Glycemic Control and Metabolic Supplements

01 What this document is, and five things it is not

This article asks one question of a large commercial category: when a product is sold for blood sugar, insulin resistance, or metabolic health, what has actually been measured, in whom, for how long, and against what comparator?

The category is unusual in one respect that makes it easier to write about than most of this series. Glucose metabolism has hard, cheap, standardised endpoints. Fasting plasma glucose, a two-hour value after a glucose load, glycated haemoglobin, and a fasting insulin-derived resistance index are all numbers, measured the same way in Beijing and Boston, with diagnostic thresholds set by the same professional bodies (American Diabetes Association Professional Practice Committee, 2026a). Nobody has to argue about whether the outcome was real.

That is also what makes the category treacherous. Because the surrogates are easy to move, a supplement can produce a genuine, statistically significant, reproducible change in a number that carries a diagnostic name — and still be irrelevant to whether the person taking it goes blind, loses a foot, or has a heart attack. This article is largely about the distance between those two things.

Five things follow immediately.

First, this is not a treatment manual. It recommends no compound, amount, route, or schedule for any person. Several of the compounds discussed here lower glucose. That is a reason for caution, not a dosing instruction, and the interaction problem in section 22 is the whole reason.

Second, it is not an ingredient catalogue. The individual compounds have their own articles in this series — Berberine, Alpha-Lipoic-Acid, Cinnamon, Fenugreek, Magnesium, Chromium, Resveratrol, Flavonoids, Dietary-Fibre — and those remain the deep references. The existing Metabolic-Nutrition article carries the dietary-pattern literature and is cross-referenced rather than restated. The work of a system article is synthesis: which outcomes matter, which mechanisms are plausible, which claims survive contact with a randomised trial, which work only in people who are already ill, and which combinations exist only because they appear together on a label.

Third, it is not a diabetes-management document. Where a compound has been tested in people treated for type 2 diabetes, that evidence is reported as what it is, and is never quietly transferred to a reader with normal glucose who has bought a bottle because a wearable sensor showed a spike after lunch.

Fourth, it is not neutral about the word "support". Blood-sugar support is a marketing phrase that maps onto no measurable quantity. Where a product's evidence consists of a mechanism, this document says so.

Fifth, it is not neutral about the shape of the literature. A large part of the apparent disagreement in this field is not biological. It is what happens when a surrogate endpoint is cheap, sample sizes are small, trial quality is uneven, and positive results are easier to publish. Section 05 treats that as a first-order fact rather than a caveat.

02 The physiology, in one page

Blood glucose is not a stock. It is a flow with a very tightly defended level, and almost everything here acts on one of four points in that flow.

Appearance from the gut. After a meal, starch and sugars are hydrolysed by salivary and pancreatic amylase and then by brush-border enzymes — maltase-glucoamylase, sucrase-isomaltase, collectively the α-glucosidases — before glucose crosses the enterocyte. The rate of that appearance, not the total amount, determines the height of the postprandial peak. Anything that slows gastric emptying, thickens the luminal contents, or inhibits α-glucosidase flattens the curve without changing the calories absorbed. This is where viscous fibre and acetic acid act, and it is the only mechanism in this document with an unambiguous human demonstration.

Insulin secretion. β cells sense the rising glucose, and release insulin in two phases. The first-phase spike is lost early in the progression to type 2 diabetes, which is why postprandial values deteriorate before fasting values do.

Insulin action. Insulin binds its receptor, triggers receptor autophosphorylation, and works through IRS proteins and PI3K to Akt, which among other things moves GLUT4 transporters to the membrane of muscle and adipose cells. Insulin resistance is a reduced glucose-disposal response to a given insulin concentration. The reference method for measuring it is a hyperinsulinaemic clamp — expensive, invasive, and used in almost none of the trials here. What is used instead is HOMA-IR, an index computed from a single fasting glucose and fasting insulin. It is a legitimate epidemiological instrument and a weak substitute for a clamp, and the distinction matters when reading section 15.

Hepatic output. Between meals the liver supplies glucose by glycogenolysis and gluconeogenesis. Failure to suppress that output after eating is a major contributor to hyperglycaemia in type 2 diabetes, and it is the principal site of metformin's action. Several compounds here are proposed to work here, mostly via AMP-activated protein kinase.

Two integrating measurements sit on top of all four. Glycated haemoglobin reflects the non-enzymatic glycation of haemoglobin over the roughly 8-to-12-week lifespan of the circulating red-cell population, so it is a weighted average of ambient glucose and cannot be moved quickly. And continuous glucose monitoring reports the interstitial glucose concentration every few minutes, which is a far richer signal than any of the above and, as section 07 shows, a signal most people have no idea how to interpret.

03 The outcome ladder

Six different things are called "improving blood sugar", and they are not degrees of one another.

A flatter curve after one meal. Measured as the incremental area under a two-hour glucose curve, in a dozen volunteers, on one morning. Cheap, reproducible, and mechanistically informative. It is a statement about the rate of glucose appearance from that meal and nothing else.

Fasting plasma glucose. A single number after an overnight fast. Sensitive to hepatic glucose output, to the last few days of diet, and to measurement noise. It moves within weeks.

A fasting-insulin resistance index. HOMA-IR and its relatives. Derived, not measured, and driven mostly by the insulin term, which has substantial assay and biological variability.

Glycated haemoglobin. A months-long average, the accepted regulatory surrogate, and the endpoint on which drug approvals rest. Moving it by half a percentage point is a real pharmacological achievement.

Incident diabetes. Whether a person with prediabetes crosses a diagnostic threshold. This requires thousands of participants followed for years. Exactly one programme reviewed here reached it, and it was not a supplement (Knowler and colleagues, 2002; Diabetes Prevention Program Research Group, 2015).

Complications and death. Retinopathy, nephropathy, neuropathy, myocardial infarction, stroke, amputation, mortality. These are the outcomes that matter, they are almost never studied in this field, and — as section 06 shows — even large trials of licensed glucose-lowering drugs have had difficulty demonstrating differences between agents on them (GRADE Study Research Group, 2022).

The claim ladder therefore runs from "my sensor was flatter after lunch" to "I did not develop diabetic kidney disease", and a product may sit five rungs below where its packaging implies. Figure 1 sets out which rung each compound here actually reached.

04 Three populations that are not one population

Almost every dispute in this field dissolves once the population is named, and one paper here states the principle so cleanly that it functions as a law.

Gibb, Sivakumar, Sloan, and colleagues assembled 35 randomised controlled trials of psyllium spanning three decades and three continents, and analysed them in eight meta-analyses stratified by baseline glycaemic status. In patients treated for type 2 diabetes, multi-week dosing before meals improved fasting blood glucose by 37.0 mg/dL (p < 0.001) and glycated haemoglobin by 0.97 percentage points, or 10.6 mmol/mol (p = 0.048). In people with prediabetes the improvement was modest. In euglycaemic subjects there was no significant glucose lowering at all. The authors' own summary is the title of the paper: the benefit was proportional to the loss of glycaemic control (Gibb et al., 2015).

That is established, and it generalises. Read it as three separate questions that happen to share a vocabulary.

People with type 2 diabetes. Elevated fasting and postprandial glucose, impaired first-phase secretion, insulin resistance, and considerable room to improve on every scale. Most of the positive trials here were conducted here.

People with prediabetes or impaired glucose tolerance. Intermediate physiology, intermediate effects, and the population in which prevention is a coherent goal (Jayedi et al., 2024).

People with normal glucose. No deficit to correct, a level defended within a remarkably narrow band, and — as the psyllium analysis and the resveratrol syntheses both show — essentially nothing to gain on any of these endpoints (Gibb et al., 2015; Liu et al., 2014). This is who buys most of the products.

Transferring a result across these three groups is the single most common error in commercial metabolic claims. The population is named in every reporting sentence.

05 Why this literature argues with itself

Four structural facts explain more of the disagreement in this field than any biological hypothesis, and all four are visible in the abstracts of the syntheses cited here.

Heterogeneity that is reported and then ignored. Allen, Bina, Chase, and colleagues pooled 10 randomised trials of cinnamon in type 2 diabetes and found high heterogeneity — I² between 66.5% and 94.72% — for every outcome except HDL cholesterol, and said plainly that this "may limit the ability to apply these results to patient care" (Allen et al., 2013). De Moura, and colleagues, pooling 28 trials in 3,054 patients a decade later, reported I² of 88% for fasting glucose, 94% for glycated haemoglobin, and 100% for postprandial glucose (de Moura et al., 2025). An I² of 100% means the trials are not estimating a common effect. A pooled mean under those conditions is an arithmetic operation, not a finding.

A quality-effect gradient. Balk, Tatsioni, Lichtenstein, and colleagues reviewed 41 randomised trials of chromium and noted that almost half were of poor quality — and that "larger effects were more commonly observed in poor-quality studies" (Balk et al., 2007). That sentence is the most useful single line in the chromium literature, and the same gradient is visible for berberine, where the largest and most consistent effects come from the trial set whose methodological quality was described as generally low (Dong et al., 2012).

Small samples, narrow provenance. Many of the trials here enrolled fewer than 50 people. Several of the most-cited berberine trials come from a small number of groups searching a database set that includes four Chinese-language registries (Guo et al., 2021). Neither fact makes a result wrong. Both mean that a single pooled estimate carries less independent information than its confidence interval implies.

