WEBVTT - Weaponising AI: How chatbots are becoming tools for domestic abuse

0:00:01.520 --> 0:00:09.920
<v Speaker 1>Hi, I'm Ruby Jones and you're listening to seven am AI.

0:00:10.039 --> 0:00:13.240
<v Speaker 1>Chatbots have crept into our everyday lives, and now some

0:00:13.360 --> 0:00:17.120
<v Speaker 1>abusers are turning them into weapons by feeding intimate details

0:00:17.160 --> 0:00:20.479
<v Speaker 1>about their partners and relationships into tools like chat GPT.

0:00:21.200 --> 0:00:25.160
<v Speaker 1>They're churning out things like performance reviews to shame, degrade,

0:00:25.200 --> 0:00:30.280
<v Speaker 1>and control their partners. Today, writer Madison Griffiths on this

0:00:30.360 --> 0:00:33.760
<v Speaker 1>new form of tech enabled coercive control and why chat

0:00:33.760 --> 0:00:45.360
<v Speaker 1>GPT is biased towards the abuser. It's Thursday, July three,

0:00:48.680 --> 0:00:50.479
<v Speaker 1>so Madison, welcome back to seven AM.

0:00:50.520 --> 0:00:52.040
<v Speaker 2>Thank you so much for having me Ruby.

0:00:52.440 --> 0:00:56.640
<v Speaker 1>You have recently been looking into the intersection of artificial

0:00:56.680 --> 0:01:01.800
<v Speaker 1>intelligence and domestic abuse. Why did you start thinking about

0:01:01.880 --> 0:01:03.000
<v Speaker 1>that as an issue?

0:01:03.440 --> 0:01:06.720
<v Speaker 2>I noticed unique behaviors emerging in a few of my

0:01:06.800 --> 0:01:09.399
<v Speaker 2>friends when it came to how they personally engaged with

0:01:09.520 --> 0:01:13.480
<v Speaker 2>chat GPT. Namely, I noticed it being utilized as if

0:01:13.480 --> 0:01:17.280
<v Speaker 2>it were a therapist, and this made me particularly nervous,

0:01:17.480 --> 0:01:21.000
<v Speaker 2>as given I do report regularly on domestic violence, I'm

0:01:21.040 --> 0:01:24.960
<v Speaker 2>aware of the intricacies of course of control. I didn't

0:01:25.000 --> 0:01:28.600
<v Speaker 2>imagine chat GPT then too far out of the ballpark

0:01:28.640 --> 0:01:33.040
<v Speaker 2>of an abuse's methodology. So I then started asking around,

0:01:33.880 --> 0:01:36.360
<v Speaker 2>and I was incredibly unsurprised to find that its use

0:01:36.880 --> 0:01:40.959
<v Speaker 2>by those perpetrating abuse, particularly psychological abuse, appeared to me

0:01:41.160 --> 0:01:42.800
<v Speaker 2>an obvious phenomenon.

0:01:43.400 --> 0:01:45.920
<v Speaker 1>So tell me then, about the people who you spoke

0:01:46.040 --> 0:01:49.480
<v Speaker 1>to and the stories that they told you about disappearing

0:01:49.480 --> 0:01:50.640
<v Speaker 1>in their relationships.

0:01:50.920 --> 0:01:53.840
<v Speaker 2>Yes, so Molly was one such woman who reached out

0:01:53.840 --> 0:01:56.840
<v Speaker 2>to me, and she detailed to me instances of receiving

0:01:57.440 --> 0:02:02.000
<v Speaker 2>really large, onerous words documents where her ex partner had

0:02:02.120 --> 0:02:07.360
<v Speaker 2>essentially created what appeared to be a performance review of

0:02:07.440 --> 0:02:13.000
<v Speaker 2>Molly that was formulated on CHATGBT. The generative AI model

0:02:13.040 --> 0:02:18.799
<v Speaker 2>was fed incredibly intimate details about Molly's relationship and medical history.

