Where a person crosses from distressed into ill is a threshold society sets — a value-laden decision dressed up as a medical discovery.
In December 1973 the board of the American Psychiatric Association voted to remove homosexuality as a general category of mental illness, and the change did not follow any new finding about the body. No laboratory or brain study had turned up a marker of illness in same-sex orientation. What changed was the association’s judgment about what should count as a disorder, together with several years of pressure from activists and from psychiatrists who found the old category indefensible. A faction that wanted the diagnosis kept forced the question to the full membership, and in 1974 the members ratified the removal by 5,854 votes to 3,810 in a mailed ballot.
That the boundary of an illness could be redrawn by a vote is uncomfortable partly because it clashes with how most people picture diagnosis. We imagine it as a discovery: a clinician examines a person, finds something that was already there, and names it, the way a radiologist finds a fracture on an X-ray. The fracture is present or absent regardless of what anyone votes, and a second radiologist looking at the same film will usually agree. Much of medicine really does work this way, since a blood culture grows a bacterium or it does not, and a tumor has a margin you can point to. When the same picture is carried into psychiatry, we assume that “depressed” or “anxious” or “disordered” names a fact about a person that the diagnosis merely uncovers.
Most psychiatric symptoms, though, come in degrees rather than in the two clean states the fracture model needs. Low mood, worry, difficulty concentrating, and loss of interest each run smoothly across a population, from none through mild and moderate to severe, with people spread all along the range and no gap in the middle that separates a healthy group from an ill one. When researchers have looked directly for such a gap in depression, using statistical methods designed to detect a hidden boundary between two natural classes, they have generally not found one, and the data look like a single continuous dimension. Suffering is real and often severe, but it is graded, and nature does not mark the point on the scale where ordinary unhappiness becomes major depressive disorder.
The diagnostic rules inherit that problem. The standard definition of major depression asks for at least five symptoms out of nine, present for at least two weeks. Someone with four symptoms for thirteen days does not meet this particular threshold, while a neighbor with five for fourteen days does, and when researchers have tested nearby cutoffs, requiring four symptoms or three weeks, the alternatives capture people just as impaired as the ones the official rule selects. The five-symptom, two-week rule organizes the continuum in a workable way, but it lands at a point the underlying data do not single out.
Since the distribution is continuous and a clinician still has to decide whether to treat, someone has to choose a cut-point, and that choice is where the value judgment enters. Consider the PHQ-9, a nine-item questionnaire used across primary care to screen for depression. Its score runs from 0 to 27, and a common convention treats a score of 10 or above as a probable case worth acting on. In the study that validated it, that cut-off caught about 88 percent of the people a diagnostic interview would call depressed while correctly clearing about 88 percent of those it would not. A later meta-analysis found that any value from roughly 8 to 11 performs about as well, so the choice of 10 reflects a tradeoff rather than a fact about the scale. Lowering the line catches more genuine cases but also flags more people who would have recovered on their own; raising it does the reverse. Symptom scores cannot say whether it is worse to over-treat the mild or to overlook the serious, because that choice depends on what a health system owes people and what it can afford.
The 1973 change shows the same threshold logic without a questionnaire score. When Robert Spitzer, the psychiatrist who did the most to broker it, argued for removal, his case was definitional. He proposed that a condition should count as a psychiatric disorder only if it regularly causes the person subjective distress or produces a general impairment in how they function. Homosexuality, on its own, did neither for a great many people, and the distress that did occur usually traced to social hostility rather than to the orientation. Under that definition, homosexuality as such no longer met the standard for a disorder, and the argument in 1973 was about where to set that bar. The category had been moving for decades, from a personality disturbance in 1952, to a sexual deviation in 1968, to removal of the general category in 1973, with a residual diagnosis for people distressed by their orientation deleted in 1987. Across these revisions the orientation did not change; the diagnostic threshold for disorder did.
Both research and activism drove the change. Research had weakened the claim that homosexuality was inherently tied to psychological dysfunction, while activism forced the profession to confront the harm the label produced and the values buried inside its criteria. The decision became possible once the question shifted from whether homosexuality departed from a social norm to whether it involved the kind of distress or impairment that psychiatry should treat as disorder.
