Google's Historically Inaccurate AI Explained
An explainer about how this could have happened and how to understand most machine learning and artificial intelligence
Google made headlines this month when Gemini, Google’s artificial intelligence (AI) product, image-synthesis feature, would produce historically inaccurate images that even would go against the prompt provided at times. Google has since paused the image-synthesis feature so it can address the issues and re-release it at a future date.
Conservative conspiracy theories were abound that said it was some plot against white people, which is as ridiculous and dumb of a notion to suggest. Conversely, others have noted that these historically inaccurate depictions also have the consequence of erasing the history of discrimination. In other words, no one is happy or likes this.
What is likely to have occurred is Gemini’s algorithm over-correcting based on system prompts that aim to reduce historically inaccurate racist and sexist depictions.
Notably, the same individuals who claimed a conspiracy against white people are not the people who have made the same criticisms about racist, sexist, or historically inaccurate non-white people. We can assume these people complaining about inaccurate depictions of white people do not think the same about inaccurate depictions of non-white people because they believe in the racist and sexist imagery.
AI/ML model/algorithm developers, including OpenAI, Meta, and others, build in system prompts to their models to correct for racist, sexist, prejudicial, and generally inaccurate images.
Prompt Injections
One reality that AI developers caught onto quickly in these past few years is that no matter what, people will attempt to use any form of AI model for malicious purposes or in ways the developer does not want them to. This can include illegal activities such as writing a phishing email or using the image-synthesis feature to produce racist, sexist, or prejudicial imagery.
The Internet is racist, sexist, prejudicial, and inaccurate
If you leave an AI/ML model to train itself on publicly available information and data on the Internet, it will quickly turn racist and prejudicial. This trait was long known before our current AI rush and has been seen for years in chatbots. If models are not built to correct this, this will increasingly produce inaccurate and false information, which would compound in time.
A Primer on AI and Machine Learning
Presently, there is no such thing as real AI. In reality, everything we refer to as AI is machine learning (ML). Machine learning has been around for a long time now, but it was not until we had the proper algorithms and understanding of AI/ML, the big data to train the AI/ML, and the affordable computing power to run these AI/ML models. When you boil down these algorithms to their basic points, they are large scale linear algebra equations. The prompts you provide can be imagined as a math problem where the solution or answer is the output you requested.
This is what has produced the boom over the last couple of years.
What Caused Gemini’s Historical Inaccuracies?
Since the massive rush to AI/ML following the success of Microsoft-backed OpenAI with ChatGPT and image-producing AI/ML models, every major tech company is trying to get in the game and catch up. This includes Google, with Gemini (formerly known as Bard). Ultimately, not all algorithms and AI/ML models are the same. In the competition to produce similar or better models, unintended consequences are due to happen.
As a result, borrowing the example of an algebra equation again, Google likely had their system prompts that reduce prompt injections or racism and sexism with greater exponential weight or with an additional +1.