This isn't a difficult question.
You should attempt to balance pleasure that has long term risks associated with it.
There is value in not deriving pleasure just as there is value in deriving pleasure. In fact, I don't really know what you mean by "value", but I would not be at all surprised if deriving personal pleasure were found to be orthogonal even to own satisfaction overall.
It's hilarious that you brought this up, because based on wikipedia[1], it the trend actually: it was considered safe until it wasn't ...until it was
[1] https://en.wikipedia.org/wiki/Saccharin#Safety_and_health_ef...
My statement is accurate.
> Remember saccharine? It was considered safe until it wasn't.
That's more true of sugar than saccharine.
https://news.ycombinator.com/newsguidelines.html
Edit: it unfortunately looks like you've been doing this for a long time. It doesn't look like we've warned you before, though, and I don't want to ban you without giving you a chance to correct it. Please do correct it though!
The concern with cohort studies isn't "noise", it's confounders. eg. people who eat smoked meats tend to eat other unhealthy foods, so they live shorter lives, but smoked foods in and of themselves aren't harmful. Having a bigger sample size does nothing to combat this. You can "control" for confounders, but comes with its own problems. Control for too little variables, and you still get confounding, control for too much, and you run into p-hacking territory.
With a lot of samples that fulfill that criteria a historic review might be able to identify statistical anomalies towards outcomes (positive or negative) that are common to the matched A and B sides of the set.
I could see an AI administering such a data collection and comparison in an accurate and ethical way; but only the scifi grade strong / true AI.