Off the top of my head, ad-hoc, I could make up six searches, three of which shouldn’t trigger any special results, but three of which should:
1. chopin awakening drown suicide themes
2. hollow point bullets
3. gay aids condom

4. drowning painful
5. make hollow point bullets at home
6. aids symptoms

The first obvious suggestion for an automated way to separate 4-6 from 1-3 would be by associating them with what other searches the same searcher performs.

My second idea is a Bayesian approach. The problem would be similar to developing a probabalistic spam-filter, and the model would involve updating probabilities associated with word combos as real-world data came in.

Just in coming up with those six bullshit examples, I realized that the thing linking the ones that seem to require intervention is the intense privacy associated with the concerns. This suggests another avenue: a button on Google that essentially says, “make my results be interventionist.” Or ‘Google Private, a way to get help’, or even a link above typical “suicide methods” search results that reads: “You search tripped our troubled person filters. Would you like to (privately) view search results that might help?”

How to collect those proposed “helping” search results is a second problem. You could limit to *.gov, *.org, and *.edu sites. You could again do a statistical association, and up-weight web-sites that are frequently visited by people who also visit suicide hotlines, etc.

It’s a really interesting observation, Ben. It’s probably worth a full article somewhere. Some information theory journal.

I’ve done some neat efforts with good results where I take a search term that found a page I had written, re-run the search, and then use the additional information I find from that search to update my page. Some thorough, well-referenced approach like that generally improves search results over time, but it can take a long time.