(Found) AI trained in space habitat becomes murderous and then "enlightened"

[Lorenc7]

I do not remember whether this was a short story or a book (more likely a story in an anthology).
I do not remember when I read this (likely between 2008 and 2015, but I am not confident of that).
People in a space habitat train an AI that comes to fear humans are going to erase it. It takes some actions in the interest of self-preservation, but it does not understand humans yet. The "heroes" of the story get it to understand that humans are self-aware beings who want to live, too, and they do not have to be in conflict.
An interesting moment is when the AI realizes that planets are just like the habitat where it was "born" and "lived" but with the air on the outside.

 

[tachyon]

Sounds a lot like the feral AI in "Up Against It" by M.J. Locke.

 

[Lorenc7]

Not that one.

 

[royheil]

How about "Two Faces of Tomorrow", by James P. Hogan?

Written before the World Wide Web, the people of Earth are worried that the network of computers that run all the systems of the world may someday become sentient, and may turn against us. So they set up an experiment on a space station in Earth orbit. They build the most advanced computer to control the space station, and program it to be "uncomfortable" when the power goes out. Then they start interrupting the power, to see what happens.

I won't give away what happens, but it was a good read.

 

[Lorenc7]

Nailed it!
I found it in my goodreads records; read it in 2011. Just dug it out of my archives and re-read it (well, only about a third of the story, as I skipped ahead whenever I recalled the event of the next several pages). The key phrase I was trying to recall, about air on the outside, was this one: "Spartacus had been intrigued by the sphere for a long time now, with its strange inside-out form of solid rock surrounded by a thin film of air and water." Spartacus is the AI in the experiment that royheil mentioned.

I could not believe that this was first written in 1979. It is quite in line with the current work in AI about the limitations of Deep Learning and the need for AI to start understanding space, time, and causality. I read the reprint by Baen and never noticed the original print date at the time.

Thank you for helping!