Two boxes of the same biscuit sit side by side on a shelf. Same brand, same weight, near enough the same price. One of them carries four small words tucked under the ingredients list — may contain nuts. The other says nothing about nuts at all. A hand looking for the safer option goes, almost without deciding to, toward the quiet box. No warning printed on it, so nothing to worry about. The choice feels like reading. It is closer to a reflex.
In 2017 a group of scientists at the United States Food and Drug Administration and the Illinois Institute of Technology set out to check whether the reflex held up. They bought dark-chocolate bars off American shelves (ordinary ones, the kind that end up in a coat pocket) and ran each of them through a laboratory test for peanut and for milk, two allergens the bars were not supposed to contain. Then they lined the results up against what each wrapper actually said.
This was one kind of food, dark chocolate; one allergen, peanut; one country; one year; fewer than ninety bars. Within those limits, the warning did not track the food in either direction. Among the bars that carried a peanut advisory, only about 8% actually contained any peanut — more than nine in ten of the warned bars were clean. And among the bars that said nothing about peanut at all, the quiet ones, 17% contained it anyway, at levels the test read without difficulty. The milk figures were sharper still: a third of the bars that made no mention of milk turned out to contain it, some of them heavily.1
None of this means the silent pack and the warned pack are an even bet. On the wider numbers, a warning is a weak sign that something might be there. A Canadian survey of more than eight hundred packaged foods found peanut in about 5% of products carrying a peanut warning, against 0.4% of products that instead claimed to be made in a peanut-free facility.2 A European survey of 569 cookies and chocolates found that a precautionary label was, if anything, tied to a higher frequency of positive results, not a lower one.3 Averaged across a whole shelf, the warned packs are marginally the more likely to hold the allergen.
But look at what the signal does when you move between the studies. The Canadian and European surveys point one way — the warned pack is the riskier one, so silence is faintly safer. The American survey of silent bars points the other: among the bars saying nothing, 17% carried peanut, more than double the warned rate. When the direction of a signal flips from one dataset to the next, it is not something a single shopper can reliably rely on at the shelf. And it was never really the shelf-wide average that mattered anyway. It was this box, in this hand. Reaching for the quiet pack is the one move the label cannot reliably reward: you cannot know whether the biscuit you have just picked up is one of the clean ones or one of the seventeen in a hundred that are not. And no test in the world could tell you.
No laboratory can tell an honest warning from a reflexive one
Picture the two boxes again, but change one thing about how they were made. Behind the first, a manufacturer ran a proper risk assessment (checked the shared line, the cleaning between runs, the airborne dust), decided a real trace of nuts could survive, and printed the words. Behind the second, someone printed the same words without doing any of that, because printing them was easier than being sure. Two identical packs, three identical words, two completely different stories underneath.
Hand both to a laboratory and ask it to tell them apart. It cannot, and not because the equipment is poor. The thing you would need to measure, the absence of contamination, is not a property a finished pack actually has in any stable, readable way.
Cross-contact allergens do not spread themselves evenly through a batch. A stray fragment of peanut is a hot-spot, not a wash. The bodies that write the testing rules are blunt about what this means: a result is only ever representative of the exact crumbs that went into the sample, and testing a blended, composite sample can dilute a real hot-spot until it drops below the level the test can see. A clean result, in other words, never proves a clean pack — it proves that the particular pinch of food you ground up happened not to contain the thing. Even an extensive sampling programme, the same guidance warns, may miss cross-contact allergens depending on how they are distributed.11
The chocolate survey caught this happening in real bars. When the scientists tested several production lots of the same product for peanut, they got mixed results — the same bar coming back clean in one lot and contaminated in another.1 Milk, which was there as an ingredient rather than a stray trace, stayed consistent from lot to lot; the trace did not. That is the whole problem in one line: the very thing the warning is about is the thing that refuses to sit still long enough to be measured.
