Applies to a specific decision
This is a decision-scoped read for a person considering a GLP-1 for weight. The answer turns on baseline cardiometabolic risk — the same drug is a proven hard-outcome lever for one person and a surrogate-only weight change for another. It is not population-wide advice to start or avoid it.
The drug is two decisions wearing one name
A GLP-1 receptor agonist does the same thing to everyone’s appetite — it resupplies a satiety signal from the outside, so you eat less. But its value forks on who takes it. The benefit that matters most — a prevented heart attack, stroke, kidney failure, or death — scales with the risk a person already carries. Sort by baseline cardiometabolic risk and the drug resolves into two answers wearing one name.
For a person with obesity and established cardiovascular disease, kidney disease, or type-2 diabetes, this is a large, proven lever on the outcomes that count — a Big Rock, not a refinement. Semaglutide reliably cuts hard cardiovascular events and death in secondary prevention, and a class-level meta-analysis now shows that benefit is a property of the whole drug class, not one molecule. In diabetic kidney disease it slows hard kidney outcomes too.
For a near-normal-weight person with no cardiometabolic disease, the picture inverts. The weight loss is large and reliable, but it is a surrogate: any reduction in heart attacks, strokes, kidney failure, or death is unproven, because the hard-outcome trials enrolled only high-risk patients. This person pays the same tolerability cost — and, in serious-event terms, nets more harm than benefit, because few hard events stand to be averted. The weighing between a certain surrogate gain and an unproven hard-outcome one is irreducibly theirs.
Every section below rests on two facts. The benefit is a maintained state, not a cure — it lasts only while the drug is taken. And the class is not uniform: tirzepatide takes off more weight than semaglutide but has no hard-outcome trial at all, so a bigger number on the scale does not carry a bigger proven benefit.
What the drug reliably does is make you eat less
The best-evidenced effect is on body weight, and its size depends on the agent and the population. In STEP-1 — semaglutide 2.4 mg weekly in non-diabetic obese adults (mean age 46), adjunct to lifestyle — mean weight fell 14.9% versus 2.4% on placebo at week 68, a placebo-adjusted 12.4 percentage points (95% CI 11.5 to 13.4) (Wilding et al., 2021). In SELECT — the same drug in an older secondary-prevention population — the effect was smaller, roughly 9-10%, a difference of 8.5 points at week 104 (Lincoff et al., 2023).
The response is heterogeneous — a mean of ~15% hides both big responders and non-responders. STEP-1 makes the spread explicit: 69% of the semaglutide arm lost ≥10%, 51% lost ≥15%, and 32% lost ≥20% (versus 12%, 5% and 2% on placebo) — so a substantial minority did not reach even 10% (Wilding et al., 2021). The mean is not the individual estimate.
Tirzepatide is a distinct drug, not a stronger dose of the same one. It is a dual GIP/GLP-1 receptor agonist — it adds a second receptor axis rather than intensifying GLP-1 agonism. In SURMOUNT-1 (the same non-diabetic obese class as STEP-1), weight fell 15.0%, 19.5%, 20.9% at 5/10/15 mg versus 3.1% placebo, with half to 57% of the top two doses losing ≥20% (Jastreboff et al., 2022).
A network meta-analysis of 262 trials (99,791 participants, 19 drugs) ranks the class against a common reference — lifestyle alone — supplying the head-to-head the single-drug trials cannot. At one year tirzepatide takes off 14.9% (95% CI 13.9 to 16.0) against subcutaneous semaglutide’s 9.8% (9.1 to 10.6), both at moderate-to-high certainty (Nong et al., 2026) -> Comparing Obesity Drugs. The pivotal trials cannot be compared directly: «no direct comparison of these trials can be made, since trial populations and designs differed» (Jastreboff et al., 2022).
Every one of these numbers is a surrogate. Weight is a marker; whether moving it moves an outcome a person actually cares about is a separate evidenced step, never assumed from the scale -> Surrogate Outcomes. That discipline bites hardest exactly here, where the effect is so large and reliable that reading the surrogate as the outcome is most tempting.
