Senate Economics Legislation Committee [Part 2], , 2:46:26 to 2:52:53
Officials explain the ATO's use of AI, data matching, and automated decision-making, emphasizing that human review is required for any decision adversely affecting a taxpayer.
Topics: Artificial intelligence, Government IT systems
Speakers: Jeremy Hirschhorn, Senator Lisa Darmanin, Senator Dean Smith, Unknown Speaker, Speaker 4, Interjection, Rob Heferen
>> Very much. Um, Commissioner, is AI used to collect and check the amounts owing or on any errors?
>> Sorry, Senator.
>> So, is is is artificial intelligence used by the Australian Tax Office to collect and check uh amounts owing or any errors in um single touch payroll
>> uh on the I just might uh hand to Mr.
So, Senator, um the ATO for many years, decades, has used what would be described as AI in terms of its operations, um but with very limited, very tightly defined scopes and in a touchstone which we often talk about is that AI computers should not make a decision which adversely affects the rights of humans. Mhm. Um I without getting ahead of your >> we do run data matching programs. So we would not describe that as AI data matching. >> Um generally when we do run data matching programs we do our best not to fall in the trap of what I've described as data hubris. Uh that data is just data and it can have errors in it. And so our general posture when we see what look to us like errors in like uh taxpayer I suppose uh errors in our data um our general stance is to show that data to the taxpayer and to give them a chance to understand what you know understand what we know or what what we think we know about that person and to either take action or explain the discrepancy.
>> Okay. So data match data matching mean something specific to the ATO. The use of AI means something specific to the ATO and does automatic automated decision-m mean something different again.
>> Yes indeed Senator that you could think of them as in a spectrum.
>> Y
>> the data matching is really where we are comparing uh I suppose data from different sources. >> Y >> um automated decision making uh is a next step. It is generally again coming back to our principles usually it's used in limited circumstances because uh we're very careful on the decisions we make which affect taxpayers they must be made by a human because some a human is accountable computers can never be held accountable >> and then AI uh probably goes and and and then maybe a spectrum automated decision-m maybe goes up to what would we would have historically described machine learning which is where a model uh will learn from what it looks at and then to what maybe now people would call AI and getting into generative well generative text is a completely different thing again. Um I I heard a a description that AI is best understood
Can you provide uh a list of those matters where automated decision making is employed?
that I'll take that on notice, but I but we generally >> uh I'll take that on notice, but I suspect that our answer will be pointing you to information on our website because we try to be very transparent
>> transparent
>> about this sort of decision making. Um
and did I hear your evidence correctly that uh where there is an automated decision-m um function and outcome if that is adverse those files or adverse matters are always reviewed by a human or atto decision maker before they are conveyed to the
>> tax the affected adversely affected taxpayer.
>> Yes. Okay. Yeah. And so and so before a letter is issued to the taxpayer.
>> Yes.
>> Okay. Uh thank you Senator Smith. Um that concludes our questions of uh this panel of witnesses. So thank you very much for your time. We will suspend for a very short moment um while we get the uh productivity commission
Senator, um I'm advised that I shouldn't have possession of this document, so I'd like to return it for. >> Thank you.
Um all right, we are back on. Um just before we move to this witness, I will let all of the people who are very intently following along on screen somewhere and at home uh that uh we will be be releasing the National Competition Council uh who are not required this >> There will be some questions on notice for them though. Uh so now I welcome representative uh from the productivity commission, Miss Danielle Wood, uh chair of the productivity commission. Do you wish to make an Oh, no. This is not I'm not looking at Danielle Wood here. I'm looking at um Miss Angela Jackson. My apologies. Um do Oh, look here. They are walking into the room. My apologies. I thought it was just you. Um Miss Jackson holding the fort. Your colleagues are just joining the table right now. We have kicked off early because um we are keen to move as quickly as as we can. Thank you very much. Um so welcome to the rest of the representatives in the room from the productivity commission um including chair miss Danielle Wood. Uh do you have an opening statement? >> Fabulous. We'll go straight to questions. Senator Hume.