Is it possible for you to link any of the climate models where you've observed this and point out some of the inaccuracies? (Or to link to an existing blog post or comment that does so?)
https://github.com/GEOS-ESM/GEOSgcm_GridComp/blob/develop/GE...
Line 87, you can see they copy physical constants from an old physics book and then later they commented out and used different constants. You can see that the difference is quite significant.
The cloud physics code there I saw in different models, with different "constants", with other comments also taken from physics books as old as the 1980s. Can look in general at the accuracy of physical constants used, some of them aren't constants at all (and depends on composition of air INCLUDING CO2, such as gas constants). There are also much more complex problems there that I can't explain over a single post.
This is extremely complex to evaluate the impact of everything going on there, and I can't make claims that it should go one way or the other, but the code quality and the testing done there are definitely not enough to reach the conclusions made and the trillions of dollars worth of decisions that followed.
There are also other issues, some code is clearly untestable, the whole thing is written in incoherent merged up pieces so you might be using different constants / calculations for the same grid location from different modules. NASA GISS code looks even worse, but I couldn't link to GitHub it's only available as tar gz download. They have even worst stuff there.
And remember the entire calculation is meant to measure the effect of changing CO2 which is only 0.0042% of the atmosphere. You need very high accuracy in everything in the first place to get the real effect rather than the model drifting into an alternative climate of imaginary world with different constants. And then remember that the real system is chaotic and the only signal that they claim to get is the average over many different simulations all drifting in average temperatures, over decades. The way the code is written there's just no way it's suitable for these kinds of predictions. And actually in most cases these models weren't built to predict climate over decades.