The classic illustration is the Literary Digest poll of the 1936 Presidential Election. They got over two million respondents and confidently predicted a heavy Roosevelt loss. George Gallup got a comparatively tiny but well sampled set of respondents and correctly estimated the Roosevelt victory. The magazine's problem was partly sampling bias (their readers were richer than the average American, amongst other things) and partly non-response bias (they got a return rate of ~25%, which is pretty good, but probably disproportionately from the most politically engaged of their readership, who hated Roosevelt).
I don't know how bad the sample is here for extrapolating to 'all those seeking dates', but the size is a distraction - the sampling is all.
(This isn't necessarily true if you're making individual profiles, searching for rare phenomena, or something else which is more in the 'big data' than 'social statistics' space, and which genuinely needs that much training data. It's also not true if you're studying the site members themselves, for whom the site members are a perfect (non-) sample. The latter is true for the in-house data science team themselves, but we shouldn't just assume their experience will extend to the rest of the world).