1900 NYC colorized by latest and unreleased DeOldify model

And the fascinating/mildly disturbing thing is that because the image looks so convincing to our expectations (which is what these sorts of models are best at achieving), everyone who sees it subconsciously accepts it as reality, even though most of these colors are probably wrong. It's not "restoring" the color, it's making it up. Even those of us who know better can fall into that trap if we aren't thinking about it.

Of course for a toy example like this it doesn't really matter. But at a large enough scale, something like this could subtly distort our collective concept of the past. More importantly, when applying ML models to much more impactful domains, they can easily create a more dangerous, subtle collective self-delusion about reality that by definition is very convincing even to its creators.

For a slightly more meaningful example: lots of people in the replies are posting pictures of their relatives, asking them to be colorized. In a very real way this model is subtly re-writing people's memories of what their loved ones looked like.