precisely why doctors did not take it seriously for about a decade.
Starcevic asks an uncomfortable question: is this even a separate illness? Formally, cyberchondria is strongly linked to health anxiety—a meta-analysis (seven studies and 3069 people entered its quantitative section, while the review as a whole covered twenty-five papers) gives a correlation of 0.63, solid but not absolute. It also has its own markers: links to obsessive symptoms, intolerance of uncertainty, and a tendency to interrupt other activities to search. But whether that adds up to a separate diagnosis or merely describes how an anxious person seeks reassurance remains an open question. Starcevic’s answer is blunt: it is still unclear whether cyberchondria can be regarded as an independent entity.
His model explains the mechanism in the language of psychology: symptom searching works here as “safety behaviour.” A person looks for confirmation that everything is fine, gets it for a second, then starts searching again a minute later because the reassurance does not work. This behaviour is sustained by intermittent reinforcement—exactly like a slot machine that sometimes pays out. The first tool for measuring cyberchondria, the CSS scale, appeared only in 2014; before that, there was simply no proper way to measure it.
There is a side to this story that the horror stories do not mention. In 2005, doctors Hangwi Tang and Jennifer Ng took twenty-six difficult diagnostic cases from the New England Journal of Medicine and tried to guess the diagnoses using ordinary Google, without peeking at the answers. They got fifteen out of twenty-six right—58%. The study is still cited by both sides of the argument: as proof that the internet helps with diagnosis, and as proof that it is dangerous. The figure honestly supports both readings—for a doctor who knows how to filter out noise, 58% is useful assistance; for an untrained person, the same 58% means they will be wrong in four cases out of ten.
The behaviour of the search industry itself was revealing, too. In June 2016, Google officially acknowledged the problem: a post on the company blog said that medical content online “tends to lead people from mild symptoms to scary and unlikely conditions,” causing unnecessary anxiety. The company redesigned symptom search results—now, for “headache,” the first thing you see is a list of likely, often boring causes: migraine, tension headache, sinusitis, common cold. The lists were compiled with the help of doctors and checked against the Mayo Clinic reference guide. Google also named the scale: around 1% of all search queries concern symptoms, which means millions a day.
Whether the treatment works has been tested as well. British researchers led by Peter Tyrer recruited 444 patients from five hospitals—people whose health anxiety scored off the charts on the questionnaire—and gave half of them a course of cognitive behavioural therapy specifically designed to teach them not to check symptoms. A year later, their severe anxiety had eased to moderate levels, and the effect held even five years on. The sessions were run by ordinary nurses, who did no worse than psychologists. One separate finding: patients in the control group died earlier, meaning that giving up compulsive checking did not lead to real illnesses being missed. Yet only one in ten sufferers receives the diagnosis—the rest are treated for anything but anxiety.
The money in this story is counted surprisingly crudely. A press release from Imperial College accompanying Tyrer’s study put the cost of unnecessary visits and tests by such patients at “up to £56 million” a year for the NHS. The same story, as retold by The Telegraph and then Wikipedia, turns into £420 million—more than seven times as much, now with the caveat “outpatient appointments only.”
The difference is most likely one of definition. One figure counts only explicit cases; the other counts every visit that can be linked to internet anxiety, and there are many ways to draw that boundary. But what is revealing is something else: nobody has managed to agree on what it really costs. Health anxiety does not fit neatly into accounting: it has no billing code, only repeat appointments and unnecessary scans, each of which looks justified on its own.
The first wave of cyberchondria was associated with search bars—at least with them, it was clear that a person received a list of links, not a verdict. The new version sounds different. In the description of the problem found in recent reviews, conversational chatbots based on large language models appear as a separate item: a polite, personal, detailed answer in which the assistant agrees with your theory and adapts itself to it. The same AI called upon to solve doctor shortages is known for the opposite trait, too—being far too good at confirming what a person already believes.
For now, this is more observation than measured effect: the review covering 87 studies of quality of life in cyberchondria could include only cross-sectional work, which says nothing about causes. But the shift is obvious. Search at least showed ten possibilities, nine of them boring—a conversational bot shows one, but it is confident and written specifically for the person it is talking to.
The malfunction so much has been written about is not that the internet stores frightening information. It is in the arithmetic that cannot be done alone. The denominator of that probability is how many thousands of people with the same headache turned out to be healthy, and a person physically cannot look into those statistics because they can see only their own head. Search answers their question with a number from an entirely different problem, and there is almost nothing to be done about it—except not ask it.