Is X/Grok the Free Speech AI?
Grok 2.0 Release Newscard
Alliance for the Future is a nonpartisan 501 c(4) which advocates for startups, open research, and free speech in AI.
2 minute read.
xAI, a branch of X (formerly known as Twitter) released a new AI model on Tuesday. Grok 2.0 is competitive with other top models such as OpenAI’s ChatGPT or Google’s Gemini.
To generate images and video, xAI partnered with Black Forest Labs, a startup backed by Andreessen Horowitz, General Catalyst, and MätchVC.
Unlike those models, Grok is far more flexible in complying with politically sensitive requests:
https://x.com/EuroWynner/status/1823723555806388389
https://x.com/SwannMarcus89/status/1823890284373004685
https://x.com/eyeslasho/status/1823763292461949160
https://x.com/jules_su/status/1823892222082732373
While distasteful, all of these depictions appear to be “awful but legal” speech protected under the first amendment and fair use. This follows Elon Musk’s criticism of “woke” censorship in AI.
Gemini, Google’s language model, was heavily criticized earlier this year for excessive political correctness, refusing to depict white people in overwhelmingly white historical periods.
Our original reporting showed that Google intentionally and systematically modified Gemini to conform to these political biases. According to an engineer interviewed by Pirate Wires, “we spend probably half of our engineering hours on this.”
Anjney Midha, a Black Forest Labs board member and investor at Andreesen Horowitz, contrasted FLUX/Grok’s image generation to Gemini’s:
https://x.com/AnjneyMidha/status/1823634936416625028
FLUX appears to greatly outperform comparable models such as Stable Diffusion or DALLE at accurately generating text in images.
xAI may be held liable for a different kind of image generation: copyright. It freely reproduces copyrighted characters from notable franchises:
https://x.com/JungleSilicon/status/1823848561575899178
https://x.com/binarybits/status/1823759192924344754
Meanwhile, other models such as ChatGPT will initially refuse to comply with copyrighted requests, but can be convinced with strategic wording:
To this non-lawyer, the Terms of Service seem to forbid using photos of "natural persons" without their permission and violating trademarks and copyrights. What is different is that the user assumes complete liability for any violation. It's probably safe to use U.S. politicians for most uses, but I'd be wary of using celebrities who are big on collecting money for their image use.