Surrogate cheapness. Because fasting glucose costs almost nothing to measure and glycated haemoglobin only slightly more, a trial can report six glycaemic outcomes and four lipid outcomes at two timepoints. Without a pre-specified primary endpoint, the probability that something reaches conventional significance approaches certainty — which is why so many abstracts in this field report a significant effect on one member of a family of correlated measures and a null on the rest.

None of this makes the field worthless. It means that when a compound's effect appears in the outcome that is easiest to move, in the population with the most room to improve, in the smallest and least well-conducted trials, the correct reading is scepticism rather than averaging.

06 What already works, and why that sets the bar

No compound here can be evaluated without knowing what a serious intervention achieves, because the interesting question is never "does it do anything" but "does it do anything worth doing".

Prevention. Knowler and colleagues randomised 3,234 non-diabetic people with elevated fasting and post-load glucose to placebo, metformin at 850 mg twice daily, or a lifestyle programme targeting at least 7% weight loss and at least 150 minutes of physical activity per week. Over a mean 2.8 years, diabetes incidence was 11.0, 7.8, and 4.8 cases per 100 person-years respectively. The lifestyle intervention reduced incidence by 58% (95% CI 48 to 66) and metformin by 31% (95% CI 17 to 43); 6.9 people had to take part in the lifestyle programme, or 13.9 receive metformin, to prevent one case over three years (Knowler et al., 2002). This is established and it is the reference standard for the whole field.

Prevention, followed long enough to be disappointing. The Diabetes Prevention Program Research Group followed 2,776 of the surviving cohort for a mean of 15 years. Diabetes incidence remained reduced — 27% by lifestyle (HR 0.73, 95% CI 0.65 to 0.83) and 18% by metformin (HR 0.82, 95% CI 0.72 to 0.93) — but the between-group differences declined over time, and by year 15 the cumulative incidence of diabetes was 55% in the lifestyle group, 56% in the metformin group, and 62% in the placebo group. The aggregate microvascular outcome did not differ significantly between treatment groups in the whole cohort: 12.4% on placebo, 13.0% on metformin, 11.3% on lifestyle. In women it did favour lifestyle (8.7% against 11.0% on placebo, a 21% reduction, p = 0.03). And the observation that carries the most weight here is the non-randomised one: participants who did not develop diabetes, however that happened, had a 28% lower prevalence of microvascular complications (RR 0.72, 95% CI 0.63 to 0.83) (Diabetes Prevention Program Research Group, 2015).

Read those two paragraphs together. The most intensive lifestyle programme ever mounted for diabetes prevention delayed the disease in a majority of a very high-risk cohort rather than preventing it, and did not produce a clear reduction in microvascular disease overall across 15 years. That is the bar. A supplement that lowers fasting glucose by a fraction of a millimole is not near it.

How much weight loss. Jayedi and colleagues pooled 44 randomised trials in 14,742 people with prediabetes, with mean weight loss between 1% and 9% over a median 24 months. Lifestyle weight-loss interventions increased regression to normoglycaemia by 11 per 100 participants (RR 1.51, 95% CI 1.27 to 1.80; 20 trials; GRADE moderate) and reduced progression to type 2 diabetes by 8 per 100 (RR 0.59, 95% CI 0.51 to 0.67; 37 trials; GRADE moderate), with a linear dose-response across the whole 1-to-9% range and no significant difference between diet, exercise, or both (Jayedi et al., 2024). Established. Note what that linearity means: there is no threshold below which weight loss stops helping, and no supplement in this document has been shown to produce weight loss of that magnitude.

Treatment, and the humbling of surrogates. The GRADE Study Research Group randomised 5,047 people with type 2 diabetes on metformin to added insulin glargine, glimepiride, liraglutide, or sitagliptin and followed them a mean of 5.0 years. There were no material differences between the four agents in the development of hypertension or dyslipidaemia or in microvascular outcomes; the overall rates per 100 participant-years were 2.6 for moderately increased albuminuria, 1.1 for severely increased albuminuria, 2.9 for renal impairment, and 16.7 for diabetic peripheral neuropathy. The groups did not differ in major adverse cardiovascular events (overall rate 1.0), heart-failure hospitalisation (0.4), cardiovascular death (0.3), or all-cause death (0.6). Only liraglutide showed a lower hazard for any cardiovascular disease (HR 0.7, 95% CI 0.6 to 0.9) (GRADE Study Research Group, 2022).

That trial is the most important context here and the least likely to appear in any commercial discussion of it. Four licensed drugs, all achieving glycaemic targets, differed from one another hardly at all on the outcomes patients care about, over five years, in five thousand people. Anyone who believes a botanical's half-point movement in glycated haemoglobin translates into clinical benefit is asserting something that a properly powered trial of real drugs could not demonstrate.

What the professional bodies say. The American Diabetes Association Professional Practice Committee sets the diagnostic thresholds and the pharmacologic algorithm annually (American Diabetes Association Professional Practice Committee, 2026a; 2026b). Samson and colleagues, updating the American Association of Clinical Endocrinology algorithm across 11 sections, place lifestyle modification and the treatment of overweight and obesity as the key pillars for both prediabetes and type 2 diabetes, and frame drug choice around comorbidities and complications rather than glucose alone (Samson et al., 2026). Neither body positions any compound here as therapy. The National Center for Complementary and Integrative Health's assessment of diabetes and dietary supplements reaches the same place from the other direction (NCCIH, 2026a), and the World Health Organization's fact sheet supplies the scale of the problem the market is addressing (WHO, 2026).

07 Measuring glucose in people who do not have diabetes

A large and growing share of the demand for these products comes from people wearing continuous glucose monitors, who have discovered that eating raises blood glucose. Two records establish what that observation is worth.

Shah, DuBose, Li, and colleagues fitted blinded current-generation sensors to 153 healthy, non-obese, non-diabetic people aged 7 to 80 for up to 10 days. Mean glucose was 98 to 99 mg/dL in every age group except those over 60, in whom it was 104 mg/dL. Median time between 70 and 140 mg/dL was 96% (IQR 93 to 98). The mean within-person coefficient of variation was 17 ± 3%. Median time above 140 mg/dL was 2.1% — about 30 minutes a day — and median time below 70 mg/dL was 1.1%, about 15 minutes a day (Shah et al., 2019).

Two things follow. Healthy people spend about half an hour a day above 140 mg/dL and a quarter of an hour below 70, and both are normal. And a 17% coefficient of variation means the same person, eating the same food, will produce visibly different curves — so a single flattened or spiked trace is not evidence about anything.

Liao and colleagues then asked whether wearing the sensor helps. Across 23 studies in 1,074 non-diabetic participants from 11 countries, continuous glucose monitoring improved mean blood glucose relative to controls (SMD −0.54, 95% CI −1.02 to −0.07; p = 0.03) but did not change body-mass index (SMD −0.25, 95% CI −0.63 to 0.12; p = 0.19). The stratified finding is the one that matters: monitoring improved glycaemic control in people with prediabetes, and produced no appreciable glycaemic benefit in healthy normoglycaemic people. The authors conclude that it should be positioned as a precision biofeedback tool inside a structured lifestyle programme, not as an intervention in itself (Liao et al., 2026).

Grade: strongly supported that normal non-diabetic glucose excursions are wider than consumers assume, and that sensor use confers no glycaemic benefit in normoglycaemic people. The practical consequence here is that the postprandial spike a healthy reader observes and wants to suppress is, in most cases, a physiological event with no demonstrated pathological significance, and the supplement bought to suppress it is being aimed at a number rather than a disease.

08 Vinegar, and the cleanest mechanism in the file

Acetic acid (CH3COOH; PubChem, 2026d) has the best-characterised acute glycaemic effect of anything here, and its evidence base is a useful template for what an honest small finding looks like.

Östman, Granfeldt, Persson, and colleagues gave 12 healthy volunteers white wheat bread containing 50 g of available carbohydrate with three levels of vinegar — 18, 23, and 28 mmol acetic acid — and a no-vinegar reference. A significant dose-response appeared at 30 minutes for both glucose and insulin: the higher the acetic acid, the lower the response. The highest level lowered glucose at 30 and 45 minutes and insulin at 15 and 30 minutes, and raised satiety ratings at 30, 90, and 120 minutes. Notably, when the glycaemic index was recalculated over 120 minutes rather than 90, the difference from the reference meal disappeared (Östman et al., 2005).

Johnston, Buller, and colleagues found that adding vinegar or peanut products to a test meal reduced the 60-minute glucose response by approximately 55%, but that the reduction was significant only for the high-glycaemic-load meal; the effect on later energy intake was weak and did not reach significance (p = 0.111) (Johnston et al., 2005).

Liatis, Sarantopoulou, Tsiakou, and colleagues then tested the boundary directly in 16 people with type 2 diabetes. With a high-glycaemic-index meal of mashed potato and low-fat milk, adding vinegar reduced the incremental glucose area over two hours from 311 ± 124 to 181 ± 78 mmol·min/L (p = 0.04), with insulin reduction of marginal significance (p = 0.056). With an isocaloric low-glycaemic-index meal of whole-grain bread, lettuce, and low-fat cheese, vinegar did nothing at all: 229 ± 38 against 238 ± 25 mmol·min/L (p = 0.56) (Liatis et al., 2010).

That is a mechanism with a stated domain of validity, which is rare enough to be worth stating plainly: vinegar flattens the peak of a rapidly digested carbohydrate meal, and has nothing to flatten when the meal is already slowly digested.