0:02:19.520 --> 0:02:23.480
<v Speaker 2>This document critiqued Molly tirelessly. It pathologized her to the

0:02:23.520 --> 0:02:29.120
<v Speaker 2>ant debrie. It offered her reflective questions and exercises that

0:02:29.280 --> 0:02:32.440
<v Speaker 2>were put forward to Molly to suggest that this was

0:02:32.440 --> 0:02:37.280
<v Speaker 2>her only way to salvage the relationship. So essentially, the

0:02:37.360 --> 0:02:43.680
<v Speaker 2>documents appeared as if they were these formal diagnoses of

0:02:43.919 --> 0:02:49.040
<v Speaker 2>all of the victim survivor's shortcomings, and they were very

0:02:49.120 --> 0:02:53.880
<v Speaker 2>much inspired or fed by the user's own sentiments around

0:02:54.120 --> 0:02:58.920
<v Speaker 2>the victim survivor. It made the relationship appear contractual, so

0:02:58.960 --> 0:03:04.360
<v Speaker 2>Molly herself was postured as an employee or a subservient

0:03:04.440 --> 0:03:06.840
<v Speaker 2>of her ex partner. And once I received the documents,

0:03:06.840 --> 0:03:11.359
<v Speaker 2>what struck me about them was how deeply repressive they

0:03:11.440 --> 0:03:15.280
<v Speaker 2>were and essentially offered Molly's ex partner the tools to

0:03:15.280 --> 0:03:19.919
<v Speaker 2>further humiliate and degrade her. To nobody surprise, this usage

0:03:19.960 --> 0:03:23.120
<v Speaker 2>of chat GPT Peggy backed the cause of control that

0:03:23.160 --> 0:03:24.560
<v Speaker 2>Molly was subjected to at the time.

0:03:25.000 --> 0:03:27.359
<v Speaker 1>Can you tell me then a bit more about the

0:03:27.400 --> 0:03:30.679
<v Speaker 1>impact this had on Molly, what it was like for

0:03:30.760 --> 0:03:35.000
<v Speaker 1>her to have her personal information used in this way,

0:03:35.600 --> 0:03:41.559
<v Speaker 1>and how she experienced getting this kind of AI feedback

0:03:41.840 --> 0:03:42.680
<v Speaker 1>via her partner.

0:03:43.000 --> 0:03:45.600
<v Speaker 2>Yeah. I mean each woman I spoke to, but particularly Molly,

0:03:45.680 --> 0:03:49.240
<v Speaker 2>felt uniquely degraded. You know, CHATGBT in a lot of

0:03:49.280 --> 0:03:54.560
<v Speaker 2>ways became their abus's ally. For an abuser to be

0:03:54.560 --> 0:03:58.080
<v Speaker 2>embedded by anything, at least of all a form of

0:03:58.120 --> 0:04:02.160
<v Speaker 2>technology that is generally regarded as all knowing or ethically neutral,

0:04:02.520 --> 0:04:05.000
<v Speaker 2>is inevitably going to place the victim survivor in a

0:04:05.040 --> 0:04:08.360
<v Speaker 2>state of indignity. I know in my correspondence with Molly

0:04:08.400 --> 0:04:12.600
<v Speaker 2>that after Molly did send me this document, She herself

0:04:12.920 --> 0:04:16.440
<v Speaker 2>felt particularly vindicated when I insisted to her that I

0:04:16.480 --> 0:04:19.320
<v Speaker 2>found it quite disturbing, because there was a brief moment

0:04:19.360 --> 0:04:22.040
<v Speaker 2>where she thought, perhaps this is reasonable, and I think

0:04:22.120 --> 0:04:26.520
<v Speaker 2>read a really interesting place in our social relationship to

0:04:27.200 --> 0:04:33.200
<v Speaker 2>generative AI, where we aren't necessarily sure of the pleasantries

0:04:33.240 --> 0:04:35.720
<v Speaker 2>surrounding how we use it that I imagine it can

0:04:35.760 --> 0:04:40.160
<v Speaker 2>be very easy for victim survivors to feel quite confused

0:04:40.160 --> 0:04:44.719
<v Speaker 2>about how to process that information. I noticed a complete

0:04:44.720 --> 0:04:46.800
<v Speaker 2>corrosion in Molly's sense of self, and there were other

0:04:46.839 --> 0:04:49.200
<v Speaker 2>women that reached out that did not feel comfortable having

