ChatGPT’s chaotic streak can be charming. Google’s new chat-style search shows text-generation technology is headed in a much tamer direction

THIS WEEK, AT its annual I/O developer conference in Mountain View, Google showcased a head-spinning number of projects and products powered by or enhanced by AI. They included a new-and-improved version of its chatbot Bard, tools to help you write emails and documents or manipulate images, devices with AI baked in, and a chatbot-like experimental version of Google search. For a full recap of the event, complete with insightful and witty commentary from my WIRED colleagues, check out our Google I/O liveblog. 

Google’s big pivot is, of course, largely fueled not by algorithms but by generative AI FOMO. The appearance last November of ChatGPT—the remarkably clever but still rather flawed chatbot from OpenAI—combined with Microsoft adding the technology to its search engine Bing a few months later, triggered something of a panic at Google. ChatGPT proved wildly popular with users, demonstrating new ways to serve up information that threatened Google’s vice grip on the search business and its reputation as the leader in AI

The capabilities of ChatGPT and AI language algorithms like those powering it are so striking that some experts, including Geoffrey Hinton, a pioneering researcher who recently left Google, have felt compelled to warn that we might be building systems that we will someday struggle to control. OpenAI’s chatbot is often astonishingly good at generating coherent text on a given subject, summarizing information from the web, and even answering extremely tricky questions that require expert knowledge.

And yet, unfettered AI language models are also silver-tongued agents of chaos. They will gladly fabricate facts, express unpleasant biases, and say unpleasant or disturbing things with the right prompting. Microsoft was forced to limit the capabilities of Bing chat shortly after launch to avoid such embarrassing misbehavior, in part because its bot divulged its secret codename—Sydney—and accused a New York Times columnist of not loving his spouse.

Google worked hard to tone down the chaotic streak of text-generation technology as it prepared the experimental search feature announced yesterday that responds to search queries with chat-style answers synthesizing information from across the web.

Google’s smarter version of search is impressively narrow-minded, refusing to use the first person or talk about its thoughts or feelings. It completely avoids topics that might be considered risky, refusing to dispense medical advice or offer answers on potentially controversial topics such as US politics.

Google deserves recognition for reining in generative chatbots’ wild side like that. But in my tests, the new search interface felt incredibly tame compared to ChatGPT or Google’s own chatbot Bard.

As the company moves the technology into more of its products, perhaps the generative AI revolution will turn out to be a lot less fun than you might expect from the early shock and awe of ChatGPT, a chatbot that has an edgy charm. Gone are the wild ravings and imaginings of powerful AI bots. In their place are new ways to populate spreadsheets, compose email pleasantries, and find products to buy.

Even if “AI doomers” warning about errant AI prove overblown, it will be interesting to watch how companies like Google and OpenAI balance the development of more powerful generative language models with the need to have them behave.

Google has invested huge sums and major resources in AI over recent years, with CEO Sundar Pichai often pitching the company as “AI first”, and the company is desperate to show it can advance the technology more quickly than OpenAI. One high-level message from Google’s stream of AI announcements was that the company is not going to hold back anymore, as it did the LaMDA chatbot that was announced long before ChatGPT appeared but not made public.

In March, some big names in AI research signed an open letter calling for a six-month pause on creating machine learning systems more powerful than GPT-4, which powers ChatGPT. Pichai was not a signatory and said in his keynote speech yesterday that the company is currently training a new, more powerful language model called Gemini.

A source at Google tells me this new system will incorporate a range of recent advances from different large language models and may eclipse GPT-4. But don’t expect to get to experience the full power or charisma Gemini can offer. If Google applies the same chaos-taming methods seen in its chat-like search experiment, it may just seem like another surprisingly clever autocomplete.