Classifying people changes them in a way that classifying rocks does not, as the philosopher of science Ian Hacking argued. “They are moving targets because our investigations interact with them, and change them,” he wrote of the kinds of people the human sciences invent. “And since they are changed, they are not quite the same kind of people as before. The target has moved. I call this the ‘looping effect’.” A person told they have a condition may come to understand their moods, choices, and history through it, may seek out others with the same label, and may adjust what they expect of themselves and what clinicians expect of them. Those changes feed back into the category, altering who fits it and what it seems to mean. Hacking’s most vivid case was multiple personality: as the diagnosis spread through the 1980s the reported number of personalities climbed steeply, so that “first, a person had two or three personalities” and “within a decade the mean number was 17”, each new expectation shaping the next round of patients.
The feedback runs in both directions, helping in some cases and harming in others. A diagnosis can give a person language for a pattern that previously felt chaotic, reduce self-blame, open access to treatment, and make suffering legible to family, employers, and health systems. It can also narrow a person’s sense of what they can do, make a passing crisis feel permanent, or lead institutions to see the label before the individual. A diagnostic label matters, then, because it sorts a graded population into the treated and the untreated, and then changes how the sorted people are handled and how they regard themselves, which changes the population the category picks out the next time it is applied.
Automated screening makes the chosen threshold more consequential, because it takes a cut-point some committee selected and applies it to many thousands of people who never sat with a clinician. A tool that reads a questionnaire, or a stream of behavior such as how someone types or sleeps, and returns a judgment about their mental health hard-codes that boundary and repeats it. A human clinician can treat a threshold as a guide, noticing a recent job loss or a bereavement before applying a code. A model has to operationalize the boundary: it needs a target label and a cut-off that converts uncertainty into an action, whether to flag, escalate, reassure, or ignore. The value judgment a clinician makes case by case, and can be argued with, is made once, invisibly, and then repeated at scale.
Once that cut-off is built into software, it can look more objective than it is. A questionnaire score or a model probability can appear more precise than the judgment behind it, and its users are strongly inclined to treat a machine’s output as an objective discovery. If the system was trained to identify “depression” or “risk”, it inherits earlier decisions about who counted as a case, who counted as a control, and how much distress justified intervention, and then applies that inherited boundary with a consistency that feels like neutrality. Consistency helps when the threshold is well chosen and the pathway after a positive result is humane, and it does harm when the threshold is poorly matched to the setting. A primary-care clinic and a crisis line do not face the same costs of false positives and false negatives. A low threshold can be appropriate when the next step is a private, well-supported offer of help, and damaging when it triggers surveillance, exclusion, or a permanent record. The statistical choice and the ethical choice cannot be separated, because the meaning of an error depends on what happens after the flag.
Building the boundary into software also makes it harder to revise when values shift, as they did around homosexuality. A human institution can revise a threshold through argument and review, while a deployed model keeps applying the old cut-point until someone changes its target, its threshold, or its deployment policy. Automated systems can also accelerate the feedback loop, because a flag can route a person to services, change how a platform treats them, or shift how they see themselves, and if those flagged responses are later reused to train or evaluate the system, the original threshold can harden into an apparent fact.
Health systems with finite clinicians and finite hours still have to use thresholds, because they cannot treat a continuum and must decide whom to see this week and whom to reassure. Cut-points let clinics compare cases and trials define who was enrolled, and they give a person something more concrete than a location on a scale. The workable position is that thresholds are necessary tools whose placement should stay visible and open to argument. Thresholds can be revisited when the evidence or the values behind them move, the way the line around homosexuality was moved and then erased, as long as they are remembered as decisions. When a threshold is instead treated as a discovered fact, it becomes harder to revise even as evidence, harms, and social values change, and it hides a choice that someone made and that someone else has to live inside. The 1973 vote is worth remembering because the profession made its choice in the open, stating that the boundary of an illness was one it had set and could reset.