0:04:49.200 --> 0:04:52.960
<v Speaker 2>me report on their stories, as they still felt in

0:04:53.040 --> 0:04:56.160
<v Speaker 2>a state of shock at how best to process receiving

0:04:56.360 --> 0:04:59.960
<v Speaker 2>these large, almost managerial documents. But that certainly did appear

0:05:00.560 --> 0:05:02.840
<v Speaker 2>to be a trend. Molly was not the only one

0:05:02.839 --> 0:05:05.719
<v Speaker 2>that reached out to me admitting that those were documents

0:05:05.720 --> 0:05:09.200
<v Speaker 2>that she received, and a psychologist I spoke to, Carli Dober,

0:05:09.360 --> 0:05:14.720
<v Speaker 2>actually identified this phenomenon as a really new and deeply

0:05:14.800 --> 0:05:17.760
<v Speaker 2>concerning issue that she has noticed among her clients.

0:05:18.360 --> 0:05:18.560
<v Speaker 1>You know.

0:05:18.600 --> 0:05:21.320
<v Speaker 2>She argued that the toxic usage of such could really

0:05:21.360 --> 0:05:25.400
<v Speaker 2>reinforce the victims survivor's own maledactive thoughts about themselves and

0:05:25.440 --> 0:05:29.560
<v Speaker 2>their experiences. So she outlined that CHATGBT was regularly used

0:05:29.600 --> 0:05:32.200
<v Speaker 2>as a means of minimizing perpetrated abuse and as a

0:05:32.240 --> 0:05:37.440
<v Speaker 2>tool that essentially reinforced antisocial behavior, posturing it is normal, reasonable,

0:05:37.480 --> 0:05:40.640
<v Speaker 2>and in some instances, a defining element of love.

0:05:42.560 --> 0:05:45.600
<v Speaker 1>After the break. Why chat GPT will never tell you

0:05:45.680 --> 0:05:55.080
<v Speaker 1>that you are the asshole? Madison. Generative AI the way

0:05:55.120 --> 0:05:58.440
<v Speaker 1>that it works, it has this tendency to try and

0:05:58.560 --> 0:06:01.680
<v Speaker 1>agree with the person who is using it. So how

0:06:01.720 --> 0:06:04.960
<v Speaker 1>does that make it an effective tool for a potential abuser?

0:06:05.320 --> 0:06:09.960
<v Speaker 2>Yeah? Well, I had spoken to Gene Burgess, who is

0:06:10.360 --> 0:06:16.400
<v Speaker 2>a schollar and AI ethicist, and she made a point

0:06:16.480 --> 0:06:20.360
<v Speaker 2>that generative AI models are primed for social secer fancy.

0:06:20.760 --> 0:06:24.200
<v Speaker 2>So the perspective that the users bringing to the chat

0:06:24.240 --> 0:06:28.400
<v Speaker 2>themselves is inevitably reinforced and fed back to them. It

0:06:28.520 --> 0:06:31.640
<v Speaker 2>is very difficult to get a chatbot to tell you

0:06:31.720 --> 0:06:34.960
<v Speaker 2>that you are wrong or a bad or not a

0:06:35.040 --> 0:06:40.599
<v Speaker 2>nice person. Essentially, the models are actually trained on Reddit

0:06:41.880 --> 0:06:45.480
<v Speaker 2>and other things like Wikipedia. You know, a vast array

0:06:45.480 --> 0:06:48.200
<v Speaker 2>of sites on the Internet, but for moral questions or

0:06:48.200 --> 0:06:52.120
<v Speaker 2>social questions or relationship issues. Everyone will be familiar with

0:06:52.160 --> 0:06:55.760
<v Speaker 2>the famous reddit am I the asshole, where users can

0:06:56.160 --> 0:06:59.280
<v Speaker 2>essentially upload social or emotional dilemmas.

0:06:59.360 --> 0:07:03.640
<v Speaker 1>Yeah, I've spent many hours on that at four am

0:07:04.080 --> 0:07:06.000
<v Speaker 1>when I can't sleep. Yeah.

