Clearly I need to add that I'm not saying to trust everything you read on Mastodon.

What I'm saying is, even if it's all done in MSPaint, the things it has the AI responding are, well...

...unfortunately close to a few things I've read in the course of my job. (Which is not at Google, to be clear. But the backend of some AI is in the vicinity of what I do for a living, unfortunately. It makes me sick.)

How these responses describe the guardrails AI usually has is how they are described in some internal documents I have read.

Yes, I had to sign an NDA.

@foolishowl Maybe it is a prank. Everyone should decide for themselves.

Here's where it gets weird, though -- how WOULD you (or anyone) decide it for yourself? You could search for arguments for or against the ideas in the responses in those screenshots. But would you be searching on Google? Would you get results that fully scoured the internet for true and reliable sources? That's not what Google search does (or any other search, consistently).

I am not saying that means we should blindly trust this thread and its posts. It's a line in the sand that we can watch and see if it turns out to be accurate.

The thread doesn't claim the bot has access to internal Google posts. That's addressed in one of the toots.

@cobweb And the guardrails (the programmatic instructions the AI companies, not just Google, put into their LLM products to keep them on a PR friendly path, and guide users away from discovering the LLMs' real use-case) are what people need to be aware of.

The existence of these guardrails are what makes an otherwise perky and helpful "agent" an actually terrible thing, not to be trusted.

Convincing people of that is really difficult.

And that is because millions of dollars and literally untold stolen personal data has already been frontloaded into making "AI" look positive. People who are very aware of human psychology planned this.

@cobweb It IS doing autocomplete, it's just doing it with its own guardrails stripped off.

@jnfingerle @Wuzzy What's funny is that I predicted this would be your response. Meh.

@pseudonym @JessTheUnstill @Emathion That "infosec horsey" illustration is the best thing I have seen in a minute. ​