Google, more specifically, Google Gemini, has come under fire from various users for having generated a slew of biased images and text responses to specific prompts in recent times. The firm, which has been looking to get ahead in the AI race, seems to be facing consistent challenges in its releases, given that the market has gotten immensely fast-paced with OpenAI in a comfortably dominant position. Google Gemini had enabled AI image generation capabilities in its chatbot, in an attempt to match the capabilities of GPT-4, which is paired with the advanced Dall-E 3. However, like with any AI tool, bias seems to have crept into several responses. The key complaint among distraught users was the reluctance of Gemini’s AI image generator to create historically accurate images, while also leaving out certain communities from its output.
The issues spiraled out of control as users began posting Gemini’s responses on social media platforms like X. Google’s upper management soon took notice, with its Senior Vice President and CEO addressing the matter on separate occasions and platforms. Acknowledging the problem, the latter also mentioned that the firm got several things wrong and that the issue was unacceptable. Meanwhile, other personalities such as Elon Musk also joined the discussion, criticizing the firm and Google’s large language model at large. The recent spate of problems with Google Gemini and other chatbots is largely associated with the issue of bias and hallucination, a problem engineers and scientists have found rather challenging to address.
AI Image Generators, Hallucination, and Google Gemini’s Blunders
Gemini’s issues hint at a delicate balance between user prompts and internal guardrails.
Google Gemini was the tech giant’s biggest release in recent times. With over three variations, the firm expected the LLM to be an equalizer in its ongoing rivalry with OpenAI. While Gemini remains a capable model, the responses provided to specific prompts and the impression of racial bias against specific ethnicities have considerably dented the model’s image. Google also struggled with issues following an instance of AI hallucination before the launch of its Bard chatbot (now Gemini), which hampered its prospects in the market. The recent events underscore the prevalence of biased responses in AI models and how machine learning protocols can extrapolate instructions and prompts to provide very different responses. It must also be noted that natural language processing functions to statistically place words and outputs based on their respective probabilities as opposed to “understanding” them in the true sense.
While this poses challenges in some cases, it isn’t inherently flawed, since artificial intelligence does not possess capabilities such as intuitive thought or critical thinking like humans. It remains complex to predict AI behavior, which was also witnessed with Microsoft Copilot’s strange and disturbing responses. While Google’s CEO, Sundar Pichai, did acknowledge that Gemini’s bias was unacceptable, the executive did mention that no AI is perfect and that Google is working tirelessly to fix the issues following the AI image generation feature being taken down temporarily. Regardless, these developments do not bode well for the firm, which has already witnessed considerable challenges in establishing itself in the AI chatbot and image generation markets.
Potential Causes and Fixes
Google is currently working on solving the issues with Gemini’s AI image generator.
Most AI chatbots have guardrails and guidelines embedded in their mechanisms to ensure they’re limited to certain use cases, while also preventing the protocols from generating harmful content. There exist numerous elements to training LLMs and adjusting their outputs based on successive rounds of supervised and unsupervised learning. Most image generators deploy generative adversarial networks, which are an advanced deep learning protocol using noise, predictive modeling, and two competing circuits to produce the best output possible. Google Gemini’s issues possibly stem from an overcompensation by the underlying ML algorithm to stay within its guardrails, while also producing content in line with user prompts. The overcompensation coupled with excessive caution in the case of some prompts invariably resulted in absurd and inaccurate results, which irked Gemini’s users.
It is just as crucial to note that Gemini, like all other prevalent language models, also relies on publicly available databases for its training. Biased datasets, in addition to complex guardrails that could be misinterpreted by the LLM, would’ve invariably resulted in problematic responses, causing the firm to pull the AI image generation feature from the publicly available Gemini Pro model. These issues also prompt the larger AI community to take a better look at the way AI models are designed, trained, and implemented. While AI was touted to be a game changer in many domains, it still struggles with several generative use cases. Additionally, incidents of this sort take away credibility not only from Google but also from chatbots and language models at large, now that artificial intelligence is being optimized to be included in various workflows.
The Future of Google’s Gemini AI
Gemini’s image generator will be back online in the near future following fixes to its protocols.
While the setback for Google and its LLM is major, the firm is said to be working on fixes to these problems, with the promise of bringing it back online in the forthcoming weeks. That being said, the emphasis on responsible AI has only been growing. However, it also comes with certain conundrums and issues that are difficult to address. Regulation is also consistently growing stricter, with the European Union looking to establish and implement the world’s first AI Act very soon. Firms like Google will have to remain wary of malfunctioning protocols to avoid further backlash and regulatory investigations, which might further cut into Gemini’s commercial prospects.
FAQs
1. Is Google Gemini safe to use?
While Google Gemini does have safety features imbued into its functional mechanisms, Google advised users not to share any personal information with the model in February 2024.
2. Why was Google Gemini’s AI image generator temporarily disabled?
Gemini’s AI image generator was disabled following numerous incorrect outputs that were historically inaccurate, along with responses and results that reflected a degree of AI bias. Google is currently working on fixes to resolve the issue.
3. What language does Google Gemini use?
The publicly available edition of Gemini uses the Gemini Pro language model, which is larger than the Nano variant yet smaller than the top-end Ultra model.