I've had a similar notion that Time() is a necessary test function. Maybe it's because of the limitations of human cognition. (We have biases and blind-spots and human intelligence itself is erratic.)

I find it's helpful to avoid conflating the following three topics:

/1/ Is the tool useful?

/2/ At scale, what is the economic opportunity and social/environmental impact?

/3/ Is the tool intelligent?

Casual observation suggests that most people agree on /1/. An LLM can be a useful tool. (Present case: someone found a novel approach to a proof.) So are pocket calculators, personal computers, and portable telephones. None of these tools confers intelligence, although these tools may be used adeptly and intelligently.

For /2/, any level of observation suggests that LLMs offer a notable opportunity and have a social/environmental impact. (Present case: students benefitted in their studies.) A better understanding comes with Time() ... our species is just not good at preparing for risks at scale. The other challenge is that competing interests may see economic opportunities that don't align for social/environmental Good.

Topic /3/ is of course the source of energetic, contentious debate. Any claim of intelligence for a tool has always had a limited application. Even a complex tool like a computer, a modern aircraft, or a guided missile is not "intelligent". These tools are meant to be operated by educated/trained personnel. IBM's Deep Blue and Watson made headlines -- but was defeating humans at games proof of Intelligence?

On this particular point, we should worry seriously about conferring trust and confidence on stochastic software in any context where we expect humans to act responsibly and be fully accountable. No tool, no software system, no corporation has ever provided a guarantee that harm won't ensue. Instead, they hire very smart lawyers.