So when a business decides whether to print may contain, it is not reading a number off a machine. It is making a judgement — an assessment of a risk that is calculated, estimated, argued over, never simply detected. Even the new international guidance, the one meant to fix all this, says a precautionary warning should be used only after a business has carried out a scientific risk assessment showing that a residual risk remains. An assessment is an act of judgement. And a judgement leaves no fingerprint on the pack. There is nothing in the printed words, and nothing a test could pull out of the food, that separates the warning made after careful work from the warning made to save the trouble of it. That is not a story about anyone hiding a result. There is no result to hide. The warning is un-gradeable by design, and the design came before any individual maker touched it.
The warning tracks the maker's legal exposure, not the food
If no instrument decides whether those four words go on the pack, a person does. On what basis, and in whose interest?
Start with what the warning legally is. In Britain the Food Standards Agency is plain that a precautionary allergen label is voluntary: a tool a business may choose to use. The agency's own paper says businesses may choose to warn consumers about this risk by using such a label.8 Nothing triggers it: there is no threshold a batch has to cross, no test that switches it on, and (this is the part that decides everything) no penalty at all for printing it when there was no real risk.
Now weigh the two ways a manufacturer can be wrong. Print the warning when the food is in fact clean, and the cost is a few millimetres of ink and a shopper who avoids a biscuit she could have eaten. Leave the warning off when a trace does get through, and someone has a reaction, and the cost is a hospital, a recall, a headline, a lawsuit. The price of over-warning is symmetric and tiny. The price of under-warning is enormous and one-directional. Set those payoffs in front of any rational business and the answer writes itself: when in doubt, print. When barely in doubt, print anyway. The words cost nothing to add and everything to have left off.
Which means the presence of may contain is not tracking the food. It is tracking the manufacturer's legal exposure — the size of the downside they are insuring against. The warning is a shield held up against the courtroom, and it works just as well whether or not there is anything behind it to guard against.
The regulator does not dispute this. The Food Standards Agency's own 2026 paper says confidence in the current approach is lacking, and that it is precisely this lack of confidence that is driving the push toward a standard. It records that consumers wrongly believe the system is more regulated than it is. And it notes, among the concerns charities have raised, a suspicion about whether the warnings are being used for the benefit of the people reading them or the businesses printing them.8 When the body responsible for the regime writes down that its own warning does not reliably reflect risk, the question stops being whether the words are noise and becomes what the noise is doing to the people who cannot afford to ignore it, and whether anyone has ever proven the makers themselves don't believe them.
When they had the choice, the makers changed the food instead
There is a proof, and it comes from watching what firms do when the incentive shifts.
In the United States, sesame became a labelled allergen: a substance that, if present as an ingredient, must be declared under the mandatory "contains" rules, the same tier as peanuts or milk. Economists studying the aftermath expected to see firms managing sesame more carefully, or reaching more often for precautionary warnings on products where cross-contact was possible. That is not mainly what happened. Across roughly twenty food categories, manufacturers responded by adding sesame to the recipe: deliberately putting the allergen into products that had never contained it, so those products would fall under the clean, mandatory "contains sesame" umbrella.5
Read that twice, because it is the market appraising its own warning label out loud. Faced with a choice between relying on a precautionary line plus careful manufacturing (a combination the firms themselves described as leaving them exposed to possible risk) and simply reformulating the food so the exposure disappeared, a documented slice of the industry chose to change the food. Adding sesame to hundreds of products across those twenty categories was, for them, the safer bet than trusting may contain to hold up.
This is not cynicism, and it is worth being fair about that. These firms did not write the rules that make over-warning rational; they are playing a payoff matrix someone else built, and playing it sensibly. That is exactly why it is so damning. When the people who print the warning would rather alter the contents of the food than depend on the warning to protect them, they have told you, in the most expensive language available, what they think those four words are worth as a shield. The answer is: not much. And if the makers won't lean on the warning, consider who is left holding it.
The people the warning is for have learned to stop reading it
The people left holding it are the ones with no choice but to read every label — and the warning, printed everywhere and meaning almost nothing, has quietly trained them out of it.
