People Are Crashing Out Over Sora 2's New Guardrails

[highplainsdem]

Source: 404 Media

Sora, OpenAI’s new social media platform for its Sora 2 image generation model, launched eight days ago. In the first days of the app, users did what they always do with a new tool in their hands: generate endless chaos, in this case images of Spongebob Squarepants in a Nazi uniform and OpenAI CEO Sam Altman shoplifting or throwing Pikachus on the grill.

In little over a week, Sora 2 and OpenAI have caught a lot of heat from journalists like ourselves stress-testing the app, but also, it seems, from rightsholders themselves. Now, Sora 2 refuses to generate all sorts of prompts, including characters that are in the public domain like Steamboat Willie and Winnie the Pooh. “This content may violate our guardrails concerning similarity to third-party content,” the app said when I tried to generate Dracula hanging out in Paris, for example.

When Sora 2 launched, it had an opt-out policy for copyright holders, meaning owners of intellectual property like Nintendo or Disney or any of the many, many massive corporations that own copyrighted characters and designs being directly copied and published on the Sora platform would need to contact OpenAI with instances of infringement to get them removed. Days after launch, and after hundreds of iterations of him grilling Pokemon or saying “I hope Nintendo doesn’t sue us!” flooded his platform, Altman backtracked that choice in a blog post, writing that he’d been listening to “feedback” from rightsholders. “First, we will give rightsholders more granular control over generation of characters, similar to the opt-in model for likeness but with additional controls,” Altman wrote on Saturday.

But generating copyrighted characters was a huge part of what people wanted to do on the app, and now that they can’t (and the guardrails are apparently so strict, they’re making it hard to get even non-copyrighted content generated), users are pissed. People started noticing the changes to guardrails on Saturday, immediately after Altman’s blog post. “Did they just change the content policy on Sora 2?” someone asked on the OpenAI subreddit. “Seems like everything now is violating the content policy.” Almost 300 people have replied in that thread so far to complain or crash out about the change. “It's flagging 90% of my requests now. Epic fail.. time to move on,” someone replied.

-snip-

Read more: https://www.404media.co/sora-2-content-violation-guardrails-error/

---

These are the same sort of disgusting AI users who let OpenAI know, when they were surveyed a year or so ago, that they did NOT want what they generated with AI to ever be labeled as AI. Of course people like that would consider all copyrighted work their plaything.

[angrychair]

1. Genuinely surprised

These AI companies have, up until now, have given copyright owners the middle finger about the abject theft of their IP.
I am still waiting on a class action lawsuit, for trillions of dollars, on all the IP that they let their respective AIs suck up to train themselves.

I think Altman is hurting himself here though because it shows he and his company are fully aware of what they are doing and that it violated copyright laws. A larger class action lawsuit may be the the best last ditch effort we have for crushing AI before things get anymore out of hand.

[hvn_nbr_2]

2. AI has generally seemed to be completely unaware of the notion of copyright.

Once recently I googled for the definition of a word. The AI summary included a 15-words long exact, word-for-word copy of content from the Merriam-Webster website, with no credit or indication that it was a direct quote.

That's not an unusual circumstance. When all the copyright infringement lawsuits get settled, Merriam-Webster may own all the AI companies.

[angrychair]

3. So little trick I came across

On Instagram believe it or not, that if in your prompt you tell it to cite sources and only use current information and to form an answer only using sources you can cite and and information that is verified as accurate and factual. It really improves the response and it will cite it's sources for it's answers.
(Not that I use it really at all. I just tested the above as an intellectual exercise)

[usonian]

4. See my Sora post

https://www.democraticunderground.com/100220707123

Rollingstone.

[highplainsdem]

5. Just recced it.

[usonian]

6. Thanks, it's just nuts.

🥜🥜🥜🥜🌰🌰🌰🌰🌰🌰