Why I am afraid to talk to Bard? The AI daycare ecosystem of Google, the search giant, and Bing at I/O
At Google’s annual I/O conference today, the search giant announced that it will infuse results with generative artificial intelligence technology similar to that behind ChatGPT. Text generation is one of the ways that the company is launching an experimental version of its prized search engine.
This failure is exemplified best by Google’s work in AI language models and its chatbot Bard, especially when compared to the launch and trajectory of Microsoft’s rival Bing.
Today, talking to Bard feels like being trapped in an AI daycare. Stray too far from its index of acceptable questions, and you’ll be politely reprimanded. I am sorry, Dave. I fear I can’t do that. Even when the system is helpful, its answers are insufferably bland. I was told that trees are an essential part of the Earth’s ecology. “They provide us with oxygen, food, and shelter.” Yes, Bard. I guess. Why not shoot me in the head when you are at it?
This difference can be seen just in basic UI choices for the two chatbots. Bing consistently offers links to other sites in its answers which encourage exploration but also show it is closer to a companion than an authority. It’s open and permissive; it makes you feel like the system is somehow on your side while you navigate the web’s vast churn of information.
The sidekick that helps you escape daycare is Bing. That’s not to say it’s some semi-sentient entity or seamlessly crafted NPC. The design encourages conversation rather than shutting it down, as it creates the illusion of personality by changing the answers unpredictablely.
It isn’t saying that some of these experiments aren’t good. Many are malicious (like deepfake pornography), and many more are simply irresponsible and poorly designed (like chatbot therapists). The total amount of work contributes to the sense that there is a technological “ecosystem” of change and experimentation. A tide that Google, for all its expertise, has completely missed.
There are two main sources for this feeling. The first is a technical ecosystem that’s iterative and comparatively open. There are many important artificial intelligence models that are open source and/or shared. Even companies that are pretty closed up, like OpenAI, push through updates with impressive speed and offer enticing hooks for developers to build on.
ChatGPT: A machine learning bot to search for information about the UK’s new King – a question and a conversation with humans
A question about the event of Britain’s new King might be answered with a couple of paragraphs. If asked about ebikes, Google’s algorithms can list bullet-point recaps of product reviews published by various websites, and link to online stores where a user can make a purchase. The revamped version of search will be accessible in the US via a new feature called Search Labs, but it will not be activated by default or for all Google users.
“The technology is very early on and we will make mistakes,” said Liz Reid, vice president of search at Microsoft, who gave WIRED a preview of the new features.
ChatGPT is powered by a machine learning model trained to predict the words likely to follow a string of text by digesting huge amounts of text, including vast numbers of web pages. Additional training, provided by humans rating the quality of the bot’s responses, made ChatGPT more adept at answering questions and holding a conversation.