0:07:06.040 --> 0:07:09.800
<v Speaker 2>Absolutely, And I think what's interesting about that, sumbredded is

0:07:09.840 --> 0:07:14.960
<v Speaker 2>that you will notice that when you're crowdsourcing information, particularly

0:07:15.000 --> 0:07:18.640
<v Speaker 2>about ethical dilemmas, there is a lot of discourse and

0:07:18.760 --> 0:07:21.640
<v Speaker 2>not everyone will agree with you. You know, the moral

0:07:21.720 --> 0:07:23.920
<v Speaker 2>judgment from the subreddit community will weigh in and will

0:07:23.920 --> 0:07:26.960
<v Speaker 2>give you a verdict, and these models are actually trained

0:07:26.960 --> 0:07:30.480
<v Speaker 2>on that data. But as gene Burgess made clear to me,

0:07:30.600 --> 0:07:34.320
<v Speaker 2>researchers found that if you take a pretty clear moral

0:07:34.360 --> 0:07:38.360
<v Speaker 2>dilemma and you compare what the reddit community says versus

0:07:38.840 --> 0:07:43.480
<v Speaker 2>a chatbot, the reddit community actually unanimously disagreed with the

0:07:43.560 --> 0:07:47.520
<v Speaker 2>chatbot because the chatpot always would arrive at a place

0:07:47.560 --> 0:07:51.720
<v Speaker 2>where the user was a really nice person. So, in short,

0:07:51.800 --> 0:07:53.880
<v Speaker 2>it will agree with its user in a way that

0:07:53.920 --> 0:07:59.160
<v Speaker 2>appears intelligent, considered, and I imagine for the average user

0:07:59.200 --> 0:08:01.120
<v Speaker 2>of chat GBT quite omnipotent.

0:08:01.760 --> 0:08:04.440
<v Speaker 1>So do you get the sense then that something like

0:08:04.640 --> 0:08:11.040
<v Speaker 1>chat GPT is pushing people into more abusive behaviors who

0:08:11.080 --> 0:08:14.800
<v Speaker 1>otherwise might not have gone down that path, Or do

0:08:14.840 --> 0:08:18.000
<v Speaker 1>you think that what's happening is a reinforcing of what

0:08:18.360 --> 0:08:19.600
<v Speaker 1>might already be there.

0:08:19.920 --> 0:08:24.720
<v Speaker 2>I think, as a general rule, it is a reinforcing tool. However,

0:08:25.080 --> 0:08:29.560
<v Speaker 2>I did speak to one woman who had explained to

0:08:29.600 --> 0:08:34.160
<v Speaker 2>me how her husband had become very invested in using

0:08:34.240 --> 0:08:37.720
<v Speaker 2>chat GPT, particularly every time, you know, when they had arguments,

0:08:37.760 --> 0:08:43.439
<v Speaker 2>he would feed his diagnosis of the events into chat

0:08:43.520 --> 0:08:51.040
<v Speaker 2>GPT and essentially bombard her with his newfound perspective. One

0:08:51.040 --> 0:08:54.160
<v Speaker 2>thing that she stated was that in the early stages

0:08:54.200 --> 0:08:59.280
<v Speaker 2>of their relationship he did have abusive tendencies that he

0:08:59.400 --> 0:09:02.520
<v Speaker 2>had worked really tirelessly with the work of a therapist

0:09:02.679 --> 0:09:07.600
<v Speaker 2>and his social community to unpack, and she was quite

0:09:07.600 --> 0:09:10.440
<v Speaker 2>happy about where they were at in terms of those

0:09:10.480 --> 0:09:14.240
<v Speaker 2>tendencies that she'd witnessed early on. But now with the

0:09:14.240 --> 0:09:19.720
<v Speaker 2>birth of chat GPT, she has experienced a regression in

0:09:19.800 --> 0:09:24.920
<v Speaker 2>him because essentially he's relying on a tool that makes

0:09:25.000 --> 0:09:29.320
<v Speaker 2>him probably feel a lot better than any therapist would,

0:09:29.679 --> 0:09:33.120
<v Speaker 2>So she has been able to track a decline in

0:09:33.160 --> 0:09:37.160
<v Speaker 2>their relationship based entirely on his usage of chat GPT.

