One week after its artificial intelligence (AI) algorithms suggested to users to consume rocks and put adhesive on pizza, Google confessed that it was essential to finetune its new generative AI search function. The incident underscores not only the perils connected with Google’s assertive pursuit to commercialize generative AI, but also the inherent and treacherous confines of this technology.
The AI Overviews feature of Google utilizes Gemini, a large language model (LLM) akin to the one that powers OpenAI’s ChatGPT. This model generates responses to specific search queries by summarizing online data. Although the current AI revolution revolves around LLMs’ remarkable capability to manipulate text, this feature might also serve to cast a persuasive spin on inaccuracies or falsehoods.
Harnessing AI to simplify online data can make search results more accessible, but this can be risky when online sources deviate, or when the derived information influences crucial decisions. Richard Socher, an AI researcher who launched the AI-focused search engine You.com in late 2021, states that streamlining LLMs is difficult because the fundamental technology doesn’t genuinely comprehend reality. Additionally, the fact that the internet is entangled with unreliable information further complicates matters.
Google’s search department head, Liz Reid, has acknowledged that despite rigorous testing before AI Overviews’ launch, relevant adjustments have been prompted in the wake of the ‘rock eating and glue pizza’ incidents. These include better detection of “nonsensical queries,” and less dependence on content generated by users.
According to Socher, You.com regularly avoids the kind of errors committed by Google’s AI Overviews, since his company employs numerous methods to ensure LLMs do not misbehave when used for searches. Despite these efforts, perfecting AI search remains challenging as expressed by WIRED’s findings that You.com failed to appropriately answer a question that normally confuses other AI systems.
The generative AI enhancement from Google to its most utilized and lucrative product trails a larger trend within the tech industry following OpenAI’s introduction of the chatbot ChatGPT in late 2022. Following its release, Microsoft, an important ally of OpenAI, incorporated its technology to improve its own search engine, Bing. Although replete with AI-generated errors and odd behavior, the upgraded Bing attempted to rival Google’s dominance in the search engine market.
Certain specialists believe that Google launched its AI upgrade hastily. They argue that the company should have better anticipated that some users would deliberately seek to confuse AI Overviews, especially given that results are displayed by default on their most valuable product.
Even as Google continues to refine and improve its AI technology, it is unlikely to ever completely eliminate errors – this is by nature, inherent in AI. This becomes even more complex when considering that the AI can sometimes rely on poor sources, such as websites from other regions or even defunct sites, something that could result in less valuable information being given to users.
Moreover, as Google introduces these changes, those working on optimizing content for Google’s search algorithm are still trying to understand their implications. Although industry experts and consumers may gradually get used to new Google search behaviors, the likelihood of the system making errors continues.
Search consultant and researcher, Daniel Griffin, who is developing tools to facilitate comparisons between different AI-powered search services, states that no one has found a universal solution to the problem of LLMs’ failure to comprehend truth or their propensity to fabricate information. This problem, as Griffin points out, is not solely a problem of AI but a reflection of the broader world we navigate.
