People may have compatibilist intuitions, but start digging into what they they think past their initial intuitions and you tend to get completely different responses. Most people never think about what free will implies, and so their intuitions are fairly uncorrelated with what they'll insist on if you start questioning them about details in my experience.
Their roomba doesn't have thoughts, beliefs, intentions or the ability to learn in any meaningful way.
Ask yourself to what extent a person who lost the ability to form new memories should be responsible for breaking a law that changed after their injury.
What's a thought and why should it matter for free will?
What's an intention, and how do you know roomba doesn't have intentions but people do?
When roomba tries to go forward but can't and changes direction - how is that different from when a human does it?
In the end all the differences are in your model of reality and none of them are in the actual reality. You categorize roomba decision process as "lacking free will" and your own as "having free will". But there's nothing in the actual data justifying that distinction, it's just artifact of our simplified model of the world. Sociall interactions were so important for us that we got huge coprocessors in our brains dedicated to recognizing and simulating decisions processes going on in other humans.
We lack such coprocessors for other decision processes like ai or corporations, so we intuitively feel they are qualitatively different. But there's no data to back that intuition.
> Ask yourself to what extent a person who lost the ability to form new memories should be responsible for breaking a law that changed after their injury.
I could ask why a memory that you form by changing weights on neurons count for a purpose of having free will and memories formed by switching transistors don't. But that's again irrelevant. You can have free will without being able to form memories at all.
Besides in all law systems I know you can be punished for a crime you weren't aware even is a crime.
> The only reason people add these conditions is to be able to continue living in an illusion of free will and of being something more than algorithm :)
Not at all. A sorting algorithm can't learn to not sort. A learning algorithm can learn to choose between right and wrong. That's why it has free will.