Chronic dosing has been tested once at reasonable length. Jasbi and colleagues randomised 45 adults at increased metabolic risk by waist circumference to eight weeks of daily red-wine vinegar or control. Fasting glucose fell (p = 0.003) and fasting insulin fell (p < 0.001); insulin resistance decreased 8.3% in the vinegar group and increased 9.7% in controls (p < 0.001). There were no between-group differences in body-mass index, weight, waist circumference, or visceral fat measured by dual-energy X-ray absorptiometry (Jasbi et al., 2019).

Grades. The acute postprandial effect on high-glycaemic-index meals is strongly supported. The eight-week improvement in fasting glucose homeostasis is emerging, from one trial of 45 people. The claim that vinegar reduces body fat is not supported, and the one study designed to detect it found the opposite of what is advertised. No trial here tested vinegar against glycated haemoglobin as a primary endpoint. Note also that Kumar and colleagues' network meta-analysis, discussed in section 21, ranked apple cider vinegar first among six herbs for fasting glucose reduction on the strength of numbers that cannot be right.

09 Viscous fibre

Fibre is the other place where the mechanism is unambiguous, and it is where the population law of section 04 was first stated cleanly.

Psyllium is a soluble, gel-forming, non-fermented fibre — a distinction that matters, because the microbiome-mediated arguments for fermentable fibre are a different mechanism with a different evidence base, treated in the companion Microbiome-Nutrition title. Psyllium's glycaemic action is physical: it raises the viscosity of the luminal contents, slowing the rate at which glucose reaches the absorptive surface.

Gibb, Sivakumar, Sloan, and colleagues' stratified analysis of 35 trials is reported in full in section 04. The essential numbers: fasting glucose −37.0 mg/dL and glycated haemoglobin −0.97 percentage points in people treated for type 2 diabetes; a modest improvement in prediabetes; no significant lowering in euglycaemic subjects (Gibb et al., 2015). Grade: strongly supported, with the population restriction as part of the finding rather than a caveat.

The authors added a sentence that section 22 will return to: because the largest effect was seen in patients being treated for type 2 diabetes, further work is needed on how to incorporate psyllium into existing treatment algorithms alongside concomitant hypoglycaemic medications (Gibb et al., 2015). An agent that lowers glucose by 37 mg/dL in people already taking glucose-lowering drugs is not a neutral addition.

Compound	Best-supported outcome	Effect where reported	Population in which it appeared	Effect in normoglycaemic people
Acetic acid (vinegar)	postprandial glucose area	iAUC 311 → 181 mmol·min/L, high-GI meal	healthy volunteers; type 2 diabetes	present acutely; no HbA1c data
Psyllium	fasting glucose; HbA1c	−37.0 mg/dL; −0.97 points	treated type 2 diabetes	none detected
Berberine	HbA1c	−0.57 to −0.73 points	type 2 diabetes	not studied
Cinnamon	fasting glucose	inconsistent across syntheses	type 2 diabetes	not studied
Fenugreek	HbA1c	−0.88 points, direction only	type 2 diabetes and prediabetes	not studied
Alpha-lipoic acid	insulin-stimulated glucose disposal	+27% clamp MCR	type 2 diabetes	not studied
Magnesium	HOMA-IR	−0.67	diabetes and at-risk	not studied
Chromium	HbA1c	−0.55 to −0.6 points	type 2 diabetes	explicitly none
Resveratrol	HOMA-IR; HbA1c	−0.34; −0.64 (SMD)	type 2 diabetes	explicitly none
Polyphenols, whole food	blood pressure, not glucose	SBP −3.69 mmHg	cardiometabolic risk	no glucose effect

Table 1. Outcome by population. The right-hand column is the one a consumer needs and the one almost nobody measured. Where it was measured — psyllium, chromium, resveratrol — the answer was no effect.

10 Berberine, and the largest positive file in the category

Berberine (C20H18NO4+; PubChem, 2026a) is a quaternary benzylisoquinoline alkaloid from Coptis chinensis and related plants, and it has the strongest and most consistently positive body of glycaemic evidence of any compound here. It is also the compound whose evidence base most rewards careful reading.

Start with the primary trials, because they are unusually informative.

Yin, Xing, and Ye conducted two studies. In study A, 36 adults with newly diagnosed type 2 diabetes were randomised to berberine or metformin, both at 0.5 g three times daily, for three months. In the berberine arm, glycated haemoglobin fell from 9.5 ± 0.5% to 7.5 ± 0.4% (p < 0.01), fasting glucose from 10.6 ± 0.9 to 6.9 ± 0.5 mmol/L (p < 0.01), and postprandial glucose from 19.8 ± 1.7 to 11.1 ± 0.9 mmol/L (p < 0.01). The hypoglycaemic effect was similar to metformin's. In study B, 48 adults with poorly controlled diabetes received berberine in addition to existing therapy; glycated haemoglobin fell from 8.1 ± 0.2% to 7.3 ± 0.3% (p < 0.001), with fasting insulin and HOMA-IR reduced by 28.1% and 44.7%. Twenty of the 58 patients, 34.5%, experienced transient gastrointestinal adverse effects (Yin et al., 2008).

Zhang, Wei, Xue, and colleagues pursued the mechanism, showing that berberine increased insulin-receptor messenger RNA and protein expression across several human cell lines, with increased insulin-stimulated phosphorylation of the receptor β-subunit and Akt; in patients, the proportion of peripheral blood lymphocytes expressing the insulin receptor rose after treatment, and fasting glucose and glycated haemoglobin fell with efficacy the authors described as similar to metformin and rosiglitazone (Zhang et al., 2010).

Those are real findings, and the receptor-expression mechanism is a genuinely different one from metformin's.

Now the syntheses. Dong, Wang, and Zhao pooled 14 randomised trials in 1,068 participants and reported that methodological quality was generally low. Against lifestyle modification with or without placebo, berberine improved glycaemia and dyslipidaemia. Against oral hypoglycaemic drugs — metformin, glipizide, or rosiglitazone — berberine did not demonstrate significantly better glycaemic control, though it showed a mild antidyslipidaemic effect. Combined with those same drugs, it produced better control than the drugs alone. No serious adverse effects were reported, and the authors ended by warning that the evidence should be carefully interpreted given low methodological quality, small samples, few trials, and unidentified risks of bias (Dong et al., 2012).

Guo and colleagues assembled 46 trials from eight databases, four of them Chinese-language, and reported glycated haemoglobin −0.73 percentage points (95% CI −0.97 to −0.51), fasting glucose −0.86 mmol/L (95% CI −1.10 to −0.62), two-hour postprandial glucose −1.26 mmol/L (95% CI −1.64 to −0.89), fasting insulin −2.05, HOMA-IR −0.71 (95% CI −1.03 to −0.39), and body-mass index −1.07, together with improvements in triglycerides, total cholesterol, LDL, HDL, and inflammatory markers. Their conclusion was that there is strong evidence for efficacy and safety, especially as adjunctive therapy (Guo et al., 2021).

Nazari and colleagues then ran an umbrella meta-analysis of the meta-analyses, motivated explicitly by inconsistency between them, and found fasting glucose −0.77 (95% CI −0.90 to −0.63), glycated haemoglobin −0.57 (95% CI −0.68 to −0.46), HOMA-IR −1.04 (95% CI −1.66 to −0.42), insulin −1.00, and reductions in interleukin-6, tumour necrosis factor α, and C-reactive protein (Nazari et al., 2024).

And most recently Shadin and colleagues pooled more than 30 trials in over 2,000 participants covering probiotics, synbiotics, and berberine, and reported modest pooled reductions — fasting glucose −0.71 mmol/L and glycated haemoglobin −0.19 percentage points — with substantial between-study variability. Probiotics alone produced approximately −0.80 mmol/L and −0.21 points. An exploratory molecular-dynamics analysis found weaker α-glucosidase binding for berberine than for acarbose, the reference inhibitor. Their conclusion positions all three as adjunctive rather than primary options and calls for larger, longer, standardised trials (Shadin et al., 2026).

Look at what happened to the glycated-haemoglobin estimate across those four syntheses: −0.73, then −0.57, then −0.19. The direction is stable; the magnitude is not, and it shrank as the evidence base widened and as pooling became more conservative.

Grade: that berberine lowers glucose and glycated haemoglobin in type 2 diabetes is strongly supported. That the magnitude is comparable to metformin, as the two early trials suggested, is emerging at best and rests on trials of 36 and 48 people with the methodological profile Dong and colleagues described. That it does anything measurable in people with normal glucose is not studied. The dedicated Berberine article carries the compound-level detail.

11 Berberine's bioavailability, and what it forces us to conclude

There is one fact about berberine that belongs in every discussion of it and appears in almost none.

Cui and colleagues, reviewing the pharmacology and bioavailability, state that poor solubility, low membrane permeability, P-glycoprotein-mediated efflux, and hepatic and intestinal metabolism give oral berberine a bioavailability of less than 1%, and that improving it is a prerequisite for clinical application. Their review is structured around the strategies — absorption enhancers, P-glycoprotein inhibitors, structural modification, salts and cocrystals, novel formulations — needed to overcome that limit (Cui et al., 2024).

This creates a genuine scientific tension rather than a debunking, and it is worth stating both halves.

On one side, a compound with sub-1% oral bioavailability producing a systemic, insulin-receptor-mediated, metformin-comparable effect requires either extraordinary potency at the target or an explanation that does not depend on systemic exposure. The most parsimonious candidates are luminal and hepatic-first-pass actions: berberine's effects on the gut microbiota, on intestinal glucose handling, and on the liver, which sees a far higher concentration than the peripheral circulation. Shadin and colleagues' pairing of berberine with probiotics and synbiotics in a single analysis is a nod in that direction, and their finding that berberine binds α-glucosidase more weakly than acarbose argues against luminal enzyme inhibition being the whole story (Shadin et al., 2026).