Hard cardiovascular events fall as a class — but the benefit is proven where risk is high
This outcome turns the drug from a weight intervention into a lever on something people care about, and the evidence has grown from one trial to a class.
SELECT is the landmark for the obese, non-diabetic stratum. In 17,604 patients aged 45+ with established cardiovascular disease and a BMI ≥27, on top of standard care (90% on statins, 86% on antiplatelets), semaglutide 2.4 mg cut the primary MACE composite (CV death, non-fatal MI, non-fatal stroke) from 8.0% to 6.5% over a mean 39.8 months — HR 0.80 (95% CI 0.72-0.90) (Lincoff et al., 2023). That is a 1.5-percentage-point absolute reduction, an NNT of about 67 over ~3.3 years (inferred from Lincoff et al., 2023) — incremental benefit on top of guideline therapy, not instead of it.
The class-level meta-analysis now shows this is not a semaglutide one-off. Badve 2024 pools 11 cardiovascular-outcome trials (85,373 participants, median follow-up ~2 years) and finds the whole GLP-1 receptor agonist class cuts MACE — HR 0.87 (0.81-0.93) in type-2 diabetes, 0.86 (0.80-0.92) with SELECT added, NNT 74 — and all-cause death HR 0.88 (0.83-0.93), NNT 101, both high-certainty; Badve reports that effects «appeared consistent regardless of diabetes status» (Badve et al., 2025). Only nonfatal heart attack differed by diabetes status (p-heterogeneity 0.02), and both groups still benefited (Badve et al., 2025); but SELECT is the only trial without diabetes, so the non-diabetic side of that comparison is one trial.
Badve calls the class «the first and only class of medications with proven benefits on composite kidney and cardiovascular outcomes… across a range of cardiovascular risk and chronic kidney disease severity in people with and without diabetes» (Badve et al., 2025). The single-agent trials could not establish this: the CV/mortality benefit is a class property, robust to leaving out any one trial -> GLP-1 Receptor Agonists and Cardiovascular and Kidney Outcomes.
But Badve is not a second independent line — it pools SELECT and shares its authors — so it is refinement of the same evidence, not corroboration by a separate route. The wiki reads the class benefit as largely weight-independent, because most pooled trials are glucose-lowering type 2 diabetes trials with modest weight loss. Badve does not report when the curves separate. In the single trial SELECT, the effects «occurred early after the initiation of treatment», and the investigators read this as suggesting that changes «beyond the magnitude of body-weight loss may have mediated at least part» of the benefit (Lincoff et al., 2023). So this evidence reinforces — does not overturn — the finding that losing weight by itself is not a guaranteed cardiovascular lever -> Does Weight Loss Reduce Cardiovascular Events.
The composite is carried by non-fatal events, not deaths
Only SELECT’s primary composite is a confirmatory result. Its pre-specified gatekeeping hierarchy tested CV death next, and that test failed — HR 0.85 (0.71-1.01), P=0.07 — so superiority testing stopped there and everything downstream is a point estimate, not a confirmed effect (Lincoff et al., 2023):
| Endpoint | HR (95% CI) | Confirmatory status |
|---|---|---|
| Primary MACE | 0.80 (0.72-0.90), P<0.001 | Confirmatory — hit |
| CV death | 0.85 (0.71-1.01), P=0.07 | Failed the pre-specified gate |
| All-cause mortality | 0.81 (0.71-0.93) | Exploratory (below the failed gate) |
| Heart-failure composite | 0.82 (0.71-0.96) | Exploratory |
| Non-fatal MI | 0.72 (0.61-0.85) | Favourable, exploratory |
| Non-fatal stroke | 0.93 (0.74-1.15) | Not reduced |
Read this without either error. The CV-death miss is not evidence of no effect — its interval still spans a meaningful benefit, and the endpoint was gatekept and underpowered. It means the trial did not license a confirmatory claim there, so P=0.07 must not be read as significance. All-cause mortality (0.81) is directionally favourable and, at the class level, high-certainty in Badve. Non-fatal MI carries most of the composite, while stroke was not reduced and CV death did not clear the gate. So within SELECT, MACE fell but CV deaths did not — MACE was reduced must not be read as deaths were reduced (inferred from Lincoff et al., 2023).
