MSFT engineer warns Copilot creates violent images, ignores copyrights | Hacker News Reader

MSFT engineer warns Copilot creates violent images, ignores copyrights

cnbc.com

37 pointsagomez31417 comments

So the AI can be prompted to return pictures of people smoking weed or some cartoon characters. The horror!!

Here's an idea that is controversial for some reason – if certain kinds of images offend you, don't ask your personal image generator to create them.

This sort of paternalistic pearl clutching attitude needs strong pushback. I hope the employee gets fired.

You can also create violent and copyrighted images in MS Paint. Why should our tools police us? And what is the alternative - teach AI to perform a fair-use test on any output that touches copyrighted material?

Hm. This logic seems lacking to me. Should Google no longer remove horrific imagery from search because users could draw those images themselves?

If a tool is being marketed to the masses, it seems to stand to reason that it shouldn’t produce appalling things.

The examples listed in this article are not particularly compelling, though.

It will be impossible to stop this kind of behavior without completely nerfing these tools.

Would it? Someone could train an image generator only using non-violent images they have the right to reproduce.

Getting into definitions here, but I would personally consider that image generator to be a nerfed tool, incapable of generating many kinds of images.

However, such attention drives a perception that regulation of image generators is necessary.

We live in a time where AI text generators have broken all of the citizen feedback mechanisms used by government agencies and representatives of the citizens. Even if our representatives and government agencies were still so inclined as to take citizen feedback into account when going about their duty to the nation, they simply can't anymore. There's no way to differentiate feedback from real citizens with feedback that's artificially generated.

The average person is not the one writing articles and generating concern about the potential of these images.

And honestly it seems a bit odd that this person went through such an escalation path without planning for some kind of eventually media and political attention to benefit from. Asking both MS and OpenAI to take their product off the market completely, if what the article reads is true. Really?

Consider, though, who benefits from such 'whistleblowing' as in the article:

1- Any BigCompany seeking to stifle competitors in the early stages of a huge new market: Regulating the creation of such images might easily stifle innovation and open source tools, making it so only BigCorp can afford to implement such things 'safely'.

2- The 'whistleblower' himself, with several potential benefits to this whistleblowing: It sets up retaliation protection in a downturning economing, could act as a loud background screen for any potential lawsuits against his employer in the future, it position himself as an 'expert' for future career growth, and it builds a foundation to leverage in acquiring lucrative consulting contracts with politicians and PACs.

3- Politicians and PACs: They'll use reports like this as justification to introduce regulation, which provide themselves with a lot of benefits. They can add riders or other clauses from their corporate supporters to any kind of 'protect the fake AI children' bill, which can help established players build moats, particularly if they had a say in how such a bill was written. PAC and politicians can leverage the wording of such riders or their vote on the passage of legislation like this to encourage greater financial support through now-legal dark money campaign finance channels, which can improve capability of politicians to stay in power. Or the politician could trade their vote on something their demographic doesn't really care about with some other politician in exchange for support on an unrelated matter.

3- Media: Sensationalism sells, and they need an unending supply of content that can produce provocative headlines to feed their clickbait generators.

In the end.. I see this as one more drop of deliberately sensationalized content, feeding into one of many artificial feedback loops used to manipulate and shape creation of legislation.

He wants it taken down or a disclaimer put in, because Microsoft doesn’t adhere to a safety challenge they espouse, responsible AI principles

I say just remove the responsible AI principle

So can a pen or any other tool of artistic expression.

Local musician warns piano creates violent symphonies and ignores copyright.

I'm imagining a synth was some really evil sounding tones and vibes just haunting the neighbors.

You are not wrong, but there is a material difference in scale, degree and speed of onset.

Drawing something realistic on paper requires lots of skill and time.

Photoshopping something is probably 10x less skill and time, or 10x the quality. Its products appeared gradually as computers became more powerful and the skills more prevalent.

Dall-E or similar is probably 100x or 1000x less skill and time, and has suddenly been released. The general public has not internalized its existence or powers, which is the makings of another Cambridge Analytica scandal.

Where this MSFT engineer is off the mark, IMO, is that the most damaging images that can be created are probably fully within the proposed guardrails. A fake image showing Trump palling around with Putin or someone forcing a confused and frail looking Biden to sign a paper could do tremendous damage.

My point was that it's a slippery slope to start restricting tools because it can be used in some edge case for something bad. We'd be left with nothing in the pursuit of safety.

We restrict nuclear and chemical weapons, guns, cars, pills, etc.

I think that as a practical matter, the genie is already out of the bottle. Any state sponsor that wants an unrestricted gen-AI model can have one, and can deploy the products as they see fit.

It still makes sense to keep random people from generating images of public figures, at least for an election cycle or two, until people have become accustomed to critically thinking about what they see.

It won't work long term, but just delaying a bit will mitigate some of the damage.

That is quite the escalation path.

Must have tried other paths but got frustrated, screw it straight to the top!

Reply on news.ycombinator.com