American chatbots: oversexed, overhyped and over here

Likewise, confidential documents could get hoovered up into Recall’s database, just by calling them up onto Copilot+ PC screen, security experts said. The UK’s Information Commissioner’s Office has asked Microsoft to explain itself.

But Microsoft isn’t the only major technology company that’s just shown off new AI capabilities that critics warn come with inherent risks.

A week before the Copilot+ PC launch, Microsoft’s AI partner OpenAI revealed a new “multi-modal” ChatGPT chatbot, based on a new AI model known as GPT-4o that’s capable of using not just text for its inputs and outputs, but also audio and video.

“The new voice [and video] mode is the best computer interface I’ve ever used. It feels like AI from the movies, and it’s still a bit surprising to me that it’s real,” said OpenAI cofounder and CEO, Sam Altman, at the time.

The new chatbot also surprised critics, who complained it was overly flirtatious in the OpenAI demo. And it surprised the actor Scarlett Johansson, who said she had been in talks to allow OpenAI to use her voice for the chatbot, but had never agreed to it, and was “shocked, angered and in disbelief” that the chatbot had “a voice that sounded so eerily similar to mine”.

(The new chatbot model, minus Johansson’s voice, will be one of the 40-plus new AI models appearing in Copilot+ PCs, and it’s the one Windows users can share their screen with to help escape the zombies, Microsoft said.)

A day after GPT-4o was launched, Google showed off a new multi-modal chatbot of its own, for now named Project Astra, which can take in vastly more information than OpenAI’s chatbot can, and reason over it to answer questions with up-to-the-minute information.

Google’s CEO Sundar Pichai said the company’s goal was to allow its chatbots to absorb “infinite context” whenever they’re answering a question, overcoming a limitation in current chatbots that means they often reply with months-old information that was gathered when the AI model was trained, rather than when the question was asked.

The new stuff is “impressive”, says Professor Anton Van Den Hengel, a former director of applied science at Amazon and the director of The University of Adelaide’s Centre for Augmented Reasoning.

A chatbot with infinite context could, for instance, help a cardiologist diagnose heart disease using “the absolute latest cardiology research”, he says.

Meanwhile, the move to multi-modal chatbots will help popularise AI in ways that were not possible with the text-only chatbots. Enabling voice and video chats is going to lead to an explosion of new user-generated AI applications that will match the explosion of user-generated video that occurred when YouTube appeared, he predicts.

But there’s a problem.

Chatbots from American companies such as Google, OpenAI, Microsoft and Meta are typically trained on 100 times more American data than they are on Australian data, he says, and using them will tend to homogenise Australian culture around American values.

The Australian public service is now looking into using Microsoft’s Copilot chatbot as part of its workflow, and emails, documents and even government policies written with Copilot “will embody the American training data (Copilot) has been built from” more than it embodies Australian data, Professor Van Den Hengel warns.

“We have to decide collectively whether we’re going to have our own cultural and digital identity, or whether we’re just going to be homogenised. At the moment, it’s pretty clear which way we’re headed,” he says.

Dr Dana McKay, acting dean of Interaction Technology and Information at RMIT, has similar mixed feelings about the new generation of chatbots.

Adding voice and video prompts and replies to the chatbots will make them “incredibly inclusive”, she says, and will open up not just AI but computing as a whole to a range of people who now find computers inaccessible.

Multi-modality is great news for people with low literacy or visual impairment or even dyslexia, she says, and could even lead to new forms of creativity as people figure out how to combine text prompts with visual ones.

But the extremely high cost of training such chatbots is a problem, she says. Not only will it tend to force American values on the world because it’s generally only US-based big tech that can afford to train them, it also risks replacing the old digital divide, accessibility, with a new one: affordability.

There’s a real possibility that people who can afford to access the best chatbots will get to pay with money, while people who can’t afford it will have to pay for it by sacrificing their privacy and turning their personal information into training data.

“This could actually increase the divide between the haves and the have-nots,” Dr McKay says.