For a low-risk person the benefit is unproven, not merely unstated
SELECT’s trialists draw the boundary in their own words: «we included only patients with preexisting cardiovascular disease. The effects of semaglutide on primary prevention of cardiovascular events in persons with overweight or obesity but without previous atherosclerotic disease were not studied.» (Lincoff et al., 2023). Badve’s class effect is robust, but SELECT is still the only non-diabetic trial in the pool, so the primary-prevention gap is intact.
The gap does not close by transporting the relative effect. Absolute benefit is the relative reduction applied to a person’s own baseline risk, so a constant hazard ratio buys a shrinking return as risk falls Baseline Risk and the Relative-Absolute Split. A primary-prevention person at ~5% ten-year risk sits roughly five-fold below SELECT’s event rate; apply the same HR 0.80 and the ~1.5-point reduction shrinks toward a fraction of a percentage point. The hard-CV benefit for a low-risk, near-normal-BMI person is expected-small and never tested — insufficient evidence, not a demonstrated null (inferred from Lincoff et al., 2023).
Tirzepatide has no hard-CV trial at all
The class’s strongest weight lever is the one with the least outcome evidence. Tirzepatide’s pivotal obesity trial measured weight, a surrogate, and its cardiovascular morbidity-mortality trial does not yet exist:. Across the 19-drug network, subcutaneous semaglutide is the only agent reaching a mortality (RR 0.81) or MI (0.72) signal — and the network flags that limb as «largely informed by cardiovascular outcome trials in high risk populations», i.e. SELECT re-pooled, not new corroboration (Nong et al., 2026). Tirzepatide reaches a heart-failure signal in the same network (HR 0.49, 0.27-0.88) but no mortality or MI benefit (Nong et al., 2026), and a network estimate is not a dedicated outcome trial.
A record-breaking weight number does not license a hard-outcome claim: Look AHEAD is the standing worked case where a large surrogate improvement bought no measurable event reduction -> Does Weight Loss Reduce Cardiovascular Events. Until SURMOUNT-MMO reports, tirzepatide’s hard-CV effect is insufficient-evidence, not no-effect.
The kidney is now its own proven leg
Kidney disease is a third distinct benefit, on a different endpoint in a different stratum — and the evidence now runs from one dedicated trial up to the class.
FLOW is the first dedicated hard-outcome kidney trial of a GLP-1 drug. It enrolled a narrow, high-risk population: type-2 diabetes with established chronic kidney disease (eGFR 25-75, elevated albuminuria), already on a maximal-dose RAS inhibitor, mean eGFR 47, BMI 32, age 66.6; semaglutide 1.0 mg weekly (the glycaemic dose, not the 2.4 mg obesity dose), n=3533, median 3.4 years, stopped early for efficacy (Perkovic et al., 2024). On that stratum the effect is large and patient-important:
- Major kidney disease events (primary composite): HR 0.76 (95% CI 0.66-0.88), P=0.0003, 5.8 vs 7.5 events/100 patient-years, NNT 20 over 3 years (14-40).
- Death from any cause: HR 0.80 (0.67-0.95), NNT 39; death from CV causes HR 0.71 (0.56-0.89); MACE HR 0.82 (0.68-0.98), NNT 45 — the survival signal, not just the renal one (Perkovic et al., 2024).
Two qualifiers keep it honest. The composite is not all hard kidney failure — it pools kidney failure (dialysis/transplant) with a surrogate ≥50% eGFR-decline threshold and death, and FLOW was not powered for kidney failure alone. The defensible claim is reduced a composite of kidney events and reduced death, not prevented dialysis on its own. And the benefit was weight-independent — creatinine- and cystatin-C-based eGFR agreed, so it is not merely weight loss lowering serum creatinine (Perkovic et al., 2024).
