What Did Go Wrong with Google's Gemini? (and what we can learn from it)
Google's Gemini was taken down after woke criticism as it was unable to generate any real images of historical figures and excluded White people. But what did Google actually do wrong?
In a recent blog post, Google addressed the ultra-woke issues of Google Gemini and apologized for it. Here’s a recap of what happened that caused the problem.
Google wanted to avoid the trap of previous models that produced general prompt images mostly of white people. If you wanted an image of a programmer or someone watching TV, models usually tend to create a White character. If you wanted an image of 30 workers in a construction site, the same story. They usually fail to keep the balance. Google took measures to impose balance, but their poor execution caused another imbalance, on the other side of the spectrum!
The measures they took to ensure the model’s outputs are inclusive, turned out to make them emphasize one range of people and totally ignore another range.
“This wasn’t what we intended. We did not want Gemini to refuse to create images of any particular group. And we did not want it to create inaccurate historical — or any other — images.”
Does the blog post contain any new info? Nope! It explains or mostly confirms what was already discussed in social media.
The Ominous Gemini
Gemini seems to be the loose tooth of Google’s endeavor towards AI. It has caused problems before. A few weeks back, Google released a demo of working with Gemini. Remember the blue duck?
It was an impressive video, the user gave prompts to Gemini using only voice, it replied in voice, and the whole response time was so fast and short that it was almost a real-time conversation. The video showed Gemini even understanding the video and telling what the user was drawing in real time. There was only a little problem with it. Those impressive things I listed, were FAKE!
They gave the model screenshots of the video, it wasn’t real-time and the prompts were not the user’s voice, but text. Someone had to sit down, take screenshots, and give them to the model with text prompts and then voice over the output. This was probably the poor marketing of the people responsible for making the video. But in any case, Gemini has a very unfortunate record for Google.
But AI has Gone Bad Before!
It’s true. The same thing, only the opposite, happened to Microsoft’s Tay. Microsoft released this AI chatbot. It started tweeting and responding to other people’s tweets and interacting with them. But while Gemini gone woke, Tay gone Fascist.
The ChatBot was so loose and uncontrolled, that it learned whatever people taught it on social media! In only 24 hours, the Twitter culture turned Tay into a Nazi Fascist who thought the holocaust is a lie and genocide is a moral action. Microsoft eventually had to put Tay down.
So What is There to Learn About This?
Google is a gigantic corpus of resources, and for something like this issue to slip through their fingers is sloppiness. It teaches us that “Responsible AI” shouldn’t bend backward (in a Disney fashion) to be inclusive, and while the matter of including all people from all ranges is moral and logical, there should also be measures to ensure the model doesn’t stray away from reality.
Everything must be put to check and not only a subset of criteria. Google has put checks to ensure an image of a man walking a dog can be of a Chinese person or a Black person, but the fact they didn’t check the outputs to be of any White person is, in fact, clumsiness and not “being Responsible.”
I agree that Google was sloppy. But you can kind of see why. They saw a problem. They applied a patch. They checked (probably) and didn't see the problem anymore. They concluded they had fixed it.
What they failed to do was to consider the issue of trade-offs. When you have more of one thing, what will you then have less of? How much less? This kind of contrarian thinking usually takes some experience and some time to consider and test. They lacked at least one of those attributes. This being AI, I would expect they lacked both.
There's a lot of that going around.