On the other side, the P-glycoprotein and hepatic-metabolism findings are exactly the profile of a compound with a high potential for pharmacokinetic drug interaction. A substance that is a P-glycoprotein substrate and is extensively metabolised in liver and intestine is a substance whose co-administration with prescription drugs should not be assumed to be inert (Cui et al., 2024). Section 22 returns to this.

Grade: the sub-1% bioavailability figure is established. The mechanism by which a poorly absorbed alkaloid produces a systemic glycaemic effect is unresolved, and any confident statement about how berberine works — including the AMP-kinase story that dominates consumer material — is running ahead of the human data.

12 Cinnamon, and three questions that are usually merged

Cinnamon is the clearest example here of a literature whose disagreement is not resolved by adding trials. Three separate questions get merged into one claim, and they have different answers. The constituent usually credited with the glycaemic effect is cinnamaldehyde (C9H8O; PubChem, 2026c), which is also the compound responsible for the flavour, so a dose expressed in grams of bark says little about the exposure delivered.

Does cinnamon lower glycated haemoglobin? The syntheses do not agree, and the disagreement is not chronological.

Allen, Bina, Chase, and colleagues pooled 10 trials in 543 patients with type 2 diabetes, doses from 120 mg/day to 6 g/day for 4 to 18 weeks. Fasting plasma glucose fell by 24.59 mg/dL (95% CI −40.52 to −8.67), total cholesterol by 15.60 mg/dL, LDL by 9.42 mg/dL, and triglycerides by 29.59 mg/dL, with HDL rising 1.66 mg/dL. Glycated haemoglobin did not change significantly: −0.16 percentage points (95% CI −0.39 to 0.02). Heterogeneity was high for every outcome except HDL, I² ranging from 66.5% to 94.72% (Allen et al., 2013).

Yu and colleagues, running a dose-response meta-analysis a decade later, found no significant effect on fasting blood glucose, glycated haemoglobin, or total cholesterol, and significant effects only on triglycerides (−7.31), LDL (−6.78), and HDL (+1.53), with non-linear dose relationships for the two lipid outcomes. A subgroup at doses of 1,200 mg or less did show a hypoglycaemic effect (−11.1) (Yu et al., 2023).

De Moura and colleagues, pooling 28 trials in 3,054 patients with GRADE assessment, reported fasting glucose −15.26 mg/dL, postprandial glucose −39.22 mg/dL, glycated haemoglobin −0.56 percentage points (95% CI −0.99 to −0.13), and HOMA-IR −0.76. Heterogeneity was 88%, 100%, 94%, and 22% respectively, and only doses of 2 g/day or less reduced glycated haemoglobin (de Moura et al., 2025).

Jafari and colleagues, in a GRADE-assessed dose-response analysis of 49 studies of cardiovascular risk factors, reported standardised mean differences of −1.28 for fasting glucose, −0.71 for glycated haemoglobin, −2.28 for postprandial glucose, −0.54 for HOMA-IR, along with reductions in blood pressure, waist circumference, and every lipid fraction, and a rise in HDL (Jafari et al., 2025).

Now read the sequence honestly. The two analyses that specifically examined dose-response found the glycaemic effect either absent or confined to a low-dose subgroup. The two that found it also reported heterogeneity of 94% and 100% for the outcomes in question. And Jafari and colleagues' result — a compound simultaneously reducing systolic and diastolic pressure, waist circumference, every lipid fraction, all glycaemic indices, C-reactive protein, and malondialdehyde, while raising HDL — is a pattern that should provoke suspicion rather than enthusiasm. A spice that improves every measured cardiometabolic parameter is more likely describing a literature with a systematic problem than a panacea.

There is one older, narrower data point that cuts against the whole set. Suksomboon, Poolsup, and colleagues restricted their meta-analysis to placebo-controlled trials of a single herb, of at least eight weeks, reporting glycated haemoglobin. Of the four herbs meeting that standard, Ipomoea batatas (−0.30 points), Silybum marianum (−1.92), and Trigonella foenum-graecum (−1.13) improved control, while Cinnamomum cassia did not (Suksomboon et al., 2011).

Grade: that cinnamon lowers fasting glucose in type 2 diabetes is emerging, with heterogeneity that undermines the pooled estimate. That it lowers glycated haemoglobin is inconsistent — genuinely unresolved, not merely unproven, with high-quality dose-response analyses on both sides. That it does anything in people without diabetes is not studied. The Cinnamon article holds the botanical detail.

13 Cinnamon's other question: coumarin, species, and dose

The second question about cinnamon has nothing to do with efficacy, and it is the one where the evidence is strongest.

Most cinnamon sold as a spice and used in the trials above is Cinnamomum cassia, not C. verum (Ceylon cinnamon). Cassia bark contains coumarin, and coumarin is hepatotoxic in a susceptible human subgroup.

Abraham and colleagues reviewed the toxicology and derived a tolerable daily intake of 0.1 mg per kg body weight from human clinical data — patients treated with coumarin as a medicinal drug — confirming a value the European Food Safety Authority had reached from animal hepatotoxicity data. Critically, the human data revealed a subgroup more susceptible to hepatotoxicity than any animal species tested, and the cause of that susceptibility is unknown. Cassia cinnamon is the main dietary source, and a telephone survey of more than 1,000 randomly selected people in Germany indicated that heavy consumers during the Christmas baking season may reach a daily coumarin intake corresponding to the entire tolerable daily intake (Abraham et al., 2010).

Grade: established that a tolerable daily intake exists, that it was derived partly from human hepatotoxicity, that a hypersusceptible human subgroup exists and is unexplained, and that ordinary culinary cassia consumption can approach the limit.

Set that beside section 12. The trials reporting benefit used doses up to 6 g/day of cinnamon for up to 18 weeks (Allen et al., 2013), and the two analyses that examined dose found the glycaemic effect confined to low doses — 1,200 mg or less, or 2 g or less (Yu et al., 2023; de Moura et al., 2025). A supplement regimen at the upper end of the trialled range, taken daily and indefinitely, with an unspecified species and no coumarin assay, is a coumarin exposure of unknown size layered on top of dietary intake. NCCIH's summary of cinnamon's usefulness and safety states the position for a general readership (NCCIH, 2026b).

This is the single most concrete safety finding in the article, and it is a species-and-authentication problem rather than a pharmacological one. Nothing about it is addressed by a label that reads "cinnamon extract".

14 Fenugreek

Trigonella foenum-graecum seed has a plausible mechanism — it is rich in galactomannan, a viscous soluble fibre, so part of any effect should resemble section 09 — and two recent meta-analyses that reach compatible conclusions by incompatible arithmetic.

Kim and colleagues pooled 10 trials in 706 participants with type 2 diabetes or prediabetes, drawing on Korean and Chinese as well as English-language databases. Fenugreek significantly reduced fasting blood glucose, two-hour post-load glucose, and glycated haemoglobin, but did not significantly reduce HOMA-IR. Total cholesterol, triglycerides, and HDL improved; LDL and body-mass index did not. No hepatic or renal toxicity was observed, and adverse events were confined to mild gastrointestinal effects in some studies (Kim et al., 2023).

Shabil and colleagues pooled 14 trials in 894 participants and reported glycated haemoglobin −0.88 percentage points (95% CI −1.49 to −0.27; p = 0.00). Their fasting-glucose and postprandial figures are a different matter, discussed in section 21, because as printed they do not support the sentence built on them. The authors' own conclusion is appropriately hedged: the quality and heterogeneity of the included studies remain a concern, and rigorous double-blind randomised trials are needed to understand fenugreek's true potential (Shabil et al., 2023).

The oldest and most restrictive analysis supports the direction. Suksomboon, Poolsup, and colleagues' single-herb, eight-week-minimum, glycated-haemoglobin-reporting criteria admitted Trigonella foenum-graecum with a pooled difference of −1.13 percentage points (95% CI −0.11 to −2.14; p = 0.03), while explicitly noting high heterogeneity for that estimate (Suksomboon et al., 2011).

Grade: that fenugreek lowers glycated haemoglobin in people with type 2 diabetes or prediabetes is emerging, consistent in direction across three syntheses and unreliable in magnitude in all three. That the effect is distinguishable from the effect of an equivalent dose of viscous fibre is not established, because no trial in this set included a fibre-matched comparator. Note also that fenugreek appears in the hormonal and male-vitality literature for entirely different claims; the Fenugreek article and the companion Hormonal-Sexual-and-Male-Vitality title separate them.

15 Alpha-lipoic acid: two literatures wearing one name

Alpha-lipoic acid (C8H14O2S2; PubChem, 2026b) is the compound here whose reputation and evidence are most badly misaligned, because its strong evidence is for something other than glycaemic control.

The glycaemic literature is weak. Jacob and colleagues randomised 74 patients with type 2 diabetes to placebo or oral alpha-lipoic acid at 600, 1,200, or 1,800 mg daily for four weeks, with an isoglycaemic glucose clamp before and after — the reference method. Significantly more subjects on active treatment showed increased insulin-stimulated glucose disposal in each dose group; pooling all active arms gave a metabolic clearance rate 27% higher than placebo (p < 0.01). There was no dose effect across a threefold dose range, which the authors handled by combining the arms, and they described their own study as explorative and in need of substantiation (Jacob et al., 1999).

