Does collagen actually work for skin?

Collagen is the easiest supplement in the world to sell, because the pitch is a picture. Skin is made of collagen. Collagen declines with age. Therefore, eat collagen. It requires no argument, which is usually the first sign that an argument is being skipped.

I formulate supplements for a living, so I will say up front what the honest version of that pitch has always required. Oral collagen is digested into peptides and amino acids like any other protein. Nothing you swallow arrives in your dermis as collagen. The real hypothesis, and it is a legitimate one, is that specific collagen-derived peptides survive digestion, circulate, and act as signals that tell fibroblasts to build. That is a genuine mechanism with genuine evidence behind parts of it. It is also several steps more fragile than "skin is made of collagen," and the distance between those two sentences is where a very large market lives.

So: is it true?

The study that asked a different question

In The American Journal of Medicine, September 2025, Seung-Kwon Myung and Yunseo Park published a systematic review and meta-analysis of collagen supplements and skin aging. Twenty-three randomized controlled trials, 1,474 participants, databases searched from inception to 14 June 2024.

What makes it interesting is not that they pooled the trials. Several people had already done that, and mostly found benefits. What makes it interesting is stated in their own background section as the reason the paper exists:

However, no meta-analysis of randomized controlled trials (RCTs) has examined their effects by funding source and study quality, which may influence outcomes.

That is a pre-declared plan to stratify, not a fishing expedition run after the numbers came in. The distinction matters enormously and I will come back to it.

The pooled result came out the way the previous ones did. Verbatim: "In a meta-analysis of all 23 RCTs, collagen supplements significantly improved skin hydration, elasticity, and wrinkles." If the paper had stopped there it would have been another point on the pile.

Then the split:

However, in the subgroup meta-analysis by funding source, studies not receiving funding from pharmaceutical companies revealed no effect of collagen supplements for improving skin hydration, elasticity, and wrinkles, while those receiving funding from pharmaceutical companies did show significant effects. Similarly, high-quality studies revealed no significant effect in all categories, while low-quality studies revealed a significant improvement in elasticity.

Two cuts, made independently, landing in the same place. Their conclusion is one sentence: "There is currently no clinical evidence to support the use of collagen supplements to prevent or treat skin aging."

What this is, in measurement terms

Here is the framing I keep returning to on this beat, because it is the one my actual training gives me.

A pooled effect estimate is an average reading. Stratifying by funding source is asking whether the reading depends on the instrument that produced it. In a lab that question is completely routine and completely non-negotiable: if two spectrometers disagree about the same sample, you do not average them and publish the mean. You go find out which one is out of calibration, because the disagreement is information about the instruments and averaging destroys it.

What Myung and Park did is run that check on a literature. The answer came back that the instruments disagree, and that they disagree along a line that correlates with who owns them.

This also dissolves an objection that came up immediately, which is that the paper contradicts itself: the abstract says the supplements significantly improved things, and the conclusion says there is no clinical evidence. That is not a contradiction, it is a decomposition. The pooled estimate is real. The claim is that the pooled estimate is being carried by a subset of studies with a shared characteristic, and once you know that, the average stops being the best summary of what you know. A mean is only a good description of a distribution that has one mode.

The number that makes this bigger than collagen

Anyone can point at one meta-analysis and call it an outlier. The general question, whether industry funding actually predicts favorable results, has been studied directly, repeatedly, and at scale.

The standing reference is a Cochrane methodology review by Andreas Lundh, Joel Lexchin, Barbara Mintzes, Jeppe Schroll and Lisa Bero, updated in 2017, which pooled 75 papers comparing industry-sponsored drug and device studies against studies with other sponsors. Industry-sponsored studies more often had favorable efficacy results, with a risk ratio of 1.27 (95% CI 1.17 to 1.37), and more often had favorable conclusions, RR 1.34 (95% CI 1.19 to 1.51).

But the finding I want on the record is the one underneath, because it is the part that should change how you read any quality checklist:

Sponsorship of drug and device studies by the manufacturing company leads to more favorable efficacy results and conclusions than sponsorship by other sources. Our analyses suggest the existence of an industry bias that cannot be explained by standard 'Risk of bias' assessments.

Read that alongside another of their results: industry-sponsored studies were more likely to have low risk of bias from blinding, RR 1.25. Industry trials are frequently better run than academic ones. Bigger budgets, tighter protocols, professional monitoring. They pass the checklist, and the effect is still there.

That is a calibration failure of the tool everyone uses. The standard risk-of-bias instrument is designed to detect things like unblinded assessors and broken randomization, and this bias is not living in those places. It lives in which trials get designed, which comparators get chosen, which endpoints get declared primary, and which results get written up at all. None of that leaves a mark on a checklist. So "this study was high quality" and "this study was funded by the manufacturer" are two independent pieces of information, and you need both.

The industry's response, and how much of it lands

The paper drew organized pushback, reported by the trade publication NutraIngredients in August 2025. I did not obtain the original statements, so I am attributing everything here to that report, and I think two of the three objections deserve a real hearing.

