I use little clues throughout a conversation. I send out test probes – sentences where I can expect a human type response. (I treat terrible behavior as an algorithm the other person is running and my job is to stop that algorithm for a moment).
If I get less than a few human type responses within a period of time (some point that’s a little less than the projected length of a conversation, which is linked to my time and/or patience and probably to theirs too), I abandon my attempts to “find the human”.
But usually I stay around long enough to get a sufficiently high-value set of human responses that confirms that I am talking to a human capable of some human level of empathy.