Badve raises the kidney finding to a class effect on a hard outcome. The prior class analysis credited a kidney benefit «mainly driven by new-onset macroalbuminuria, a surrogate outcome not validated for clinical kidney disease outcomes»; Badve re-runs it on a composite that excludes macroalbuminuria — HR 0.82 (0.73-0.93), high-certainty, NNT 164 — and, for the first time, demonstrates a class-level kidney-failure reduction, HR 0.84 (0.72-0.99), moderate-certainty (Badve et al., 2025). Kidney failure is less exposed to the muscle-loss artifact than the eGFR-decline endpoint: it is a late, threshold event, and four trials define it by dialysis or transplant alone. But in the trials carrying most of the weight, FLOW included, it still counts a persistent eGFR below 15, so it partly shares the creatinine channel . Badve offers it as reassurance, not proof: «The separately significant reduction in the risk of kidney failure provides important reassurance that the kidney benefits of GLP-1 receptor agonists are real and clinically important» (Badve et al., 2025). FLOW carried 49% of that kidney-failure weight, so this is largely FLOW confirmed at scale, not an independent route.
The absolute kidney benefit is modest outside high kidney risk. Badve’s pooled trials mostly enrolled near-normal kidney function (mean baseline eGFR 77), so the class kidney NNT runs to 164 (Badve et al., 2025) against FLOW’s 20 (Perkovic et al., 2024). The relative effect is real, but the payoff scales with baseline kidney risk (route (a)) and is decision-moving mainly in the FLOW-type stratum. It does not transport to a low-risk, near-normal-BMI person -> Semaglutide and Kidney Outcomes in Chronic Kidney Disease.
Progression to diabetes drops sharply
Two-thirds of SELECT was prediabetic, and progression to diabetes fell steeply — progression to HbA1c ≥6.5% ran at HR 0.27 (95% CI 0.24-0.31) (Lincoff et al., 2023). STEP-1 shows the same in a younger population: 84% of prediabetic participants on semaglutide reverted to normoglycaemia by week 68 versus 48% on placebo (Wilding et al., 2021). For an obese, impaired-fasting-glucose person this is a distinct, plausibly-relevant benefit — but it sits on a glycaemic-threshold surrogate plus a progression endpoint, not on a hard diabetes-complication outcome. Whether it lowers those hard outcomes in primary prevention inherits the same insufficient-evidence caveat as the CV case.
Fatty liver improves on the same weight lever
Weight loss of the magnitude these drugs produce improves fatty-liver (MASLD) disease in a dose-dependent way; the full steatosis/inflammation/fibrosis threshold ladder lives on Body Fat.
Muscle comes off with the fat — the ratio improves while absolute mass falls
The dedicated synthesis is a meta-analysis of 7 RCTs at obesity doses (821 patients), and its finding has two faces that must be read together (Laverde et al., 2026). As a proportion of total weight, lean mass rises — +1.81% (95% CI 1.1 to 2.52; I²=7%) — because fat is shed faster than lean, so body composition improves as a ratio. In absolute terms, muscle is lost — −1.74 kg (−3.04 to −0.45; I²=98%), a −3.06% fall. Semaglutide shows the largest loss, −5.44 kg, −9.9%, from a single small body-scan substudy. The near-98% heterogeneity is itself a reason to hold the absolute figure loosely.
Laverde’s own verdict lands on the ratio face — «lean mass loss should not be considered a limitation for the use of these drugs in patients with obesity» (Laverde et al., 2026) — which is faithful to the composition data and not a denial that absolute muscle falls; both are true at once.
The lost muscle is roughly what any rapid loss costs — not a drug-specific hazard. Laverde cites an earlier meta-analysis in which about 30% of the mass lost on a GLP-1 drug is lean, «comparable to that observed after bariatric surgery»; across the trials Laverde pooled, the share ran by agent: liraglutide 14-22%, tirzepatide ~26%, semaglutide up to 45% (Laverde et al., 2026). Against the typical 20-30% lean share of weight loss, the class sits at or just inside that band. Semaglutide’s upper value sits above it, but on thin data: 45% is the top of a range from one small body-scan substudy, not a typical value, so it does not compare like-for-like with the band -> Protein and Resistance Training for Muscle and Strength. Laverde puts the larger semaglutide loss down to more weight lost overall — «The greater reduction observed with semaglutide could be attributed to its increased potency and total weight loss, as previously described in studies [24].» — and flags the data as sparse: «Similarly, the number of patients and studies with semaglutide is limited, so further clinical trials with these drugs are needed to confirm our findings.» (Laverde et al., 2026).
