That they can’t replicate indicates that the sampling population isn’t consistent representation of population demographics and so there’s no real signal there that can be applied to a general population like most failure to replicate studies
I saw one the other day from a professional polling firm in which exactly 7% of "the public" said they had attended a protest, no matter what the protest was about. They asked about six or seven things people might protest about, and for every single one, 7% of the panel said they'd been on such a protest. Taken at face value it led to nonsense like concluding millions of people had attended protest marches about CDBC and 15 minute cities.
Unfortunately, the panels these firms use are extremely unrepresentative of the public but are advertised as perfect proxies. People then accept this claim unthinkingly. It's a problem.
Tangentially related, Mechanical Turk and its various less well known clones which come down to the exact same thing, are increasingly the status quo for social science studies. They keep making really shocking discoveries with like 99.99999% statistical certainty given the huge sample sizes these services enable. Kind of weird how often they fail to replicate.