Google exec left red-faced as chatbot interjects in Senate hearing

Senators grilled Lucinda Longcroft after her ‘Google Assistant’ interrupted – and appeared to offer instructions – during her opening statement on the adoption of AI | Watch the video.

Google’s director of government affairs and public policy for Australia and New Zealand, Lucinda Longcroft. Picture: NCA ewsWire / Martin Ollman

An errant chatbot that loudly interjected during a Senate hearing into artificial intelligence (AI) has led to Google’s grilling over safety, privacy and “creative accounting’’.

Google’s director of government affairs and public policy for Australia and New Zealand, Lucinda Longcroft, was caught off-guard when her loud-mouthed “Google Assistant’’ interrupted – and appeared to offer instructions - during her opening statement to Friday’s Senate hearing on the adoption of AI.

“AI will be vital for all organisations in our society,’’ she told senators.

“At Google, we take a responsible approach at each stage of the AI life cycle.

“This approach starts with research to inform model development, that encompasses topics such as data protection and adversarial testing.’’

Senators were startled by a male voice with an American accent, heard over the remote computer link, interrupting Ms Longcroft’s speech.

“Let’s explore responsible AI across the life cycle,’’ the AI assistant said. “That’s a great statement – it highlights the ….’’

Ms Longcroft laughed and said, ‘There is Google stepping in to be responsible.’’

But some senators were less than impressed, with Independent Senator David Pocock asking: “What interrupted you earlier when you were talking and we heard a voice?”

“Is that AI helping you on this call?’’ he said. “Have you got a screen there that’s helping with your answers? (Is it) listening in to this conversation and prompting?’’

Google’s director of product management for AI project Google DeepMind, Tulsee Doshi, told senators the voice “was simply the assistant going off in the background.’’

Committee chairman Tony Sheldon, an ALP senator, asked Ms Longcroft to hand over any transcripts the AI bot had generated before or during the hearing.

“There’s a serious question about what is preparation, and what are genuine answers to question, and what (are) ingenuous answers to questions,’’ he said.

Senator Sheldon asked Ms Longcroft if she had used AI to prepare for her Senate appearance and predict what questions she might be asked.

She replied that “the intelligence was not artificial’’.

“I and my colleagues went through the transcripts of the Hansard of the hearings of this committee that have been held in past weeks,’’ she said.

“We have engaged with policy makers and regulators to understand the key concerns that will be before this committee, in your terms of reference

“We applied the collective intelligence of our global teams as they deal with policy makers like yourselves in jurisdictions around the world that are grappling with these issues in how to govern AI responsibly, in order to be as prepared as possible and as helpful as possible in answering your questions today.’’

Ms Longcroft said the Google Assistant built into Google products, apps and phones uses machine learning.

“I can categorically state that AI has not been involved in the answers to any of these questions here today,’’ she said.

“I turned my phone off so that the assistant wouldn’t try to be helpful throughout this hearing.

“So aside from the small interjection from the assistant while I was delivering up the opening statement, the assistant was not active at any point in this testimony.

“AI is embedded in all of Google’s tools, whether it’s search or maps or assistants, so it is helping us and all Australians that use our products and services every day in finding information.’’

Both Google executives told Senators that the tech giant was scraping “public information” – including Facebook and LinkedIn posts, YouTube videos and people’s blogs – to train its generative AI.

“How do you let us know whether photos of my kids are going to be used or end up somewhere in one of your (AI) models?’’ Labor Senator Lisa Darmanin asked.

“How do you let me know that, and how do you take steps to exclude that?’’

Ms Doshi, of Google’s DeepMind project, replied that “while we do pull publicly available data from the web to train our models’’, online publishers could “choose to remove or opt out of that content being part of the publicly available dataset’’ through Google Extended.

Adults could also use Google’s “results about you” page to “take down specific content that is about you’’.

Senator Pocock asked why Google did not pay tax in Australia on $7.1 billion worth of annual gross revenue classified as “advertising resale revenue’’ in 2022.

“I have so many people raise concerns about seeing $8 billion come from Australians through advertising, and then the company gets to ship $7.1 billion offshore to a parent company,’’ he said.

“Is the parent company a Singaporean company? Where does that $7.1 billion in 2022 go?’’

Ms Longcroft said she would have to take that question on notice, but told Senators that in 2023, Google had paid $332m in tax on a pre-tax profit of $417m and complied with all Australian tax laws.

After Ms Longcroft repeatedly refused to tell Senators what share of the online market Google controls, Senator Pocock said that 94 per cent of search queries in Australia use Google.

“What we hear from Google is that your

AI is world-leading because you’re big and you can draw on this large data set, you’ve got the best minds in the world doing it,’’ he said.

“One of the fundamental questions everyone is grappling with is, how do you trust this AI?

“Why should we trust Google’s AI if you are disputing the fact you have an incredible market share of 94 per cent when it comes to search?

“At the same time we see … around your tax some incredibly creative accounting.

“Sure you might not be breaking any laws but there’s a lack of transparency. Why should we trust your AI?’’

Ms Doshi told Senators that Google had developed AI “principles around safety, transparency and privacy’’.