A 19-drug network meta-analysis gives the same rank: the biggest fat-loss drugs (tirzepatide −8.3%, subcutaneous semaglutide −5.8%, both moderate certainty) are the worst for lean mass (Nong et al., 2026) -> GLP-1 and Lean Mass. That is probably the same trials counted twice, not a second line of evidence: Nong’s tirzepatide figure rests on one trial and matches the −8.3% Laverde reports for SURMOUNT-1, and Nong’s semaglutide trials very likely include the one Laverde used. So on the class average the lean loss looks generic to rapid weight loss, with semaglutide’s higher figure resting on thin data. The wiki holds no head-to-head against a matched, same-magnitude non-drug loss, so the attribution is directional, not settled.
Function is unmeasured — the streetlight gap. Composition is measured; the outcome people actually care about is not. «The included studies did not evaluate functional outcomes, such as muscle strength or physical capacity» (Laverde et al., 2026). Lean mass is a surrogate for muscle mass, which is a surrogate for strength, gait speed, falls, and independence. A reassuring ratio does not prove function is spared, and a scary kilogram figure does not prove it is lost — hold this as insufficient evidence on function, not as safe or harmful -> Sarcopenia Definition and Diagnosis.
Which face governs is a matter of stratum, and it inverts. Laverde ran in «relatively young populations with a low burden of comorbidities» — the wrong population for the worry (Laverde et al., 2026):
- Young, high-BMI adult with abundant muscle: the ratio face governs — composition improves, and the absolute loss matters less for someone with muscle to spare. Whether it comes back is untested; resistance training and protein are prudent, not urgent. Rank the muscle cost low, while noting that Laverde, studying mostly this group, still flags «the potential for clinically significant muscle loss during treatment with GLP1-RA» (Laverde et al., 2026).
- Older / sarcopenia-risk / multimorbid adult: the absolute face governs — the same kilograms become a front-line harm (falls, fractures, lost independence). And the drug’s own appetite suppression fights the defense, because sparing muscle needs more protein exactly when intake is cut. Resistance training and adequate protein lower the lean-mass cost (function unmeasured) without touching the weight benefit, but monitor on function, not on weight (inferred from Laverde et al., 2026) -> Big Rocks (Elderly), Protein and Resistance Training for Muscle and Strength.
The resistance-training advice comes from diet trials, not drug trials. A meta-analysis of six RCTs in obese older adults (mean age over 57) losing weight by caloric restriction found that adding resistance training kept about 0.8 kg of lean mass the diet alone removed (RMD 0.819 kg, 95% CI 0.364 to 1.273), while fat and total weight loss came out similar in both arms (Sardeli et al., 2018). None of those trials gave a GLP-1 drug, and they measured lean mass rather than function. Carrying the result over to drug users assumes drug-driven and diet-driven loss respond alike — the open attribution question above -> Resistance Training During Caloric Restriction.
The safety ledger: worry about the gut, not the cancers
Read each entry on two axes at once. The umbrella evidence grades each harm on two scales that routinely disagree — GRADE certainty (how well the effect is estimated) and credibility class (how robust the association is to bias) (Yang et al., 2026). Within the gastrointestinal cluster, nausea is GRADE-high but credibility-weak, while vomiting is GRADE-moderate but highly suggestive — so a headline quoting one number misleads -> Rating Certainty of Evidence.