In a 2011 study, researchers sat down with nut-allergic adults and asked how they actually handled may contain. One participant described the slide exactly: it is the over-usage of it, he said, the boy-who-cried-wolf syndrome, and after a while you become blasé, and you just go, well, I'm going to eat it anyway. The same testimony turns to the other side of the trap: if he avoided everything that said may contain traces of nuts, he would be left with a very narrow spectrum of food that he could eat.4 That is the choice the noise forces on a person with a serious allergy. Take every warning literally and your diet collapses to a handful of safe boxes. Learn to discount the warnings so you can eat, and you have taught yourself to walk past the one that was true.
And the discounting has a price that shows up in the injury figures. In the allergy literature, one study found that of the people who had had an accidental reaction, 8% attributed it to having ignored a precautionary label — a warning that was, for once, worth heeding, drowned in all the ones that weren't.4
This is the line the story will not cross: un-gradeable is not the same as ignorable. That a warning cannot be graded, cannot be told apart from the reflexive one, does not make it safe to disregard. For someone whose reaction runs to anaphylaxis, an unreliable flag is still the only flag there is, and the honest response to the noise is not to stop looking but to know exactly how little the looking tells you. The cost runs in both directions at once. Habituation is dangerous. So is a life narrowed to a short list of trusted packets. The regime manages to inflict both.
It is not only allergy labelling that ends up here. In California, the state's own environmental regulator looked at a different threshold-free warning — the short "this product can expose you to chemicals" notice on countless goods — and conceded, in its own rulemaking record, that the warning was being used prophylactically because it protected businesses from litigation, and that this over-use was diluting the impact of legitimate warnings.12 A regulator, on the record, describing a free-to-print warning training the public to ignore all warnings. The mechanism is not unique to food. It is what any warning does when it costs nothing to give and nothing to overuse.
The number that would make the warning mean something arrived in July
For decades the obvious question has been whether this can be fixed at all, whether the problem is the words themselves or something missing underneath them. It is the second. What has been missing is a number.
To grade a warning, you need something to anchor the judgement to: a measured quantity of allergen below which the risk is treated as residual and above which the warning is genuinely earned. Toxicologists call the reference points reference doses, and for most major allergens the science to set them has existed for years. Gluten was the last holdout — the one allergen without an agreed figure, and so the one loose thread keeping the whole standard from being tied off.
This investigation continues below.
Want the next one in your inbox?
In November 2025 a panel of experts convened by the Food and Agriculture Organization and the World Health Organization met in Rome to close it. The reference doses they work with are what specialists call ED05 figures, and the meaning matters more than the arithmetic: ED05, in the Food Standards Agency's own words, means that 5% of people with the relevant allergy who are exposed at that level will react.8 Holding that meaning firmly in view, the experts recommended a reference dose of 4 milligrams of gluten, replacing an older value pinned to 5 milligrams of wheat protein.7 With gluten resolved, the last gap closed. On 7 July 2026, at its 49th session, the Codex Alimentarius Commission, the international body that writes food-labelling standards, adopted risk-based guidance on precautionary labelling as an annex to its general labelling standard.6
Here is where it is easy to say the wrong thing, so it is worth saying the right one carefully. A reference dose does not make the shelf safer. It works by giving a firm permission to drop the warning wherever an honest assessment puts the residual risk below the line — and because 5% of allergic people still react at that line, adopting the standard would strip may contain off some packs that would still send one in twenty allergic people into a reaction. For the most sensitive reader, a regime that flags everything is, in the narrow sense of coverage, more protective than one that flags only above a threshold. What the number buys is not safety. It is legibility. Once a warning has to correspond to a measured quantity, a firm can no longer print it for free, and the words stop being a shield and start being information. That is the whole of the gain, and it is a real one, but it is a gain in what the warning means, not in what it protects.
