The symptom that was named after its own relief.

In June 2023, the New York Times reported that patients taking semaglutide kept describing the same thing: a constant internal commentary about food had gone quiet. Videos tagged with the phrase had already been viewed 1.8 billion times on TikTok. At that moment the term appeared in zero approved drug labels, zero registrational endpoints, and zero validated instruments.

Here is the thesis. Food noise did not enter medicine the way clinical concepts normally do. The usual sequence runs mechanism, then construct, then measurement, then drug. Food noise ran backwards: drug, then patient report, then press, then marketing, then construct, then — two years later — measurement. Every instrument now being built is being fitted retroactively to a symptom that was defined by the experience of its own relief.

I think that inversion is the most underrated problem in metabolic medicine right now, and it is not a semantic complaint. A construct built backwards from a drug's effect profile will tend to describe what that drug does rather than what the underlying biology is. That is a measurement error with a very long half-life.

The minimum background. The phrase has no identifiable author. It surfaces in web search records as far back as 2006, sat at negligible volume for fifteen years, and rose sharply from 2022 alongside the first wave of semaglutide prescribing for weight. It was formally incorporated into a scientific model only in November 2023, when Hayashi and colleagues published What Is Food Noise? A Conceptual Model of Food Cue Reactivity in Nutrients, proposing what they called the CIRO model — cue, influencer, reactivity, outcome.

The literature already had a word for this, and it was not food noise

This is the part that gets skipped in almost every piece written on the subject. Researchers were not silent about intrusive food cognition before 2022. They were studying it under a different name, with better-developed tools, for roughly two decades.

Food cue reactivity is the conditioned response to internal cues, like a stomach signal, and external ones, like the sight or smell of food. It is measured, it is reproducible, and in meta-analysis it has been identified as the single strongest measurable behavioural predictor of weight gain. Hedonic hunger — eating driven by palatability rather than energy need — has had a validated instrument since 2009 in the Power of Food Scale. The Control of Eating Questionnaire, a nineteen-item instrument covering craving control, craving for sweet, craving for savoury, hunger, fullness and mood, has been running inside obesity trials since long before anyone had heard of Ozempic.

Note what Hayashi's team actually did in 2023. They did not propose food noise as a new entity. They located it inside food cue reactivity, as a cognitive expression of a construct that already existed. That was the careful move, and the field has largely ignored it in favour of treating food noise as a discrete symptom with its own presumed lesion.

The measurement arrived two years after the marketing

The first validated instrument specific to the term is the Food Noise Questionnaire, published by Diktas and colleagues in Obesity in 2025 — five items, psychometrically tested, presented as the first tool to quantify intrusive food-related thoughts in people with overweight and obesity.

What happened next is the detail I find most revealing, and it is almost never reported. Hayashi and colleagues — the authors of the conceptual model the questionnaire was built on top of — published a formal Response in the same journal contesting it. The people who defined the construct and the people who built the instrument are in open, published disagreement about whether the instrument measures the construct.

Then, in late 2025, a second scale: the RAID-FN Inventory, developed with academic obesity researchers and launched by Ro, with data presented at ObesityWeek 2025 and versions released in English and Spanish. I want to be precise here, because the point is structural rather than accusatory. Ro is a telehealth company whose revenue depends on this drug class. It now co-owns one of the two leading instruments for measuring the symptom that class is credited with relieving. That is not fraud and I am not alleging any. It is a conflict of interest sitting at the definitional layer of a field, which is a harder problem than one sitting at the trial layer, because everything above it inherits the shape.

What the trial data actually shows

The best long-run evidence comes from the STEP 5 control-of-eating substudy, published by Wharton and colleagues. Participants on semaglutide 2.4 mg or placebo completed the nineteen-item CoEQ at baseline and at weeks 20, 52 and 104. In the substudy population — 88 on semaglutide, 86 on placebo — mean weight change at two years was −14.8% versus −2.4%.

The craving domains separated and stayed separated. Craving Control and Craving for Savoury were significantly better on semaglutide at all three timepoints, weeks 20, 52 and 104, at p < 0.01. Craving for Sweet and Positive Mood separated at weeks 20 and 52, at p < 0.05.

Now the finding almost nobody quotes. Hunger and fullness separated at week 20 at p < 0.001 — the strongest early signal in the whole substudy — and are not reported as holding through week 104. The appetite signal faded. The craving signal did not. If food noise were simply hunger by another name, those two lines should have decayed together. They did not, and that dissociation is the most interesting number in the CoEQ literature.

One honest caveat on the same data. Craving improvements correlated with weight reduction, and a correlation like that runs in both directions at once. It is not possible to tell from it whether quieter cravings produced the weight loss or two years of successful weight loss produced quieter cravings. This is the same interpretive trap that shows up whenever a subjective endpoint travels alongside a visible one — a problem I have written about before in the context of how the STEP-4 results get quoted.

The mechanism is inferred, not observed

Here is what makes this a mechanism story and not just a nomenclature story. There is no biomarker for food noise. No imaging endpoint, no assay, no physiological correlate that a trial can register. The entire mechanistic account is assembled indirectly, from where GLP-1 receptors are known to sit — the hypothalamic arcuate nucleus, the area postrema, regions relevant to reward valuation — and from the behaviour of people who report the change.

That inference has a specific weakness. Semaglutide is a large peptide with limited blood-brain barrier penetration, so much of its central action is understood to route through circumventricular organs and vagal afferents rather than through direct engagement with the cortical machinery that would plausibly generate an intrusive thought. The distance between "receptors exist in appetite-relevant circuits" and "this molecule reduces the frequency of food-related cognitions" is being crossed by questionnaire, not by measurement.

And the questionnaire is administered to people who are visibly losing weight and, in most real-world settings, know exactly what they are taking. Functional unblinding is a well-understood hazard for subjective endpoints in obesity trials. It does not invalidate the finding. It does mean the field currently cannot separate three live explanations: genuinely reduced cue reactivity, reduced hunger reallocating attention elsewhere, and the ordinary psychological relief of a number that is finally moving in the right direction.

The strongest case against everything I have just argued

Patient-coined language has an excellent track record in medicine. Tinnitus, phantom limb pain, brain fog — plenty of durable clinical constructs began as descriptions from people who had the experience and no vocabulary handed to them. The absence of a term before 2022 may say more about medicine's inattention than about the concept's validity. On this reading, my objection is a purity argument about sequencing, and sequencing does not determine truth.

There is a stronger version still. Language follows utility. Before there was an intervention that made the distinction between hunger and intrusive food cognition actionable, the field had little reason to separate them. Once there was, the separation became worth naming. That is not backwards. That is how clinical vocabulary has always worked.

I take that seriously, and I would accept it if the construct were converging. It is not. Convergence looks like competing instruments producing correlated scores and a shared definition settling into place. What we have instead is two scales inside eighteen months, a published dispute between their authors, a commercial stakeholder at the definitional layer, and a term whose reach in culture — billions of views — exceeds its reach in evidence by several orders of magnitude.

What I think this changes

Something real is being measured. The STEP 5 craving dissociation is not noise, and dismissing patient reports because the vocabulary is untidy would be its own kind of error. But the honest description of the current state is that we have a well-evidenced effect attached to a poorly-specified construct, and the field is using the strength of the first to borrow credibility for the second.

The way out is not a better name. It is a discriminant validity study: does a food noise score predict anything that food cue reactivity and the Power of Food Scale do not already predict? Until someone runs it, food noise remains the most-cited concept in metabolic medicine and one of the least understood, and those two facts are related.

Ozemback — August 2026