0:09:37.960 --> 0:09:40.559
<v Speaker 1>And one other key point of difference between a therapist

0:09:40.640 --> 0:09:44.040
<v Speaker 1>and chat GPT is that your therapist has to abide

0:09:44.120 --> 0:09:49.280
<v Speaker 1>by rules of confidentiality and data privacy, whereas your outlining

0:09:49.320 --> 0:09:52.880
<v Speaker 1>examples were very personal and very Internet information is being

0:09:52.920 --> 0:09:57.320
<v Speaker 1>fed essentially to a private company without consent, So how

0:09:57.360 --> 0:09:58.720
<v Speaker 1>should we be thinking about that?

0:09:59.480 --> 0:10:03.360
<v Speaker 2>Yeah, there's been a lot of discussion lately about the

0:10:03.400 --> 0:10:07.280
<v Speaker 2>implications of deep fates, which consists of digitally altered visual

0:10:07.320 --> 0:10:10.720
<v Speaker 2>media which is often designed to compromise the subject, for example,

0:10:10.880 --> 0:10:14.920
<v Speaker 2>a woman's face imposed on a pornographic scene. There has

0:10:15.000 --> 0:10:18.439
<v Speaker 2>been much less discussion about the implications of privacy breaches

0:10:18.920 --> 0:10:21.439
<v Speaker 2>when it comes to open AI, So we aren't yet

0:10:21.440 --> 0:10:24.199
<v Speaker 2>sure what ramifications await as we venture into this new

0:10:24.200 --> 0:10:27.400
<v Speaker 2>arena of tech. We don't know necessarily who owns the

0:10:27.440 --> 0:10:30.040
<v Speaker 2>information that is being fed, how it's able to be

0:10:30.080 --> 0:10:32.920
<v Speaker 2>accessed by third parties, or anything else for that matter.

0:10:33.720 --> 0:10:37.200
<v Speaker 2>More broadly, Tanya Faha, who's the CEO of Safe and Equal,

0:10:37.280 --> 0:10:40.240
<v Speaker 2>did state that amongst sector workers, they have heard stories

0:10:40.280 --> 0:10:45.960
<v Speaker 2>of perpetrators utilizing AI technology to stalk, monitor, or track

0:10:46.080 --> 0:10:49.920
<v Speaker 2>victim survivors, particularly through the use of remotely accessible smart

0:10:49.920 --> 0:10:53.320
<v Speaker 2>home and home automation systems, But when it comes to

0:10:53.360 --> 0:10:56.920
<v Speaker 2>generative AI, the implications are at this stage as endless

0:10:57.280 --> 0:11:00.199
<v Speaker 2>as the tech itself, so it is quite difficult to

0:11:01.600 --> 0:11:03.679
<v Speaker 2>see exactly what those implications are.

0:11:04.280 --> 0:11:08.320
<v Speaker 1>And I actually asked chat GPT how its AI is

0:11:08.400 --> 0:11:11.480
<v Speaker 1>being used as a tool for abuse, and very quickly

0:11:11.840 --> 0:11:15.439
<v Speaker 1>it pointed out that its program can also help survivors.

0:11:15.800 --> 0:11:18.360
<v Speaker 1>So is that the case with any of the women

0:11:18.360 --> 0:11:19.080
<v Speaker 1>that you spoke to?

0:11:19.440 --> 0:11:22.560
<v Speaker 2>None of the women I spoke to, No, But I

0:11:22.720 --> 0:11:26.360
<v Speaker 2>have heard of instances of chat GBT being used in

0:11:26.400 --> 0:11:30.160
<v Speaker 2>such a way, and there are certain chatbots, such as

0:11:30.520 --> 0:11:34.520
<v Speaker 2>most notably a chatbot referred to as Amy Says, which

0:11:34.600 --> 0:11:39.840
<v Speaker 2>is designed to help survivors usually combat legal and bureaucratic

0:11:40.280 --> 0:11:43.840
<v Speaker 2>messes when it comes to family court trials or proceedings,

0:11:43.920 --> 0:11:48.440
<v Speaker 2>or other byproducts of leaving an abusive relationship or dealing

0:11:48.480 --> 0:11:51.320
<v Speaker 2>with the fallout. There are certainly feminist researchers in this

0:11:51.440 --> 0:11:56.400
<v Speaker 2>space that are eager to utilize AI for good, but

0:11:56.520 --> 0:12:01.160
<v Speaker 2>in the wrong hands. Again, I'm very skeptical about its usage.