De Oliveira and colleagues then ran a four-month, four-arm, double-blind, placebo-controlled trial in 102 patients with type 2 diabetes: 600 mg lipoic acid, 800 mg α-tocopherol, both, or placebo. Lipid fractions and the HOMA index improved in the lipoic-acid group, but not significantly, and the authors concluded that supplementation alone or in combination did not affect the lipid profile or insulin sensitivity (de Oliveira et al., 2011).

Gosselin and colleagues gave 600 mg daily for 30 days to 12 prediabetic, dyslipidaemic, overweight or obese adults in a randomised placebo-controlled crossover pilot. Serum glucose did not change. Fasting insulin fell (p = 0.04) and HOMA-IR moved in the same direction without reaching conventional significance (p = 0.07). No lipid fraction changed (Gosselin et al., 2019).

Three trials, four weeks to four months, 12 to 102 participants: one positive on a clamp with no dose-response, one null, one equivocal in twelve people. Grade: plausible that oral alpha-lipoic acid improves insulin sensitivity in type 2 diabetes; not supported as a glucose-lowering agent; not studied in people with normal glucose.

The neuropathy literature is strong, and it is a different intervention. Ziegler and colleagues meta-analysed four randomised, double-blind, placebo-controlled parallel-group trials — ALADIN I, ALADIN III, SYDNEY, and NATHAN II — comprising 1,258 patients (716 on active, 542 on placebo) who received 600 mg intravenously per day for three weeks. The relative difference in Total Symptom Score favouring alpha-lipoic acid was 24.1% (95% CI 13.5 to 33.4) and in the lower-limb Neuropathy Impairment Score 16.0% (95% CI 5.7 to 25.2). Responder rates, defined as at least 50% improvement in symptom score, were 52.7% against 36.9% (p < 0.05). Pain, burning, and numbness all improved, as did pin-prick and touch-pressure sensation and ankle reflexes. Adverse-event rates did not differ (Ziegler et al., 2004a).

Ziegler's critical review of the same programme adds the oral data: treatment for four to seven months tends to reduce neuropathic deficits and improve cardiac autonomic neuropathy, and the intravenous meta-analysis represents the largest sample of diabetic patients ever treated with a single drug class for neuropathic symptoms, at the highest level of evidence (Ziegler et al., 2004b).

Grade: established that 600 mg/day intravenously for three weeks reduces symptomatic diabetic polyneuropathy to a clinically meaningful degree. That finding is about a three-week hospital infusion in people who already have nerve damage. It is not evidence that an oral capsule improves blood sugar, and the transfer between the two is the most common misuse of this compound's literature. The Alpha-Lipoic-Acid article carries both files in full.

16 Magnesium: repletion, not pharmacology

Magnesium is a cofactor for hundreds of enzymes, including those of glucose phosphorylation and the insulin-signalling cascade, and low magnesium status is common in type 2 diabetes. The mechanism is not in question; the size and nature of the supplementation effect is.

Simental-Mendía and colleagues pooled randomised trials in both diabetic and non-diabetic individuals across 22 treatment arms for glucose, 14 for glycated haemoglobin, 12 for insulin, and 10 for HOMA-IR. The result is unusually clean in its pattern: HOMA-IR improved (WMD −0.67, 95% CI −1.20 to −0.14; p = 0.013), while plasma glucose (−0.20 mmol/L; p = 0.119), glycated haemoglobin (p = 0.756), and insulin (p = 0.556) did not. A subgroup analysis showed that trials of four months or longer produced significant improvement in both fasting glucose and HOMA-IR where shorter trials did not (Simental-Mendía et al., 2016).

Veronese and colleagues, restricting to double-blind randomised trials in people with diabetes or at high risk of it, found that magnesium reduced fasting plasma glucose in people with diabetes, improved fasting and two-hour post-load glucose in people at high risk, and improved insulin-sensitivity markers (Veronese et al., 2021).

Ye and colleagues supply a different population and an instructive caution. Pooling 11 randomised trials in 618 women with polycystic ovary syndrome across magnesium, chromium, zinc, and selenium, they reported fasting glucose SMD −0.34, fasting insulin −0.72, HOMA-IR −0.75, total cholesterol −0.35, and triglycerides −0.58 — and also a small but statistically significant reduction in HDL (SMD −0.19, p = 0.04), with no effect on LDL (p = 0.55). Their own summary is that effects on lipids were mixed (Ye et al., 2026). A pooled analysis that lumps four different minerals cannot attribute any of that to magnesium specifically, and the HDL finding is a reminder that "improves metabolic parameters" is not a single direction.

The NIH Office of Dietary Supplements fact sheet sets out requirements, deficiency, and the upper intake level for supplemental magnesium (NIH ODS, 2026b).

Grade: strongly supported that magnesium supplementation improves a fasting-insulin-derived resistance index in people with diabetes or at risk, with effects concentrated in trials of four months or more. Not supported that it meaningfully lowers glycated haemoglobin. The pattern — a derived index moves, the integrated months-long average does not — is exactly what section 03 warns about, and it is most consistent with correction of a deficit rather than a pharmacological action. The Magnesium article carries the full nutrient file, and the Cardiovascular-Nutraceuticals and Bone-Health titles carry its other roles.

17 Chromium: the quality-effect gradient in one compound

Chromium is often presented as this field's cautionary tale. The literature is more interesting than that, and its real lesson is methodological.

Balk, Tatsioni, Lichtenstein, and colleagues reviewed 41 randomised trials, of which almost half were of poor quality. Among participants with type 2 diabetes, chromium improved glycated haemoglobin by 0.6 percentage points (95% CI −0.9 to −0.2) and fasting glucose by 1.0 mmol/L (95% CI −1.4 to −0.5), but not lipids. There was no benefit in individuals without diabetes. There were indications of a dose effect and of differences between chromium formulations. And the sentence that matters most: "larger effects were more commonly observed in poor-quality studies." The authors listed poor study quality, heterogeneity in methodology and results, and the absence of any agreed way to assess chromium status as limitations, and concluded that further studies addressing those limitations are needed before definitive claims can be made (Balk et al., 2007).

Suksomboon, Poolsup, and Yuwanakorn pooled 25 randomised trials seven years later and found glycated haemoglobin −0.55 percentage points (95% CI −0.88 to −0.22; p = 0.001) and fasting glucose −1.15 mmol/L (95% CI −1.84 to −0.47; p = 0.001), with chromium monotherapy also improving triglycerides and HDL. Effects were most evident with chromium picolinate, at more than 200 µg daily, and in patients whose baseline control was inadequate. Adverse-event risk did not differ from placebo, and data on combination products were limited and inconclusive (Suksomboon et al., 2014).

Georgaki and colleagues reviewed the field again through January 2024. Doses ranged from 50 to 1,000 µg/day for two to six months, in the form of picolinate, chromium yeast, chloride, or nicotinate. Several trials reported decreases in fasting glucose, insulin, glycated haemoglobin, and HOMA-IR, mainly in longer interventions, with improvements in HDL, triglycerides, and total cholesterol and little effect on LDL. The authors emphasised inconsistent dosage, chromium form, formulation, and study duration as significant limitations, and called for well-designed high-quality research on both efficacy and the potential risks of trivalent chromium exposure through supplements (Georgaki et al., 2024).

Grade: emerging that chromium supplementation lowers glycated haemoglobin in people with type 2 diabetes and inadequate baseline control, consistent across three syntheses spanning 17 years, with the magnitude systematically inflated by poor-quality trials. Established, and stated explicitly by the largest review, that there is no benefit in people without diabetes (Balk et al., 2007). The NIH Office of Dietary Supplements fact sheet notes the absence of a validated status assessment, which is why no trial here could enrol deficient participants preferentially (NIH ODS, 2026a).

That last point is the reason chromium's evidence cannot be resolved. If the effect is repletion of a deficit — the pattern that section 04 predicts and that the no-benefit-without-diabetes finding supports — then a trial that cannot identify deficient participants is guaranteed to produce heterogeneous results forever.

18 Resveratrol and dietary polyphenols

Resveratrol is the single best illustration here of the population law, because three independent syntheses tested it in both populations and agreed.

Liu, Zhou, Wang, and colleagues pooled 11 randomised trials in 388 subjects. Resveratrol significantly reduced fasting glucose, insulin, glycated haemoglobin, and HOMA-IR in participants with diabetes, and had no significant effect on any glycaemic measure in non-diabetic participants — a null that survived subgroup and sensitivity analysis by body-mass index, study design, dose, duration, and Jadad quality score (Liu et al., 2014).

Delpino and colleagues reviewed 30 studies, of which almost 60% showed at least one significant diabetes-related effect, and meta-analysed the glycaemic outcomes: HOMA-IR −0.34 (95% CI −0.64 to −0.04; I² = 70%) and glycated haemoglobin −0.64 (95% CI −1.22 to −0.07; I² = 90%), with fasting glucose significant only in individuals with diabetes (−0.85; I² = 90%) (Delpino et al., 2022).

Zhang and colleagues pooled 15 randomised trials in 896 patients with type 2 diabetes and found HOMA-IR −0.99 (95% CI −1.61 to −0.38; p = 0.002) with no statistically significant effect on total cholesterol or triglycerides and no difference in adverse events (Zhang et al., 2021).

