Decoding the language AI Safety: What is being said to Australians?

My name is Andrew McAlister. I'm the public officer of MapleTree AI, where we're a not-for-profit, we're a charity and we're a collective of privacy lawyers and governance practitioners. And we do our governance assessments from New South Wales to anyone that needs it. We've done a little bit of research, and I just wanted to present that to you for five minutes or so. And it's around decoding the language of AI safety and thinking about what is it that we're actually saying to people [inaudible] in Australia at this point in time.

The International AI Safety Report had 100-plus experts, 30 countries. I think they disagreed on a lot. One thing they did agree on is that constructive AI discourse is necessary so that we get the benefits of this technology and manage the risk. The report mentioned information asymmetry was one of the four challenges of AI governance. And that information asymmetry they broke down into institutional transparency versus algorithmic transparency. Institutional transparency around what institutions disclose, versus algorithmic transparency around the explainability of those systems. You can score highly on one and low on the other.

And so we're interested in that gap around what is being said at this moment in time. So we went and collected all the AI disclosures we could find, put them on our website under some explorer tabs and made them available. And I just wanted to talk about some of those results very, very quickly. Just see if you can spot a trend like we did here. We looked at the current privacy policies in Australia, mapping ASX-listed companies — about 101. Obviously the automated decision-making reforms via the Privacy Commissioner are coming in December.

So organisations are going to be a lot more accountable for ADM. But we see about 32% state anything around AI. So there's a real dearth of information at the moment. Only 3% name the technology that does that thing to the customer or to the consumer in Australia. And only 1 in 100 offer any sort of opt-out. So the transparency statements also follow a similar theme. We see about 1 in 6 are buried quite deeply in terms of reading age. None of them meet that sort of age-appropriate reading for a transparency statement — about 15 — for the public.

Most of them require a postgraduate-level understanding, which is concerning. And then within the public submissions, you see the same sort of trend coming down. People support transparency in principle, but it comes with caveats. And then around 10% of us, even those experts doing submissions to government, are able to name the enforcement mechanism that we'd like to see in place. So, decoding the language. A couple of tips here. We see a lot of new language coming in. About 1 in 10 of the words in those transparency statements are new terms, new concepts.

So this is being developed right now for Australians. And there's a lot of confusing lexical density, hedging — well, that's obvious — a lot of modal verbs and legalese in there. There's a lot of institutional centring around what organisations are doing to others. We'd see that's probably going to change once there's contestable mechanisms and opt-outs and things like that. But at the moment, I don't think those things exist. So obviously it's very much centred on what the organisations are doing to you. Not a lot of original content in some of those standards.

For example, every statement in the Department of Health's transparency statement aligns to another somewhere else. So there's zero content there. And anyway, transparency we think is a prerequisite to understanding AI safety. We're a not-for-profit. If you want to see where your organisation sits, just jump on the website, check it out and challenge the evidence. So thanks.