Credit where it is due: the Food Standards Agency got the order of operations right. At its board meeting in December 2025, an official, James Cooper, presented a recommendation to support Codex's work to standardise precautionary labelling internationally, but only on the condition that the treatment of gluten in the standard was resolved first. The board agreed on exactly that condition and asked for an implementation proposal to follow once Codex reached agreement.9 It declined to back a risk-based scheme while the science was still incomplete, set the precise gap that had to be filled, and waited. The gap it named was the one the Rome experts had, weeks earlier, just filled.
Then Codex agreed, on 7 July. A month on, the implementation proposal the board asked for has not appeared, and no date has been set for Britain to adopt the annex.9 Codex standards are voluntary; no country is obliged to take them up, and Britain has not yet done so. The right way to read this is not as a dereliction (the standard is four weeks old and the agency has behaved carefully throughout) but as a watch-this-space. The last scientific reason to wait expired in July. The domestic move has not yet come. If the proposal is published next month, this part of the story changes with it. What does not change on any calendar is the reason the warning has meant so little for so long: that reason sat in the incentive, not the science, and it is still sitting there today.
Only "gluten-free" carries a number you can check
Which leaves a fair question for anyone standing at the shelf right now, before any of this is adopted: is there anything on a British pack today that carries a number you can actually check? There is exactly one.
"Gluten-free" is not a precautionary hedge. It is a legally defined claim. Under the regulation retained in UK law, the words gluten-free may only be used where the food as sold contains no more than 20 milligrams of gluten per kilogram, or 20 parts per million.10 That is a concentration a laboratory can measure and, more to the point, a claim the manufacturer is legally liable for having made. Beside it on the shelf, may contain gluten carries no number at all.
The difference between the two is not that the gluten-free test is somehow better. The same sampling limits that make may contain un-gradeable apply to a gluten-free test as well — a composite sample can miss a hot-spot of gluten just as it can miss a hot-spot of peanut. The difference is legal and economic, not analytical. "Gluten-free" is a positive claim a maker chooses to assert, tied to a fixed number, and answerable for if it turns out to be false. "May contain" is a negative hedge the same incentive structure pushes firms to over-apply, tied to nothing, and answerable for mainly if it is absent. The reliability of the one comes from a defined figure plus liability for a positive assertion; the unreliability of the other comes from no figure plus a liability that runs the wrong way. The surveys bear the contrast out: where a pack made a positive free-from claim, it largely held. None of the nine free-from chocolate bars carried peanut, and the peanut-free-facility products came in at 0.4%. It was the warnings that were noise, not the claims.
And the two numbers, when both finally exist, will not sit on one scale. The 20-milligrams-per-kilogram gluten-free threshold is a concentration in the food as sold, the same whether you eat a mouthful or the whole packet. The 4-milligram gluten reference dose is a mass of protein per portion, a quantity that depends on how much you eat, and a figure the pack does not even print.8 They measure different things with different rulers. The experts in Rome pointed out that a food eaten in large enough portions could, at the same time, count as gluten-free by concentration and still cross the reference dose by amount.7 Two instruments, two questions, never one line on a chart.
What the four words are, and are not, telling you
Back to the two boxes. The presence or absence of may contain on the pack in your hand tells you almost nothing about whether that pack is contaminated. Not because a maker cheated, but because the warning is applied by the logic of legal exposure and can be printed for free, so a silent pack is not a clean pack, only a firm that chose not to warn, and a warned pack is usually a firm that chose to be careful about something that probably isn't there. The one cue on the whole shelf that means what it says is a positive claim carrying a number and a liability behind it (gluten-free, 20 milligrams per kilogram), not the blank space where a warning happens not to be.
None of which is permission to ignore the four words. For anyone whose reaction is severe, an un-gradeable warning is still the only flag there is; the honest move is to keep avoiding, to trust the numbered claim wherever one exists, and to stop reading safety into silence. What has changed is only what you can know while you do it. The words are a record of who might be sued, not a reading of what is in the food. The sentence that would finally put a measured number underneath them, and turn the shield back into information, was written on 7 July 2026. Britain has not yet reached for it.