This is almost always the result of who gets included/excluded in the study. Scroll all the way down to criticisms in this link and you’ll see a reasonably good discussion of it. Basically smokers tend to be leaner and have crappy outcomes, sick people or those with pre-existing health issues tend to lose weight etc. Including “lean” people who have other risk factors is the entire effect.

For the record this same effect is present in the “is one drink a day good for you?” debate. If you include people who can’t drink because of medication or extreme illness the alcohol use looks good. Toss all the abstainers who aren’t in peak health, and alcohol use looks very bad. The results are driven entirely by the inclusion criteria.