Google’s Gemini: What Is “True”? And Who Decides?

Critics of Google Gemini suggested that the chatbot should just simply “tell the truth” after it produced biased responses.

Just tell the truth – fair enough. But – what is true? And who decides?

In Western culture, we establish truth in a collaborative process. Free speech and public debate are designed to eventually arrive at a commonly accepted take on any given subject. This is because it is recognized that even given the same set of facts, different people will arrive at different interpretations of them, and that no-one’s personal truth can ever be everyone’s truth.

As opposed to society, Gemini (or any chatbot) cannot arrive at the common wisdom resulting from millions of people debating and learning from each other. Instead, it learns based on a set of training data. But one could argue, if the data set is big enough, that it is a representation of either the conclusions of public debates, or of the arguments, pro and con, in ongoing ones.

Googlers working on Gemini have argued that they had to manually adjust the bot’s responses in order to make up for biases in the training data. Google certainly knows Gemini’s training data better than I do. But I can’t help but wonder what Google’s test for bias is – how does it determine what is biased and what is not? Are there rules? Who made the rules? Or do Gemini devs decide based on their gut feeling? Either way, the decisions on what constitutes biases are subjective ones.

Why even engage in bias hunt? If any data set is big enough to be a reasonable representation of the real world, it is Google’s. Therefore, if Google just let Gemini generate answers -minus corrections for hallucinations and obvious factual errors- based on that data set, we would get responses reflecting society’s truths as crystallized in the data.

But that is not what Google is doing. It does not accept its data set as a proper reflection of culture. Instead, it lets a team of a few hundred developers decide what is true for the 300 million Americans and the 5 billion people worldwide using the Internet.

Given the extreme importance and power #artificialintelligence will have in just a few years – do we want a few thousand people at the most powerful corporations in the world decide what is true? Can we trust employees legitimized by nobody and accountable to no-one, and of whose economic and cultural incentives we have no idea?

Instead, do we need to find ways to open-source finding the truth for chatbots? Here are some ideas:

• Community Moderation and Input: If we adopted a model similar to Wikipedia, where community members contribute and moderate content, it could democratize the process of establishing truth. This would represent a broader range of perspectives and reduce the influence of any single entity.
• Transparent Algorithms and Datasets: Making the algorithms and datasets used by AI chatbots publicly available could enhance transparency and trust. This openness would allow experts and the public to identify and correct biases.
• Decentralized Decision-Making: Implementing decentralized governance structures for decision-making about content could distribute power more equitably.
• Independent Oversight Bodies: Establishing independent oversight bodies with representatives from diverse backgrounds could provide external checks on corporate decisions. These bodies would review and audit chatbot responses, ensuring they meet agreed-upon ethical standards and reflect a wide range of perspectives.

But perhaps none of this matters. Some think the market take will care of it. They expect users to abandon those bots they perceive to be lying to them, favoring those that don’t. But I fear users might not be equipped to discern bias effectively, and without transparency in decision-making they will not be able to make informed choices.