On the measured evidence they look good for it rather than bad. Pooled across eleven randomized trials in adults with type 2 diabetes and fatty liver disease, adding a GLP-1 lowered ALT by 6.71 U/L at moderate certainty. HbA1c fell by 0.48 percentage points, graded low certainty — the two numbers do not carry equal weight, and the review says so.
Liver cancer is the outcome people mean when they ask this, and it has been measured twice with the same shape of answer. A target trial emulation found five-year hepatocellular carcinoma risk LOWER on a GLP-1 than on every comparator, by 0.37 to 0.52 percentage points [2]. Those differences stay under 0.6 points because the cancer is rare, which is the reading in half a percentage point over five years.
A second study found liver cancer slightly HIGHER and dismissed it for being negligible, with a risk difference rounding to zero [3]. That reasoning is right, and applied to the same study’s benefits it makes those real and small rather than transformative.
Most reviews end by telling you which option is best. This one ends by explaining why it will not, and the explanation is more useful than a ranking would have been.
What was measured
Eleven randomized trials in Asian study settings, in adults with type 2 diabetes and metabolic dysfunction-associated steatotic liver disease, all adding a study drug to comparable background glucose-lowering therapy. [1] 692 participants were randomized into eligible comparisons and 642 entered the coprimary analyses. The coprimary outcomes were change in ALT, a liver enzyme, and change in HbA1c.
Adding a GLP-1 reduced ALT by 6.71 U/L, 95% CI −11.53 to −1.89, graded moderate certainty. It reduced HbA1c by 0.48 percentage points, 95% CI −0.83 to −0.14, graded low certainty. Adding an SGLT2 inhibitor reduced ALT by 7.56 U/L, 95% CI −11.24 to −3.88, at moderate certainty, with little or no clinically important further effect on HbA1c.
Where the endpoint was the liver itself
A separate 2025 meta-analysis went past the enzyme to the tissue, pooling 13 phase 2 and phase 3 trials with 1,811 participants, four of which diagnosed the disease by biopsy and nine by MRI [4]. In people with steatohepatitis and moderate to advanced fibrosis, GLP-1 drugs beat placebo at resolving the inflammation, an odds ratio of 3.48 across three trials. They also beat it at improving fibrosis, at 1.79, and the authors single out semaglutide 2.4 mg a week among the drugs tested.
The same paper marks where the benefit stopped. In the one trial of people whose disease had already reached compensated cirrhosis, semaglutide did not resolve the steatohepatitis or improve fibrosis against placebo, and the authors say the effect on liver-related events over the long term still needs study.
Two grades for one drug
Notice that the GLP-1 estimates carry different certainty ratings. Moderate for the liver enzyme, low for the blood sugar measure. That is not an inconsistency; it reflects how many trials contributed to each comparison, how consistent they were, and how precise the pooled estimate is.
A summary that reported both numbers without both grades would treat them as equally settled. They are not, and the authors went to the trouble of grading them separately — which is the sort of distinction a pooled figure usually hides.
The three refusals
First, no class ranking. The authors state the available evidence does not support ranking SGLT2 inhibitors, GLP-1 agonists and pioglitazone against one another — several comparisons rested on sparse evidence, and the two classes were not tested head to head.
Second, no claim about durability or clinical significance. Follow-up was short and the endpoints were surrogates, so the paper says both of those remain uncertain rather than projecting forward.
Third, and most striking: comparative safety could not be assessed reliably, because adverse-event reporting across the trials was heterogeneous and incomplete. Eleven randomized trials, and the authors could not compare how safe the drugs were relative to each other. That is a statement about the literature rather than about the drugs, and it is the kind of thing that never survives into a summary.
Why a refusal is worth reading
Because the demand for rankings is what produces bad ones. A reader wants to know which drug to take, and a marketer wants to say theirs. A review that declines to answer is doing the harder and more honest thing — a comparison nobody ran cannot be reconstructed from two separate ones.
The practical version for anybody reading a claim about these drugs and the liver: ask which endpoint moved, how certain the authors were, and whether they compared the drug to the alternative or only to background care. Almost nothing published by a seller answers any of the three, which the disclosure scorecard measures from the other end.