Grade: strongly supported that resveratrol improves fasting-insulin-derived resistance indices in type 2 diabetes; strongly supported that it does nothing measurable to glycaemia in people without diabetes, which is the population that buys it. The heterogeneity of 70% and 90% in the two pooled glycated-haemoglobin estimates means the magnitude is not established.

The wider polyphenol question has a more useful answer, and it is not about glucose. Kiyimba and colleagues meta-analysed 46 randomised trials in 2,494 adults at cardiometabolic risk, deliberately comparing whole polyphenol-rich foods against purified polyphenol extracts. Whole foods reduced systolic pressure by 3.69 mmHg (95% CI −4.24 to −3.15) and diastolic by 1.44 mmHg; purified extracts did not. Purified extracts produced the larger effect on waist circumference and were the arm responsible for the significant reductions in total cholesterol (−9.03 mg/dL) and triglycerides (−13.43 mg/dL). And neither whole food nor extract significantly affected fasting blood glucose, LDL, HDL, interleukin-6, or C-reactive protein. The authors flag high heterogeneity and risk of bias (Kiyimba et al., 2023).

That is a genuinely important negative here. The best available head-to-head comparison of food-form and extract-form polyphenols found no glycaemic effect from either. Whatever polyphenol-rich diets do for cardiometabolic risk — and the blood-pressure finding is real — it is not mediated by fasting glucose. The Flavonoids, Resveratrol, and Cardiovascular-Nutraceuticals titles carry those literatures; the existing Metabolic-Nutrition article carries the dietary-pattern evidence.

19 The other botanicals in the catalogue

Three more subjects appear often enough in the commercial category to require disposal, and each illustrates a different evidential failure mode.

Gymnema sylvestre. Gaytán Martínez and colleagues randomised 30 patients with impaired glucose tolerance to 300 mg twice daily or placebo. In the Gymnema group, two-hour glucose fell from 9.1 ± 1.2 to 7.8 ± 1.7 mmol/L (p = 0.003) and glycated haemoglobin from 5.8 ± 0.3% to 5.4 ± 0.4% (p = 0.025), the Matsuda index of insulin sensitivity rose from 1.8 ± 0.8 to 2.4 ± 1.2 (p = 0.008), and body weight, body-mass index, and LDL fell; 46.7% of patients reached normal glycated haemoglobin (Gaytán Martínez et al., 2021). Fifteen participants per arm. Grade: emerging, unreplicated, and a textbook case of section 05's small-sample problem — a genuinely promising result that would be irresponsible to treat as established.

Momordica charantia (bitter melon). This is the field's cleanest demonstration that adding trials to a bad literature does not fix it. Ooi, Yassin, and Hamid's Cochrane review found four randomised trials in 479 participants, only two published as full peer-reviewed papers, with generally high risk of bias. No meta-analysis was performed, because the preparations were so variable that no identical preparation was tested twice. There was no statistically significant difference in glycaemic control against placebo, and none against metformin or glibenclamide. No trial investigated death, morbidity, quality of life, or costs (Ooi et al., 2012). Zhang and colleagues, pooling eight trials in 423 patients twelve years later with GRADE assessment and the Hartung-Knapp adjustment, reported fasting glucose −0.85 mmol/L (I² = 73.4%), postprandial glucose −2.28 mmol/L (I² = 66.9%), glycated haemoglobin −0.38 percentage points (I² = 37.6%), and total cholesterol −0.38 mmol/L, with no effect on triglycerides, HDL, or LDL, and called for further high-quality studies (Zhang et al., 2024). Grade: emerging. The standardisation problem Cochrane identified in 2012 — that no two trials tested the same preparation — is not addressed by pooling more trials of different preparations.

The comparative ranking exercise. Kumar and colleagues attempted a network meta-analysis of 44 trials in 3,130 participants across six herbs, combining direct and indirect evidence to rank them. Apple cider vinegar, cinnamon, curcumin, and fenugreek all significantly reduced fasting blood glucose against placebo, and apple cider vinegar was ranked the most effective herb for that outcome; only apple cider vinegar and fenugreek were found effective for glycated haemoglobin. The authors concluded that health professionals should be encouraged to incorporate these herbs into standard care (Kumar et al., 2023). The numbers underlying that ranking are addressed in section 21, and the recommendation is not one endorsed here.

Compound	Mechanism claimed	Mechanism status in humans	Human outcome status	Population actually studied
Acetic acid	slows gastric emptying, blunts glucose appearance	strongly supported, dose-responsive	acute postprandial only; 8-week pilot	healthy volunteers; type 2 diabetes
Psyllium	luminal viscosity	established, physical	HbA1c −0.97 in treated T2D; nil if euglycaemic	all three, stratified
Berberine	insulin-receptor expression; AMPK; microbiota	receptor expression shown; systemic route unresolved at <1% bioavailability	HbA1c −0.19 to −0.73	type 2 diabetes only
Cinnamon	insulin-signalling effects of cinnamaldehyde	plausible; not demonstrated in humans	inconsistent between GRADE-assessed syntheses	type 2 diabetes
Fenugreek	galactomannan viscosity; other constituents	viscosity plausible; not isolated	HbA1c direction consistent, magnitude not	T2D and prediabetes
Alpha-lipoic acid	antioxidant; glucose-transport effects	clamp-demonstrated once, no dose-response	glycaemia weak; IV neuropathy established	type 2 diabetes
Magnesium	enzyme cofactor in insulin signalling	established biochemistry	HOMA-IR yes; HbA1c no	diabetes and at-risk
Chromium	glucose-tolerance factor, insulin potentiation	no validated status assay exists	HbA1c yes in T2D; explicitly nil otherwise	T2D; non-diabetic tested
Resveratrol	sirtuin activation, AMPK	speculative at achieved exposures	HOMA-IR yes in T2D; nil in non-diabetic	both, tested separately
Polyphenols, dietary	multiple	varied	no glucose effect, whole food or extract	cardiometabolic risk

Table 2. Mechanism against outcome. The two columns are independent. Magnesium has the best-established biochemistry and does not move glycated haemoglobin; berberine has the largest outcome file and an unresolved route of action; chromium has an outcome signal and no way to identify who is deficient.

20 Blends: evidence against co-occurrence

The workbook underlying this review contained a large number of multi-ingredient blood-sugar and metabolic formulations. Not one of them was the subject of a trial of itself. This is the normal state of the category, and it has a specific consequence: a blend inherits neither the efficacy evidence nor the safety data of its constituents.

A typical product contains berberine, chromium, cinnamon, alpha-lipoic acid, bitter melon, Gymnema, and a bulking fibre. Read against the preceding ten sections, the following are all unknown for that product: whether any ingredient is present at a dose used in any trial; whether the cinnamon is cassia or verum and what its coumarin content is; whether the berberine formulation resembles any of the poorly bioavailable preparations tested; whether the glucose-lowering actions of seven ingredients are additive, sub-additive, or supra-additive; and what the combined effect is in a person also taking metformin, a sulfonylurea, or insulin.

What is knowable is that any glucose-lowering the product achieves is most plausibly attributable to whichever constituent has the largest documented effect at the dose present, and that the blend's own evidence base is empty.

Two combinations in this file do have direct evidence, and both are narrower than the practice they are used to justify.

Berberine added to oral hypoglycaemic drugs. Dong and colleagues found that berberine combined with metformin, glipizide, or rosiglitazone produced better glycaemic control than those drugs alone (Dong et al., 2012). That is a supplement-plus-prescription-drug finding in people with diagnosed diabetes under medical supervision, and it is the strongest reason here for the interaction concern in section 22 — not a licence for unsupervised combination.

Chromium plus biotin, and chromium plus vitamins C and E. Suksomboon and colleagues identified three trials of combined products among 25, and reported the combination data as limited and inconclusive (Suksomboon et al., 2014).

Everything else is co-occurrence. The regulatory frame is worth stating plainly, because it is frequently misunderstood: in the United States a dietary supplement is not reviewed for efficacy before sale, the manufacturer is responsible for substantiating its own claims, and a structure-function claim on a label is not a finding by any agency (FDA, 2026).

21 Numbers that do not survive being read

Three of the syntheses cited here contain internal inconsistencies in their own published abstracts. Reporting them is not pedantry; it is the difference between citing a literature and repeating it.

Kumar and colleagues' network meta-analysis reports, for the reduction of fasting blood glucose against placebo, standardised mean differences of −28.99 for apple cider vinegar, −9.73 for cinnamon, −13.15 for curcumin, and −19.64 for fenugreek. A standardised mean difference is expressed in units of the outcome's standard deviation. A value of −28.99 would mean the treated group's mean lay twenty-nine standard deviations below the control group's — an effect that does not occur in biology and would not require a trial to detect. The most likely explanation is that raw mean differences in mg/dL have been labelled as standardised mean differences. The same abstract states that "only ACV (SMD = −2.10) and fenugreek seeds (0.84) were found significantly effective in reducing HbA1C" — quoting a positive number as a reduction (Kumar et al., 2023). Since the ranking of six herbs is the paper's entire contribution, and the ranking depends on these magnitudes, this review does not use the ranking. The paper's qualitative finding — that several of these herbs have some effect on fasting glucose in type 2 diabetes — is consistent with the rest of the file and is reported on that basis alone.