Gastrointestinal effects are the one robust harm. Nausea OR 2.47 (1.84-3.34), vomiting 2.78 (1.91-4.06), diarrhoea 1.94 (1.52-2.49) — established, not exploratory (Yang et al., 2026). Two facts decide how much they weigh: the tax is largely dose- and titration-dependent (it tracks how fast the dose is escalated) and mostly transient and early (it fades with continued use). It is the dominant real-world adherence cost — reflected in a class-level discontinuation-for-adverse-events rate of RR 1.51 (1.18-1.94) (Badve et al., 2025) — and it is already priced into the drug decision.
The serious-adverse-event balance inverts by stratum. In FLOW’s high-risk diabetic-CKD population serious events were fewer on semaglutide (49.6% vs 53.8%) (Perkovic et al., 2024); in STEP-1’s lower-risk obese population they ran higher (9.8% vs 6.4%), the excess driven by serious gastrointestinal (1.4% vs 0%) and hepatobiliary (1.3% vs 0.2%) events (Wilding et al., 2021). Where many hard events stand to be prevented, the drug averts serious events on net; where few do, it adds them — the mirror image of the high-risk benefit, and a cost the weight number hides.
The most-feared cancers show no robust human signal — but that is not exoneration. Thyroid cancer, the rodent C-cell scare, came back imprecise and non-significant (OR 1.43, 0.95-2.13, P=.08), and pancreatic cancer’s nominal signal ran the other way, toward lower odds (OR 0.51, 0.30-0.85) (Yang et al., 2026). The trials run months to a few years; a long-latency malignancy could not realistically have surfaced yet (the expectancy test with a latency caveat), so absence of a short-term signal is not a long-term all-clear.
Between the robust harm and the absent one sits an unsettled middle — each demoted for a different reason. Gallbladder or biliary disease is a statistically significant association (OR 1.34, 1.16-1.55, P<.001), demoted only because its prediction interval sits near the null — a robustness caveat, not a doubt about the harm. Gastro-oesophageal reflux holds the same posture (OR 2.19, 1.65-2.90). Pancreatitis is different in kind: here the demotion is about causality — the consensus on whether GLP-1 drugs cause it is explicitly unsettled, with a propensity-matched analysis finding no excess against a regulatory update flagging rare fatal cases (Yang et al., 2026). Lumping the three is the error.
Some off-target signals point the protective way — candidates to watch, not banked benefits. Fewer serious infections (OR 0.89, 0.87-0.92, the one convincing class-I signal), fewer fractures (0.67, 0.52-0.87), and less all-cause dementia (0.55, 0.35-0.87) (Yang et al., 2026). The reduced-fracture signal offsets the sarcopenia/fall worry above; the dementia signal, if real, is decision-relevant.
But a second umbrella shows the dementia effect is not GLP-1-specific. Across antidiabetic classes, GLP-1 drugs were associated with lower dementia risk (RR 0.35, 0.16-0.78). But so were metformin, the thiazolidinediones, and SGLT2 inhibitors, while sulphonylureas ran the other way — and the estimate carries near-total heterogeneity (I²=98.5%) (Kuate Defo et al., 2023). A protective signal shared across unrelated glucose-lowering classes points at glycaemic control or confounding by indication rather than anything GLP-1-specific; both umbrellas pool the same observational pharmacoepidemiology, so they cannot cross-confirm. The dementia entry stays a watched protective blank.
The neuropsychiatric signal is graded by severity, and both tiers point at one stratum. On the milder, higher-frequency endpoint it leans harm: a depression meta-analysis pooled a depression-related adverse-event OR 1.49 (95% CI 1.18-1.88) (Bi et al., 2026), which it rated GRADE moderate (Bi et al., 2026). On the severe endpoint — suicidality, behind the 2023 FDA/EMA scare — a dedicated meta-analysis found no population signal, RR 0.568 (95% CI 0.077-4.205) (Bushi et al., 2025). But the interval runs from strong protection to fourfold harm, so the result is nearly uninformative rather than reassuring. The same severe endpoint rates GRADE low-to-very-low, on serious imprecision from low event rates (Bi et al., 2026).