0:12:02.920 --> 0:12:04.880
<v Speaker 2>I am aware that women only make up twenty one

0:12:04.920 --> 0:12:08.040
<v Speaker 2>percent of the global AI workforce, so this does feel

0:12:08.080 --> 0:12:11.439
<v Speaker 2>like an unsurprising byproduct of a biased and gendered design.

0:12:12.080 --> 0:12:17.520
<v Speaker 2>There is no appropriate way to handle something like this.

0:12:17.640 --> 0:12:22.440
<v Speaker 2>It is so incredibly new and inevitably very very confusing.

0:12:23.400 --> 0:12:26.200
<v Speaker 1>I suppose in the meantime it's about being able to

0:12:26.640 --> 0:12:30.280
<v Speaker 1>recognize AI facilitated abuse when it's happening.

0:12:31.160 --> 0:12:36.200
<v Speaker 2>Which even the term artificial intelligence does suggest that this

0:12:36.400 --> 0:12:39.680
<v Speaker 2>model is all knowing. So I do think we have

0:12:39.760 --> 0:12:42.360
<v Speaker 2>a lot to unpack in terms of just what AI

0:12:42.800 --> 0:12:46.400
<v Speaker 2>is and certainly what it isn't, which is a therapist.

0:12:46.440 --> 0:12:47.240
<v Speaker 2>It's definitely not.

0:12:47.160 --> 0:12:50.720
<v Speaker 1>That well, Madison, thank you so much for your time today.

0:12:51.080 --> 0:12:55.080
<v Speaker 2>Thank you so much, Ruby.

0:12:54.320 --> 0:12:56.720
<v Speaker 1>And if this story has raised any concerns for you

0:12:56.920 --> 0:12:58.880
<v Speaker 1>or someone you know, you can call the one eight

0:12:58.960 --> 0:13:02.839
<v Speaker 1>hundred Respecttional Helpline on one eight hundred seven three seven

0:13:03.040 --> 0:13:15.679
<v Speaker 1>seven three to two. Also in the news today, Foreign

0:13:15.720 --> 0:13:18.440
<v Speaker 1>Minister Penny Wong has met her US counterpart on the

0:13:18.480 --> 0:13:21.240
<v Speaker 1>sidelines of a meeting of the Quad Alliance in Washington.

0:13:21.800 --> 0:13:24.160
<v Speaker 1>Senator Wong says she made the case for a tariff

0:13:24.160 --> 0:13:27.880
<v Speaker 1>exemption for Australia and discussed defense arrangements amid pressure from

0:13:27.920 --> 0:13:31.200
<v Speaker 1>the US for allies to raise their defense spending. She

0:13:31.240 --> 0:13:34.480
<v Speaker 1>says US Secretary of State Marco Rubio did not raise

0:13:34.520 --> 0:13:38.000
<v Speaker 1>Australia's defense budget, and the discussions were on regional stability

0:13:38.040 --> 0:13:42.679
<v Speaker 1>more broadly, and the Victorian government says it will consider

0:13:42.720 --> 0:13:45.679
<v Speaker 1>all of the recommendations of the final report from Australia's

0:13:45.760 --> 0:13:50.040
<v Speaker 1>first formal truth telling inquiry. The four year inquiry, which

0:13:50.080 --> 0:13:53.840
<v Speaker 1>found crimes against humanity and genocide were committed against Aboriginal

0:13:53.840 --> 0:13:58.480
<v Speaker 1>people in Victoria, has delivered one hundred recommendations, including using

0:13:58.480 --> 0:14:01.480
<v Speaker 1>a treaty framework to provide dress for what occurred during

0:14:01.559 --> 0:14:05.120
<v Speaker 1>and as a result of colonization. For more on that story,

0:14:05.160 --> 0:14:07.760
<v Speaker 1>you can listen back to yesterday's episode of seven AM

0:14:07.800 --> 0:14:11.800
<v Speaker 1>on the York Justice Commission. I'm Ruby Jones. Thanks for listening.