Shabil and colleagues report their fenugreek results as "a reduction in fasting blood glucose levels (MD: 3.70, 95% CI of −27.02, 19.62; p = 0.76), postprandial blood glucose (MD: −10.61, 95% CI of −68.48, 47.26; p = 0.72), and HbA1c (MD: −0.88, 95% CI −1.49, −0.27; p = 0.00)". The first figure is positive, and both of the first two confidence intervals cross zero with p-values above 0.7. Only the glycated-haemoglobin result is statistically significant. Describing all three as reductions is not supportable from the numbers printed (Shabil et al., 2023). This review therefore cites Shabil and colleagues for the glycated-haemoglobin direction only, and section 14 says so.

Guo and colleagues report the berberine lipid results as "the reduction of TG (MD = −0.5...), TC (MD = 0.64, 95% CI (−0.78, −0.49)) and LDL (MD = 0.86, 95% CI (−1.06, −0.65))" — two positive point estimates sitting outside their own entirely negative confidence intervals, which are almost certainly dropped minus signs (Guo et al., 2021). The glycaemic estimates in the same abstract are internally consistent, and those are what section 10 uses.

The general rule applied here: where a published abstract contains an arithmetic or sign inconsistency, the finding is reported as a direction and never as a magnitude, and the inconsistency is stated. A pooled estimate that its own authors have transcribed incorrectly is not a quantity anyone should carry into a clinical or purchasing decision.

22 Interactions, and the reason a glucose-lowering supplement is not neutral

This is the most important safety section in the article, and it follows directly from the efficacy findings rather than sitting apart from them.

The logic is simple. Berberine lowers glucose in people with type 2 diabetes with efficacy the primary trials described as similar to metformin (Yin et al., 2008; Zhang et al., 2010). Added to metformin, glipizide, or rosiglitazone, it produced better glycaemic control than those drugs alone (Dong et al., 2012). Psyllium lowered fasting glucose by 37 mg/dL in people being treated for type 2 diabetes, and Gibb and colleagues explicitly flagged the need to work out how to incorporate it into treatment algorithms with concomitant hypoglycaemic medications (Gibb et al., 2015). Chromium's effect was concentrated in patients with inadequate baseline control — that is, people on treatment (Suksomboon et al., 2014). Cinnamon reduced postprandial glucose by 39 mg/dL in the largest synthesis (de Moura et al., 2025).

Every one of those is a demonstration of additive glucose-lowering in medicated patients. The predictable consequence of additive glucose-lowering, when one of the agents is insulin or an insulin secretagogue such as a sulfonylurea, is hypoglycaemia — and unlike metformin, agents in those two classes cause it on their own. The American Diabetes Association's pharmacologic algorithm and the American Association of Clinical Endocrinology algorithm both organise treatment intensity around that risk (American Diabetes Association Professional Practice Committee, 2026b; Samson et al., 2026). A supplement that adds an unquantified, unlabelled, batch-variable amount of glucose-lowering on top of a titrated prescription is, by that logic alone, a clinically material exposure.

A second interaction channel is pharmacokinetic rather than pharmacodynamic. Berberine is a P-glycoprotein substrate subject to efflux and to extensive hepatic and intestinal metabolism (Cui et al., 2024). That profile is the standard signature of a compound with the potential to alter the disposition of co-administered drugs, and it applies to every prescription a person takes, not only their diabetes therapy.

A third is mundane and well documented: viscous fibre taken with medication can alter the absorption of that medication, which is why the timing question Gibb and colleagues raised is not trivial (Gibb et al., 2015).

Grade: established that several compounds here produce additive glucose-lowering when added to prescription glucose-lowering therapy, because that is what the trials demonstrating their efficacy were designed to show. Plausible and unquantified for pharmacokinetic interaction via P-glycoprotein and hepatic metabolism. NCCIH's assessment of diabetes and dietary supplements states the practical position for a general readership (NCCIH, 2026a).

The asymmetry is worth naming. The evidence that these products lower glucose and the evidence that they can cause harm are the same evidence. A category cannot claim the first while treating the second as hypothetical.

23 Safety, tolerability and populations

Beyond interaction, four concrete signals appear in this file.

Gastrointestinal intolerance with berberine. Twenty of 58 patients — 34.5% — experienced transient gastrointestinal adverse effects in Yin and colleagues' studies, with no liver or kidney damage observed (Yin et al., 2008). Guo and colleagues assessed serum creatinine, blood urea nitrogen, and adverse events across 46 trials and concluded that berberine was safe as used (Guo et al., 2021). Grade: established that gastrointestinal effects are common at trial doses and strongly supported that no organ toxicity emerged in the trial record.

Coumarin exposure from cassia cinnamon. Section 13. A tolerable daily intake of 0.1 mg/kg body weight derived partly from human hepatotoxicity data, a hypersusceptible human subgroup of unknown cause, and culinary intake that can approach the limit before any supplement is added (Abraham et al., 2010; NCCIH, 2026b).

Trivalent chromium exposure. Georgaki and colleagues raise the potential risks of chromium supplementation explicitly alongside its possible benefits, across doses of 50 to 1,000 µg/day and four different chemical forms, and note that no consensus exists on how to assess chromium status (Georgaki et al., 2024; Balk et al., 2007; NIH ODS, 2026a). Grade: insufficient. The absence of a status assay means neither benefit nor risk can be targeted to the people it applies to.

An unexpected lipid direction. Ye and colleagues' pooled mineral analysis in polycystic ovary syndrome found a small but statistically significant reduction in HDL cholesterol (SMD −0.19, p = 0.04) alongside the glycaemic improvements (Ye et al., 2026). Whatever its interpretation, it is a reminder that "improves metabolic markers" is a directionless phrase.

Populations not studied. Pregnancy, lactation, childhood, and adolescence are essentially absent from this entire evidence base. So are people with hepatic or renal impairment, in whom both coumarin exposure and drug-interaction consequences would be most serious. Ye and colleagues' polycystic ovary syndrome analysis is the only trial set here in a specifically female metabolic population, and it pools four minerals (Ye et al., 2026).

Concern	Evidence	Grade	What is not known
Additive hypoglycaemia with insulin or sulfonylureas	berberine improved control added to oral agents; psyllium −37 mg/dL in treated patients	established as a mechanism; unquantified as a risk	incidence; which products; what dose
Berberine pharmacokinetic interaction	P-glycoprotein substrate; hepatic and intestinal metabolism	plausible	magnitude for any specific co-medication
Coumarin from cassia cinnamon	human-derived TDI 0.1 mg/kg; hypersusceptible subgroup	established	species and content of retail supplements
Berberine gastrointestinal effects	34.5% transient in 58 patients	established	tolerability at chronic use
Trivalent chromium exposure	50–1,000 µg/day used; no status assay exists	insufficient	who is deficient; long-term safety
HDL reduction with mineral supplementation	SMD −0.19, p = 0.04, pooled across four minerals	emerging	which mineral; clinical meaning
Fibre–drug absorption timing	flagged by the psyllium meta-analysis authors	plausible	which drugs, what interval
Blend interactions	no trials of any blend as a product	no evidence	everything
Pregnancy, lactation, children, organ impairment	essentially unstudied	insufficient	risk estimates

Table 3. Safety ledger. The row with the strongest evidence is the first one, and it is strong precisely because it is the same evidence the category uses to advertise efficacy.

24 What actually improved outcomes

Reduce this article to interventions that changed something a patient would notice, in a trial large enough to believe, and the list is short. None of the entries is a supplement.

A lifestyle programme prevented diabetes. 58% reduction in incidence over 2.8 years, needing 6.9 participants to prevent one case in three years — more effective than metformin (Knowler et al., 2002). Over 15 years the reduction persisted at 27%, though a majority of the cohort developed diabetes regardless, and the aggregate microvascular outcome did not differ significantly between groups (Diabetes Prevention Program Research Group, 2015).

Weight loss prevented diabetes, linearly. Across 44 trials in 14,742 people with prediabetes, each increment of weight loss from 1% to 9% moved both regression to normoglycaemia and progression to diabetes, with no threshold and no difference between diet, exercise, or both (Jayedi et al., 2024).

An intravenous antioxidant relieved neuropathic symptoms. 1,258 patients, 600 mg/day for three weeks, responder rates 52.7% against 36.9% (Ziegler et al., 2004a). This is the only symptom-level clinical benefit in the entire article, it required a hospital infusion, and it is not a blood-sugar finding.

Continuous glucose monitoring helped people with prediabetes and not people with normal glucose (Liao et al., 2026).

Set beside those, the best supplement results in this file are: glycated haemoglobin lowered by roughly 0.2 to 0.7 percentage points in people with type 2 diabetes, by compounds whose pooled estimates disagree with one another by a factor of four, in trials whose heterogeneity the authors themselves describe as limiting. And recall from section 06 what happened when four licensed drugs, all achieving glycaemic targets, were compared head to head for five years in 5,047 people: essentially no difference in microvascular outcomes, cardiovascular events, or death (GRADE Study Research Group, 2022).

Three inferences follow.

First, the surrogate is not the disease. A compound can move glycated haemoglobin and have no demonstrated effect on any outcome a person experiences, and the trial evidence for licensed drugs is the reason to take that seriously rather than a rhetorical hedge.

Second, the effect sizes available here are small even when real, and they are concentrated in the population least likely to be self-medicating without supervision.

Third, the interventions that worked are behavioural and pharmaceutical. The existing Metabolic-Nutrition, Exercise-Intervention, Strength-Training, and Dietary-Fibre articles in this series carry those literatures; the Microbiome-Nutrition title carries the fermentable-fibre and probiotic arm that Shadin and colleagues found performed comparably to berberine (Shadin et al., 2026); and the World Health Organization's fact sheet supplies the population context (WHO, 2026).