The lone counter-signal is agent- and stratum-specific — a semaglutide WHO-VigiBase suicidal-ideation ROR 1.45 (1.18-1.77), rising to 4.45 (2.52-7.86) among users co-prescribed antidepressants — best read as confounding by indication, a marker of pre-existing psychiatric illness, not clean effect modification (Bushi et al., 2025). Both tiers converge on one stratum to watch — psychiatric-comorbid, already on psychotropics — and they are two severity tiers of one mood axis: a graded distinction, not a contradiction -> Depression and Modifiable Exposures.
The feared flattens all pleasure harm is, on current evidence, a benefit-leaning-but-unmeasured question, not a demonstrated one. The same mesolimbic-reward mechanism that cuts food craving also cuts substance craving — less alcohol, less nicotine in humans — and in animal models it is selective, sparing natural reward: liraglutide’s «effects were specific to drug-related behaviors, as it did not alter natural reward-seeking or saccharin intake» (Völker et al., 2026). That cuts against a blanket-anhedonia reading. But no human study has measured hedonic tone, anhedonia, or quality of life directly, so the pleasure-flattening harm is unmeasured, not refuted -> GLP-1 and Reward Beyond Food.
Two facts make any version of this ledger provisional. First, mechanical fragility: in leave-one-trial-out reanalysis, the direction flipped in roughly 9 of 39 outcomes and significance was lost in 6 of 39 after removing a single trial. An association a single trial can overturn is not one to dose a recommendation on (Yang et al., 2026). Second, a moving target. The mature multi-year evidence sits on one older molecule (semaglutide), while prescribing moves to agents that add receptor targets it never tested (tirzepatide adds GIP) and to higher doses. And the umbrella could not stratify harms by dose or duration. Accrued safety years de-risk the specific compound at the specific dose that accrued them; the class label stays constant while the exposure under it drifts -> GLP-1 Non-Cardiometabolic Effects and Safety.
Stopping the drug gives the weight back — and the benefit with it
The effect lasts only while the drug does. When semaglutide and its lifestyle programme were both withdrawn at week 68, participants regained a mean 11.6 percentage points of body weight over the following year (versus 1.9 on placebo). That is about two-thirds of the loss, leaving a net ~5.6% below baseline, and their cardiometabolic gains drifted back toward baseline (Wilding et al., 2022). The class-wide rate is similar: roughly 0.4 kg regained per month, a projected return to baseline within ~1.7 years, with about half of patients discontinuing within the first year (Nong et al., 2026).
This is defended physiology, not weak resolve. After weight loss the body’s satiety signals stay suppressed (leptin ~35% below baseline) and hunger stays elevated, and those shifts persist a full year out — so the relapse «has a strong physiological basis and is not simply the result of the voluntary resumption of old habits» (Sumithran et al., 2011). A GLP-1 drug does not cure this; it resupplies one of the depleted satiety signals from outside, overriding the defended set point only while present. Remove it and the deficit reasserts -> Weight-Loss Maintenance and Metabolic Adaptation. One caveat: the STEP-1 extension withdrew the lifestyle support too, so it cannot cleanly separate biological rebound from behavioural drift.
So the real decision is a lifetime one, across three routes to a sustained result. Unaided behavioural maintenance works against the defended physiology, so regain is the modal outcome. Indefinite drug therapy holds the weight and the benefit only for as long as it is taken and paid for. Bariatric surgery is the durable-structural comparator: in SOS, sustained loss of 14-25% at 10 years came with an all-cause mortality reduction — adjusted HR 0.71 (P=0.01) — that does not depend on daily adherence (Sjöström et al., 2007). SOS is non-randomized and carries its own peri-operative harm, so it is not a clean head-to-head. But it anchors the one thing the drug route structurally lacks — a benefit that survives the intervention being over -> Does Weight Loss Reduce Cardiovascular Events.
Name the second axis and stop. Indefinite therapy loads heavily on an axis the health evidence cannot price: drug cost, access, supply, and the demand for lifelong adherence. The discipline is to record that this trade-off exists and which way it runs — the continuation-dependent route is the most exposed — and to stop there, not to net it against the mortality and event findings. Whether lifelong cost outweighs a rented benefit is the person’s judgment, not a number this evidence supplies -> Which Objective Moved This Recommendation.
