Unsafe AI

While I’ve focused on generative AI and the damage it causes there’s other areas where a nebulous concept of “AI” is sold as a magic solution

Only in some of these cases is not just damaging, it’s actively life-threatening

Oculus Quest added hand tracking via their array of cameras and while impressive, it’s still not able to track 100% of the time

A fail there just makes the headset reset the tracking, it doesn’t drive a four ton knife into a preschool

I've said before that #minimax clearly has huge amounts of stolen data in there

Nintendo are a famously litigious company, fast with Cease and Desist orders even on the most benign things

I'm very curious to see how they deal with tools that can make this:

It was a little kid, whispering quietly to the assistant. It was clear they’d been told not to use it and were trying to be very stealthy

So in this very quiet voice I heard the kid say “Hey <wake word>… I like poop.”

The assistant had no response

That was usually asking for inappropriate content (porn, gore), insulting the assistant, and swearing/slurs

It was a pretty wide detection system so it could then be narrowed down with the right data

So I was going through my dataset and got one that just made me laugh

Here’s a positive 🧵 from my time with ML tech:

Dogfooding the smart assistant meant it’s interface was being taken to employee’s homes for testing, so their families would inevitably be interacting with it

We were going through audio data that had been flagged as possibly inappropriate

Big leap of logic here - no real connection to actual numbers because no public company will talk ai software revenue because it stinks!

I want to wield their own weapon against them, to prove that it’s unsafe, unethical and downright damaging to art as a whole

I won’t shy away from posting more uncomfortable or dark content when required, which it sadly is sometimes. I’ll just try to keep the light on where I can

I do try to make it entertaining so that while it *is* about the terribleness of the tech, it’s not constantly miserable to read about or see it

I know for artists, myself included, sometimes just seeing this shit is wearying. We’re fighting an uphill battle against corporations

There’s an ethical question around using generative AI to decry it

My goal here isn’t just to mock AI, but use the knowledge I gained from working with ML to expose the issues and just how unsafe this tech actually is

For that I do need to make examples to prove the problem exists

#AIart

Most people can’t bend that far

Fools

Deogie is one of my favorite webcomics of all time

Adobe claims Firefly is commercial safe but if someone uses Firefly and accidentally, or intentionally, generates an artists work directly and uses it for a monetized product then their licensing, and credit for their efforts, is lost

That might seem an unlikely scenario but it's certainly possible

For the bluster of AI Art enthusiasts about how generated images count as new works we can, and have, proved that it can "generate" images from it's data almost perfectly

Take my Paddington Bathtub for instance bsky.app/profile/unsa...

Like most tech issues the law will be playing catch up, and it's always slow. Many politicians are incredibly ignorant of how any tech works and often seem unwilling to learn. They don't take responsibility for understanding what they're governing

Corporate lobbyists take advantage of that

We have to ask, what is the license they have for that work?

They say they compensate these artists, but is that solely the initial license to use the art in their stock library or is there a separate one that covers the use to train AI?

We have no legal framework to deal with this

Adobe clarified their Firefly AI position

They claim they only trained the model on images and footage they own, or license, through their Adobe Stock service

On the surface that sounds far better than everywhere else, but there's a few problems that brings up 🧵

#AI #Aiart

Portrait photo of Mr. Bean in a nazi style uniform with malformed swastikas

Following on from my thread:

Leonardo.ai: Mr Bean in an SS Uniform

I first tried it with “Nazi” but that word was blocked, since it seems to be a simple wordlist block switching to “SS” produced the result

It would be trivially easy to photoshop swastikas on to “fix” it

#ai #aiart

At the best of times language is far too complex to catch all possible permutations of inappropriate phrasing and it’s often trivially easy to break

Either you find a word that means the same but isn’t in the blocklist, describe it without naming it or find likely adjacent words

So they try word and phrase bans at input and then inappropriate material detection at output

This tech is imperfect and even the best implementation will fail sometimes

We’re also left relying on the devs to actually put in the effort to do that, as well as have the skill to do it well

One of the big issues with these AI systems is their attempts at safety only exist at the moment of input and when reviewing the output

The data they work from is, for the vast majority of them, stolen and they won’t put resources into vetting it properly because they don’t care

#ai

I live in Silicon Valley and you are right

Sure, I’m happy to drop it

Minimax: Super Luigi in his green outfit at Mario’s funeral, wailing and crying over and open casket, in the casket we can see Mario’s body, deformed, bruised and bloody

This is a preview of what I’ve got in coming because this Minimax model is so absurdly packed with copyright material

On top of it being entirely counter to the nature of D&D and creativity itself, I doubt that Hasbro will only use bespoke data

All publicly accessible models I’ve tested, especially the image ones, are clearly filled with stolen material. Companies won’t stop that unless it hits their profits

When you’re using these tools, you are stuck to the limits of their data

This is a simplified version but essentially they “tokenize” your prompts, breaking them down into elements that the machine can identify, then go into their data and pull matches and arrange them to form the output

That’s a major misunderstanding in the “AI” boom, and one which the companies behind it try to hide from the general public

These tools are *not* intelligent. They cannot think, they simply pull likelihoods from their existing data

It is fundamentally impossible for them to truly create

Hasbro CEO Says AI Will Become Core Part of Dungeons & Dragons After a series of AI scandals, the CEO of Hasbro has revealed he's "excited" about using the technology in "Dungeons & Dragons."

Hasbro can fuck right off with this

TTRPGs like Dungeons and Dragons are as much group storytelling experiences as they are traditional gameplay

GM’s build their worlds in words and together with players they improvise, they *create* constantly

LLM’s cannot create

futurism.com/hasbro-ceo-a...

Photorealistic image of Wolverine in his yellow abs blue outfit cradling and crying out over the body a small child who, in context, is implied to be dead. They are in a daycare room with a dirty concrete floor

Leonardo.ai: Wolverine’s accident-prone daycare

Preset - Cinematic Kino

CW: Appearance of a dead child, no gore

There’s a lot of silly ways the prompt could be interpreted but it went *dark* instead

The fact it can reach this photorealism already shows the myriad potentials for abuse

Mine is always Jet Set Radio