25 The "blood sugar support" sentence

The market sentence says: these compounds lower blood sugar; berberine is nature's metformin; cinnamon fights insulin resistance; alpha-lipoic acid repairs metabolic damage; chromium fixes cravings; a spike on your sensor is harm you can prevent with a capsule; and because these are foods and spices, combining them is safe.

The file says otherwise, and it is worth setting the two side by side sentence for sentence.

Berberine lowers glycated haemoglobin in people with type 2 diabetes by an amount that fell from 0.73 to 0.19 percentage points as the evidence base widened, on a trial set whose methodological quality its own reviewers called generally low, with a route of action unresolved at under 1% oral bioavailability (Guo et al., 2021; Nazari et al., 2024; Shadin et al., 2026; Dong et al., 2012; Cui et al., 2024).

Cinnamon's glycated-haemoglobin effect is genuinely unresolved between GRADE-assessed syntheses, absent in the two analyses that examined dose-response, and reported alongside heterogeneity of 94% and 100%; the strongest thing established about cassia cinnamon is its coumarin content (Allen et al., 2013; Yu et al., 2023; de Moura et al., 2025; Suksomboon et al., 2011; Abraham et al., 2010).

Alpha-lipoic acid's strong evidence is a three-week intravenous course for symptomatic neuropathy in people who already have nerve damage; its oral glycaemic evidence is one positive clamp study with no dose-response, one null four-month trial, and an equivocal result in twelve people (Ziegler et al., 2004a; Ziegler et al., 2004b; Jacob et al., 1999; de Oliveira et al., 2011; Gosselin et al., 2019).

Chromium's effect is real in treated diabetes, larger in worse studies, and explicitly absent in people without diabetes — and no one can identify who is deficient (Balk et al., 2007; Suksomboon et al., 2014; Georgaki et al., 2024; NIH ODS, 2026a).

Magnesium moves a derived resistance index and not glycated haemoglobin, which is the signature of correcting a shortfall (Simental-Mendía et al., 2016; Veronese et al., 2021).

Resveratrol works in diabetes and does nothing in people without it, tested and confirmed in three separate syntheses; dietary polyphenols, whole-food or extract, did not affect fasting glucose at all (Liu et al., 2014; Delpino et al., 2022; Zhang et al., 2021; Kiyimba et al., 2023).

Vinegar and viscous fibre do exactly what their mechanisms predict, only on the meals and in the people where there is something to slow (Östman et al., 2005; Johnston et al., 2005; Liatis et al., 2010; Gibb et al., 2015).

Gymnema rests on 15 participants per arm (Gaytán Martínez et al., 2021). Bitter melon's Cochrane review could not pool four trials because no preparation was tested twice (Ooi et al., 2012; Zhang et al., 2024). Fenugreek's direction is consistent and its magnitude is not (Kim et al., 2023; Shabil et al., 2023). Blends have no evidence of themselves, and the safest thing that can be said about combining several glucose-lowering agents with a prescription is that nobody has measured it (FDA, 2026).

And a healthy person's glucose is above 140 mg/dL for about half an hour a day, with a 17% within-person coefficient of variation, and wearing a sensor confers no glycaemic benefit on such a person at all (Shah et al., 2019; Liao et al., 2026).

Neither "these work" nor "nothing works" survives contact with the trials. What survives is narrower: the metabolic-supplement category contains two mechanisms that do exactly what physiology predicts on the meals where there is something to slow, several compounds that lower a surrogate in people who already have diabetes by amounts their own meta-analyses cannot agree on, one compound whose strongest evidence is an intravenous course for nerve damage rather than a glycaemic effect, an explicit and replicated finding of no benefit in people whose glucose is normal, and a safety profile whose most solid element is the additive hypoglycaemia implied by the very trials used to advertise it — while the interventions that prevented diabetes were weight loss and a lifestyle programme, and four licensed drugs achieving glycaemic targets could not be shown to differ from one another on the outcomes that matter.

26 Standing constraint

27 References

28 Evidence handling

The numbered list above is generated at build from records retrieved from NCBI and from regulator, agency, and chemical-database sources whose URLs were resolved and checked; nothing in it was typed from memory. Two candidate sources failed verification and were replaced rather than cited from recollection: a publisher landing page for the current Standards of Care returned HTTP 403, and was replaced by the peer-reviewed chapter records themselves; two Food and Drug Administration pages on tainted products returned HTTP 404, and were replaced by the agency's dietary-supplement questions-and-answers page.

Study types are named in the sentence that uses them. Cell and receptor work, single-meal crossover studies in volunteers, isoglycaemic clamps, small parallel-group trials, multi-week supplementation trials, meta-analyses, dose-response meta-analyses, umbrella reviews of meta-analyses, network meta-analyses, Cochrane reviews that declined to pool, multi-year prevention programmes with incident-disease endpoints, five-year comparative-effectiveness trials with microvascular and cardiovascular endpoints, sensor-reference studies in healthy volunteers, toxicological risk assessments, and professional-body algorithms are not interchangeable. Where results conflict, both are reported: cinnamon's four syntheses are set out in full rather than summarised as mixed evidence, and berberine's shrinking estimate is stated as a sequence.

Where a synthesis reported heterogeneity, the heterogeneity statistic is given, and an I² of 100% is treated as invalidating the pooled mean rather than as a footnote. Where a published abstract contains an internal inconsistency, it is recorded rather than silently corrected — section 21 sets out three, in Kumar and colleagues, Shabil and colleagues, and Guo and colleagues, and each affected finding is reported as a direction rather than a magnitude.

Populations are named every time, because the central failure mode of this literature is transferring a result from a treated diabetic population to a reader whose glucose is normal. Every reference resolves to a record retrieved from NCBI or to an agency or database page whose URL returned HTTP 200 and the expected document title at build. Project 06 was consulted read-only as a general-science corpus and is not the evidence spine for this title; no write was made to it or to project06_catalog.sqlite. News reporting is not treated as scientific evidence anywhere in this document.

Adversarial interrogation — the required adversarial challenge list — is disposed as follows.

Which claims are overstated? ACCEPTED as the organising question. The metformin-equivalence claim for berberine, the insulin-resistance claim for cinnamon, the metabolic-repair claim for oral alpha-lipoic acid, and the whole notion of blood-sugar benefit in normoglycaemic people are each named and refused in the sections that report their evidence.

What contrary trials exist? ACCEPTED and given priority. Allen and colleagues' null on glycated haemoglobin and Yu and colleagues' null on fasting glucose are reported at length beside the positive syntheses; de Oliveira and colleagues' null is given equal weight to Jacob and colleagues' positive clamp; Ooi and colleagues' Cochrane review is reported as a refusal to pool rather than as absence of evidence; and Kiyimba and colleagues' finding of no glucose effect from any polyphenol form is stated as a negative rather than omitted.

Are effect sizes clinically meaningful? ACCEPTED as a standing doubt and answered structurally in section 06. The GRADE trial's failure to distinguish four licensed glucose-lowering drugs on microvascular and cardiovascular outcomes over five years is the reason no surrogate movement here is described as clinical benefit.

Is the evidence population relevant? ACCEPTED as the article's spine. Section 04 states the three-population problem through Gibb and colleagues' stratified analysis, and every subsequent grade names its population. The right-hand column of Table 1 exists to make the omission visible.

Is there publication bias? ACCEPTED as a structural expectation given the sample sizes and quality distribution in section 05, and asserted as a detected finding nowhere, because none of the syntheses cited here reported a funnel-plot analysis in its abstract. Balk and colleagues' observation that larger effects were more commonly observed in poor-quality studies is reported as what it is — a quality gradient, which is a related but distinct phenomenon.

Are commercial claims stronger than the evidence? ACCEPTED. Figure 3 states the gap claim by claim.

Are mechanisms being mistaken for outcomes? ACCEPTED. Table 2 separates the two columns explicitly; magnesium's well-established enzymology beside its null glycated-haemoglobin result, and berberine's large outcome file beside its unresolved route of action at sub-1% bioavailability, are the reference cases.

Are safety concerns underrepresented? ACCEPTED, and treated as inseparable from efficacy. Section 22 derives the interaction hazard from the efficacy trials themselves rather than from a closing caveat, and Table 3 places additive hypoglycaemia first.

Are doses and formulations genuinely comparable? ACCEPTED and treated as decisive three times: cassia against verum cinnamon and the coumarin question; intravenous against oral alpha-lipoic acid; and bitter-melon preparations of which no two were alike.

Does any paragraph imply certainty not supported by evidence? ACCEPTED as a drafting constraint. Every efficacy statement carries an in-place grade, "inconsistent" is used where two GRADE-assessed syntheses disagree rather than defaulting to "mixed", and "not studied" is used where the population that buys a product was never enrolled.

Unresolved science remains substantial: an adequately powered trial of any single compound here against a clinical endpoint rather than a surrogate; a trial of berberine with a pharmacokinetically characterised formulation and a stated hypothesis about its route of action; a cinnamon trial that reports species, coumarin content, and cinnamaldehyde dose; a fenugreek trial with a viscosity-matched fibre comparator; a validated assay for chromium status without which that literature cannot be resolved; an oral alpha-lipoic acid trial powered for the neuropathy endpoint the intravenous programme established; a trial of any commercial multi-ingredient formulation as a product; and a systematic quantification of hypoglycaemia in people combining these supplements with insulin or sulfonylureas, which is the question this review would most like answered and the one on which it found no evidence at all.