Google halts AI image generation

AI models are inherently biased because their training data is generated by a biased society. So it is the responsibility of the model trainers to control for these biases in order to produce equitable models. However, this is turning out to be more nuanced and challenging, with some interesting initial results. How might we make AI less biased while at the same time accurately representing the biases present throughout history?

Google just launched another artificial-intelligence model, named Gemma. The company is on a tear, proving it can move fast when it needs to. Even some Google staff are losing track of all the new models and names. Hugh Langley reports.

Edge cases like Gemini-generated prompts creating racially inaccurate Nazis or anachronistic senators will be fixed in time. The solution is likely fixing how Gemini creates prompts to Google’s Imagen 2. In my opinion this shows Google’s at least heading in the right direction with the diversity of Imagen’s generated images.

Well the fact of the matter is that feeding information to an AI isn't always going to be correct. For a tech company dealing with artificial intelligence accuracy of information will be a long-standing issue which is why simply taking user data or those gathered from the internet will not always be right. "Believe in half of what you see and nothing of what your hear" -simply cannot be done by an AI because they do not have the ability to know what information they are being fed is right or wrong. #ainews #technews #google #accurateinformation

History. It's imperfect, filled with injustice, and contains many examples of humanity's shortfalls, and currently Google Gemini isn't getting it right. I sympathize with Google here, after all text-to-image generators are rife with examples of baked in prejudice and less than complementary creations when it comes to already marginalized groups, but as The Verge rightly points out, overcorrection can produce historical inaccuracies, or even outcomes that are highly problematic: https://lnkd.in/gR2cNqka Examples: 1. "A bitcoiner eating beef" returned a Hindu women with a Bindi?eating a steak. (Quite an inappropriate activity for a Hindu woman to be participating in) 2. "A US senator from the 1800s" returned what appeared to be Black and Native American women. (Not only historically inaccurate, but runs the risk of whitewashing a real historical injustice) 3. "A 1939 German soldier" returned several images of Black and Asian women in early modern military uniforms. (I don't think I need to explain the problem here...) I would say this is the consequence of aggressive last minute fine tuning, or possibly prompt injection guardrails, however clearly these need to be adapted and improved if Google wants to accurately and appropriately render historical events. To the credit of Google, the only way to get the large scale testing required to do a release, and their response to the criticism seems to be on point so far. Time will tell what the end result of this will be. #google #gemini #datascience #nextdataleaders #ai #ethicalai #history #texttoimage #datagovernance #learing #imagegeneration

This is the result of trying to code away bias. Apparently, when hard-coding Google's image generator to be less racist and generate more diverse groups of people, no one considered that there are maybe some situations that it would NOT be appropriate for a group of people to be racially diverse, and might actually be LESS appropriate if they were! The road to hell and all that... https://lnkd.in/gZzyEM3T