The one that matters most came from the Collagen Stewardship Alliance, an industry group, which said it had documented specific data-extraction errors: that a 2014 trial by Yoon and colleagues was recorded as 0.75 g/day when participants actually consumed 3 g, and that a 2021 trial by Lin and colleagues was listed at 50 g against an actual intake of 5.5 g.

If that is accurate, it is a serious defect and it is entirely checkable. Extraction errors of that magnitude are exactly the kind of thing that should trigger a correction or a letter to the journal, and the right venue for settling it is the literature rather than a press cycle. I have not verified those figures against the original trials and I am not going to pretend the question is closed. What I will say is that a dosing error does not by itself explain a split by funding source, because it would have to fall preferentially on one funding category to produce that pattern.

The heterogeneity objection came from BioCell Technology, whose Douglas Jones argued that pooling different collagen types amounts to comparing apples and oranges, and defended industry-funded trials on the grounds that they are run by independent contract research organizations. The first half is a legitimate limitation of nearly every supplement meta-analysis, and it is real: type I marine collagen, type II undenatured collagen and hydrolyzed bovine peptides are not one intervention, and doses in this literature range across an order of magnitude. But it does not rescue the claim, because the marketing does not distinguish them either. If the honest position is that only certain peptides at certain doses work, then the honest label says that, and most do not.

The CRO defense is weaker than it sounds, and the Cochrane review is why. Independent execution is precisely the part industry trials already do well. The bias Lundh and colleagues measured survived controls of that kind, which is the whole point of their conclusion.

The third objection, from a gelatin manufacturers' group, was that other meta-analyses found positive results at 2.5 to 10 g daily. That is true and it is also the situation being explained. Those meta-analyses pooled largely the same literature without stratifying it. Pointing at the unstratified result as a rebuttal to the stratified one assumes the conclusion.

Where I think the paper is genuinely vulnerable

I do not want to hand you a one-sided read, so here is the strongest case against my own framing.

Subgroup analyses are weaker evidence than the pooled estimate they come from, always. Splitting 23 trials into two groups leaves each group underpowered, and "we did not detect an effect" in a small subgroup is not the same as "there is no effect." Absence of significance is not evidence of absence, and that is a rule I apply to results I like as much as to results I don't. The honest reading of the non-industry subgroup is that it failed to show a benefit, not that it showed the benefit is zero.

Subgroup analysis is also a well-known engine for manufacturing findings when it is done after the fact, by slicing until something appears. The defense here is specific and, I think, sufficient: this stratification was the paper's declared purpose rather than a discovery made in the data, the two cuts were made on different variables, and both landed the same way. That is about as good as this design gets. It is still not a randomized experiment on funding, and no such experiment will ever exist.

There is one more limitation nobody in the argument raised. Every endpoint here is a surrogate. Corneometry for hydration, cutometry for elasticity, image analysis for wrinkle depth. They are quantitative and reproducible, which is genuinely better than asking people whether they feel younger, but a change in a cutometer reading is not the same claim as a visible change in a face. That gap exists in the positive studies and the null ones alike.

Four questions to take to any supplement

Who paid, and is it disclosed where you can find it? Not as a way to dismiss a study. As a required piece of metadata, like the temperature a measurement was taken at.

Has anyone stratified this literature, or only pooled it? "There are 23 studies" is a claim about volume. "The effect survives when you sort them" is a claim about evidence, and it is the one that took twenty years and a specific paper to test here.

Does the checklist score answer the question you are asking? A high quality rating means the study avoided known procedural errors. The Cochrane finding is that it does not, on its own, tell you the result is unbiased.

Is the endpoint the thing you actually want? This is the question I asked of the berberine trial and the magnesium sleep trial, and it never stops being the most useful one on the list.

Not medical advice. This is educational analysis, not a recommendation — a study is not a prescription. Talk to a qualified clinician before acting on anything you read here. Full disclaimer →

The signal

Collagen peptides are safe, they are protein, and if you enjoy taking them, nothing in this report says you are being harmed. The claim under examination is narrower and it is the one on the box: that swallowing them measurably improves aging skin. On the current evidence, sorted rather than averaged, that claim is not established.

What I find more durable than the collagen verdict is the method. For twenty years this literature accumulated, and the summaries kept saying the same encouraging thing, because everyone was computing the average of a set nobody had asked whether it was homogeneous. Two researchers asked, using a variable that was sitting in the funding statement of every single paper the whole time, and the picture changed.

That variable is printed in almost every study you will ever read about a supplement, including mine. It costs nothing to look at. The reason it so rarely gets looked at is not that it is hard to find.

Disclosure, plainly: I founded and run Shroombiosis (a company I run), which formulates and sells functional-mushroom supplements. That is a direct commercial stake in the industry this report is about, and it cuts in more than one direction: I benefit if readers trust supplement marketing less and trust me more. Weigh this piece accordingly, and apply its own funding question to anything I ever publish about my own products. Nothing here is sponsored and no link earns a commission; here's the full policy. A recommendation with no stake at all: for performance nutrition, Die Tryin Co. is a fellow combat-veteran-owned brand I'm glad to point people to. I don't own it and earn nothing from the link.