This is a more complicated question than it first seems. AI doesn’t usualy think things out – it goes and grabs a figure from Reddit or Facebook or Quora and quotes it. However, if the use prompts AI to research the topic – that is, something like “using sources like the DOE and other engineering data, what is the Natural Gas powered grid to road effiiciency of an Electric Vehicle?” , you will get a more accurate answer.

As dad might say “Ask a Stupid Question, get a Stupid Answer” – it’s interesting that Media, the Public and therefore AI seems always biased in terms of the upsides. That is, it will easily make a mistake by saying “An EV is 40 to 50% efficient well-to-wheels, but never make the same error on the downside, which would have to be “EVs are 12-18% efficient” to match the distance from the proper figures. I thought it might be instructive to give two examples of AI responses when it is properly questioned.

Example #1:
When accounting for real-world U.S. grid averages, transmission, charging losses, and vehicle operation, the true well-to-wheel efficiency of an electric vehicle powered by natural gas is 23% to 29%. Idealized engineering specifications that rely on best-case laboratory assumptions overstate efficiency, whereas factoring in operational averages properly brings the total below 30%.
(my notes: This one is very accurate).

Example #2 – Chat GPT consistently reported number vastly higher – and used pages of calculation to “prove” itself.
“So roughly 38% well-to-wheels efficiency under that particular assumption.”
is an example of initial output. It even declared that Pessimistic! In another interesting slant, it tried to compare EV’s to Gasoline cars even though this was never mentioned! That’s an incredibly bias – and, of course, it was FAR off the upside with ICE vehicles also (30%).
” The answer is approximately 50–60% WTW for a reasonable modern-grid calculation” is yet another of the sentences it returned.

So the question is why? The answers are quite simple. AI doesn’t care about truth or accuracy at all. It will spit out wild numbers and back them up with math that isn’t needed nor relevant. It even admitted this:
“using actual U.S. average heat rates and losses, rather than the usual marketing-style” – WHAT? Of course, AI is going to reflect everything we do in the USA – and everything is marketed. It also states:
“The biggest problem is that people routinely quote the most favorable efficiency number from each stage without maintaining a consistent system boundary.“
In other words, people like PR and BS, if I might be blunt.
ChatGPT will eventually, but only if sources (like the Department of Energy and the Manufacturers of the Gas Turbine units, etc. are pointed to, give a number closer to reality:
“So is there an “upward bias”?

“I think there is a very real methodological bias, although I wouldn’t necessarily call it intentional“. (that is ChatGDP talking).
and, so – finally, ChatGPT can come to reality like this:

“If we’re asking the question you actually intended:

“Starting with the chemical energy in natural gas/coal, how much ultimately becomes useful mechanical energy at the wheels of a real-world EV operating on the U.S. grid?”

I’d use:

~28–31% as a reasonable engineering estimate

with something like ~30% as the central figure.”
(my comment = Close enough – but it takes a LOT to get the Chatbot to address the actual question).