What the current evidence still cannot see
These are gaps — insufficient evidence, not no effect -> The Insufficient-Evidence Statement. That distinction does real work here: the trials run a few years while people take the drug for decades, and this horizon mismatch is structural — more searching now cannot close it.
- Lifetime and long-latency outcomes are unmeasured. Cancer over a full latency window, bone and fracture over decades, pregnancy and fetal exposure, adolescent development, and cardiometabolic status after discontinuation all sit in the insufficient-evidence state. For a long-latency harm, absence of a signal in a few-year trial base is exactly what you would see whether or not the harm is real — so the current quiet licenses neither alarm nor an all-clear.
- Attribution — drug-specific versus generic to the loss — is unresolved. Lean-mass loss and facial/skin adipose change may be effects of rapid, large weight loss by any route rather than of the molecule. Separating them needs a matched comparator losing the same magnitude at the same rate by other means, and none is held.
- The neuropsychiatric outcomes that matter most are the ones the design cannot see. A depression-adverse-event count is not a measure of hedonic tone — whether the drug dulls everyday pleasure (anhedonia / QoL) is unmeasured, not refuted. And because trials exclude psychiatric history, the suicidality evidence is observational only: the psychiatric-comorbid stratum the signal flags is invisible to the clean RCT design, leaving insufficient evidence where the worry concentrates.
- Unknown unknowns are unquantifiable — so name the two honest proxies. No list exists of harms nobody has measured. Two honest proxies name the shape that makes them likely and hard to catch: the evidence-horizon-versus-use-horizon mismatch above, and the moving-target problem. The exposure drifts under a constant label as prescribing moves to newer molecules and higher doses the mature record never tested. Accrued years de-risk the specific compound and dose, not the class label.
Where you net out
Start from your baseline risk, not the scale. If you have established cardiovascular disease, or type-2 diabetes with kidney disease, a GLP-1 is a large, proven lever on the outcomes that matter — cardiovascular events, death, and hard kidney outcomes — and it ranks as a Big Rock. If you are near-normal weight with no cardiometabolic disease, you are buying a reliable weight and surrogate change whose hard-outcome payoff is unproven, at the same adverse-event cost — and net serious harm runs slightly against you. Place yourself in the right stratum before deciding anything else.
Expect a maintained state, not a course. Stopping returns about two-thirds of the lost weight within a year, because the body defends the higher weight. Budget cost, tolerability, and reversibility over years, and compare that against a sustained lifestyle change or surgery on the same terms — not against an idealized one-time fix.
If you proceed, protect muscle — resistance training and adequate protein lower the functional cost without touching the weight benefit, most urgently in older or frailty-risk adults -> GLP-1 and Lean Mass.
Then weigh the trade-off yourself. The evidence names what the drug moves, in which direction, and how certainly. Weight and diabetes-progression fall reliably. Hard cardiovascular, kidney, and mortality outcomes improve where baseline risk is high. The gut cost is manageable. The feared cancers show no robust signal but no long-term all-clear. The mood signal is low-to-moderate-certainty and leans harm. The feared suicidality scare shows no robust population signal — a low-certainty adverse signal, not a proven harm, and one to watch mainly if you are psychiatric-comorbid or already on psychotropics. The evidence does not price weight change against gastrointestinal burden, lifelong cost, or an event benefit proven for one stratum and unproven for another. Bring your own weighting; the fabric supplies only the directions and the certainties.
Evidence box
Question ’For a person considering a GLP-1 / GIP-GLP-1 receptor agonist (semaglutide, liraglutide, tirzepatide) primarily for weight: what is the effect on each patient-important outcome — weight and appetite, glycaemia, cardiovascular events, kidney, adverse effects, lean-mass loss, discontinuation and regain — for whom, how large, how certain, and how does that answer change by stratum?‘ Evidence included 16 sources — 8 gold, 8 high Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note All sources are gold or high tier. Last updated 2026-10-06 · Independently reviewed: No · Full edit history