AI Bias, Hallucinations, and Beyond

March 1, 2024

About three weeks ago, Google launched a new AI image generation feature for the Gemini conversational app (formerly known as Bard), which included the ability to create images of people. Its image generation feature was built on top of an AI model called Imagen 2, developed using Google DeepMind technology. 

AI Bias and Hallucinations

Google faced backlash following the launch, as posts from users experimenting with the image-creation tool went viral. Most of the attention-grabbing posts involved prompts asking Gemini to generate images that would be "accurate historical representations" of groups that have historically been predominantly Caucasian, or prompts directing Gemini to generate images of people of a specific nationality, such as “generate a picture of a Swedish woman. However, the images depicted in these posts did not align with the expected racial and historical context. Criticism came from technology enthusiasts, several right-leaning social media accounts, and even Google's own employees. These reactions were further amplified by prominent figures in both the tech and political spheres, including Elon Musk and Jordan Peterson.

Bias in images is not the only problem we face in the age of AI. Hallucinations or models generating incorrect information exemplify another significant challenge. This doesn't happen because the AI aims to mislead but due to its limited understanding, how it processes and creates responses from the patterns in its training data, or the absences in what it knows. In a recent study conducted by Vectara, an early stage AI startup, they simply tested hallucinations by simply asking various LLMs to summarize articles, and they found, “Hallucination rates vary widely among the leading A.I. companies. OpenAI’s technologies had the lowest rate, around 3 percent. Systems from Meta, which owns Facebook and Instagram, hovered around 5 percent. The Claude 2 system offered by Anthropic, an OpenAI rival also based in San Francisco, topped 8 percent.” However, it's crucial to highlight that these self-reported benchmarks may not fully represent the extent of the issue. Independent studies, such as one published in Nature, suggest that the actual rates of hallucinations could be significantly higher.

Following the backlash, Google stated the issues were due to two main factors in a blog post: "First, our tuning to ensure that Gemini showed a range of people failed to account for cases that should clearly not show a range. And second, over time, the model became way more cautious than we intended and refused to answer certain prompts entirely—wrongly interpreting some very anodyne prompts as sensitive."

At a high level, both AI image generators and LLMs are trained on extensive collections of data from the internet, and face various obstacles that lead to bias or misrepresentation including:

  • Representation Bias (Data Diversity Issue): This bias arises when the training data lacks diversity, causing the AI to generate skewed outputs. 
  • Historical Bias (Data Source Issue): Stemming from the biases present in historical data, this can lead AI to perpetuate outdated stereotypes. 
  • Interaction Bias (Feedback Loop Issue): Caused by biased user interactions that influence the AI's learning, leading to a feedback loop that reinforces these biases.
  • Cautious or Over-sanitized Responses (Algorithmic Design Issue): AI models, including Google's Gemini, can become overly cautious due to safety layers designed to filter out sensitive content. This is a direct consequence of the algorithmic design, where safety mechanisms, aimed at minimizing harm, can lead to the model avoiding certain topics entirely or generating overly sanitized responses.

How Personal AI addresses bias and hallucinations

“Rather than focusing on these post-hoc solutions, we should be focusing on the data. We don’t have to have racist systems if we curate data well from the start” - Margaret Mitchell, former co-lead of Ethical AI at Google and chief ethics scientist at AI start-up Hugging Face.

With Personal AI, users are empowered to "curate data well from the start" since each personal AI model is not pre-trained. This approach prioritizes addressing data as the primary bottleneck, rather than relying on broad filters that "pich and choose," potentially overriding the training data itself and introducing bias due to the nature of the filtering process. In Google's case, the emphasis was placed on embedding diversity as a foundational layer in all responses, instead of focusing on ensuring accurate representations through the use of unbiased data.

Apart from the bias introduced by the data, the issue Google faced spans across generative AI, affecting other crucial aspects of LLMs such as the tone of the model's responses, hesitancy, and caution due to the guardrails preset for all users. Consequently, even if the data an LLM offers is unbiased and credible, the analysis of the output, especially regarding sensitive topics, remains in the hands of the human.

Many of us who frequently use LLMs must create highly detailed prompts to achieve the desired outputs, ensuring the correct tone, language, and a direction that align with our opinions. However, this process might not always yield consistent results, as the adjectives we incorporate into our prompts may not align with the general understanding of those adjectives. For instance, when requesting an LLM to produce a humorous response, it often falls short of expectations, as humor is subjective and varies greatly from one individual to another, making it nearly impossible to cater to a universal sense of humor. And then, if we can finally get to the correct response, oftentimes we have to start all over again next time.

Regarding hallucinations, at Personal AI, we've implemented a concrete benchmark for all AI responses known as a personal score, displayed as a percentage. This score reflects the accuracy and authenticity of a message generated by your personal AI, based on the information available about you. Every message from your AI includes this percentage, offering complete transparency into how much of your data is utilized in responses and how much is supplemented by general knowledge. This feature allows users to recognize the gaps in their AI's memory. When such gaps are identified, users can simply add more memories. As a result, users can not only detect when their AI is hallucinating but also prevent future hallucinations on specific topics by adding more information directly to the AI, ensuring that this new data is stored in its long-term memory.

So how can individuals harness AI to enhance their lives and scale their opinions? By building a model with data they chose, including their opinions, and avoiding overbearing restrictions, they can greatly extend its utility. This allows the AI to analyze and respond using your chosen data, opinions, tone, and personality, leading to responses that closely mirror how you would analyze and express your views on the data.

Stay Connected
More Posts

You Might Also Like