This sort of thing, to me, points at the deficiencies of appealing to purely statistical patterns without any guard rails placed on top. After all, statistically, the most likely relevant categories after "pre-natal" are going to be surrounding children who _were_ born safely. So if you appeal just to data alone, you won't likely solve this problem because it is relatively rare.
The biggest failure, to me, is the response of the system when the user goes out of her way to say that the "pre-natal" ads are not relevant. Statistically it assumes successful birth, but that doesn't reflect her actual intent. A simple dialog tree would maybe suffice: "not relevant" -> "this specific thing is irrelevant" vs "suppress this and all related content". One signal says maybe don't show that particular brand again, and the other says to pivot the relevance model significantly away from that topical section of the ad space entirely.
(Or if we can't manage that, maybe at least we could somewhat curtail the misbehavior of the ML systems they foist on the rest of us.)
I was challenging myself to imagine what tactics they might try to discourage others from abandoning their platform in cases of extreme grief.
That being said, I would never agree with it. Let's not pay companies to solve problems they created in the first place.
It strikes me that an ad algorithm that forces grieving people to opt out of social media at a time when their friends and family are most interested in checking in on their well-being may encourage a broader pattern of falling back on non-ad-supported communication channels in the weeks or months to come.
It seems like there's a strong case to be made here that some ad-supported communication channels should consider either an ad-free grief mode when users are in distress, or at least sell a giftable ability to use the site without ads, in which case friends of the grieving could chip in to avoid this kind of distress.