Isn't the volume so low on that and the latency requirements so undemanding that ML is overkill to get the answer as well as the accuracy issues with false positive for "not deadly" being somewhat asymmetric?

I still think of ML as having a niche of "mostly right is ok" and useful for very, very large volume otherwise one or more people does better for less. That may change in time. Self driving cars are very close now they say, but i suppose we've been hearing that for a few years without getting there.

I don’t believe modern state of AI is good enough for the job. At least not yet.

A human expert might look at your photo and instead of saying “you may / should not eat that” start asking questions, about local ecosystem, climate and weather. Or they might ask you to shoot more photos, e.g. bottom of the cap or microscopic photo of spores.

I’m not an expert in AI but based on what I know they aren’t smart enough for that, at least not yet.

Already exists, I heard about it on NPR the other day (perhaps the idea got implanted into your subconscious) https://mushroomai.ml/#mushroomai

Definitely a neat idea, but how would this work? Most identification starts with a spore print. What if instead of an AI to identify mushrooms it was software that used a classification system like what is described in "Mushrooms Demystified".

Edit: Should clarify most reading I have done on mushroom ID involves taking spore prints as an essential step but it's not always the rule.

What if you limited the range from I don’t know if it’s poisonous to X% likely poisonous vs is/is not poisonous?

People empirically do not understand probability in a useful way. Folks will, in aggregate, read that as "X% poisonous", not "X% chance to be kill you."

"Oh, that's only 25%, it'll be fiiine."

Even if you can avoid being liable for it, I'd think most people wouldn't want someone's death hanging over their head for reasons beyond the legal.

Yikes! Some species require microscopes to differentiate. Thanks for not making that app!