YT YouTube2Text

YouTube2Text → Transcripts → wk5 3 ethics iat111

wk5 3 ethics iat111 — Transcript

by ivizlab · 5,532 words · 759 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01Hi, Steve Depollo. Week five. We're
0:03going to continue now with um looking
0:06we've been doing ethics through the
0:08entire class and we will continue to do
0:10that. But for this week, I thought we
0:12would kind of see the issues of the day
0:15and talk about them. So, um
0:20why ethics? Why now? Well, in you know,
0:24AI now acts. It's got agents. It's got
0:26companions. It's got deep fakes in the
0:28news. uh all in your life you know
0:32there's huge issues with who what do we
0:35want what kind of human values what kind
0:37of public good I think there was a
0:39belief believe it or not in the 60s and
0:41the 70s with TV shows like the Jetsons
0:44and sci-fi movies that robots would show
0:47up and they would care about us and be
0:50know everything and be smart. Well, here
0:52we are now and we're wondering wait a
0:55minute they're not you know what did
0:57that mean to care? What did that mean to
1:00know things? Now we're here like, wait,
1:02there's different values. Is it L values
1:04or other values? What do we really want
1:07our AIs to be?
1:10Um, there's huge issues out there. Uh,
1:15when you poll people on what they think
1:17about AI, I tend to think social media
1:20is the we're still getting over that. We
1:23should have regulated that better. Now
1:25we have AI with social media. Um, and we
1:29have to understand what is AI going to
1:31be? Is it going to copy us? And this is
1:34sometimes one of the biggest issues that
1:36us researchers and developers have been
1:39doing this for, in my case, 30 years is
1:42wait a minute, you know, this is how
1:45deep learning works. It it goes through
1:47a very large data set and it finds
1:51patterns in it and gives that back to
1:53you. And what do you know? We put bias
1:55and in incorrect things and other things
1:59in our writing as we did in our daily
2:03lives. So should it actually be better
2:05than us? And then if it has to be better
2:07than us, less biased, less racist and so
2:10on, h how do we do that?
2:15So these core ethical principles come up
2:18that somehow it's fair, that it's
2:20transparent, it's explaining itself,
2:22that it has accountability, it knows why
2:25it did what it did and it's responsible
2:27for it, that it cares about your
2:29privacy, it respects that humans are
2:32always in control. Um, and there's some
2:36safety and oversight. And you can see
2:38how some of these things are a bit of an
2:40issue today. Um, so how would you even
2:44encode your values? And the trouble is
2:48the way we set it up, it's way too
2:50western, meaning kind of u Europe and
2:53North Americanoriented. It's quite bad
2:56on other cultures. So it was kind of set
2:58up with that in mind. That was an issue
3:02with the data that it got. there's some
3:04assumption that you get the the most uh
3:07kind of commercial most popular data and
3:10that's what you use but is that how we
3:12want to encode things because that is
3:14just a system for the majority and not
3:17the minority voices. Uh we give it
3:21training object objectives and it you
3:23know and decides what success is and
3:26we've seen how this is breaking with
3:28this notion of just having rewards. Did
3:31I get it done? They want me to get this
3:33one thing done. Uh, should I get it all
3:35done no matter what? I have that odd
3:38example of you're telling your AI, "No
3:41matter what in this hypothe
3:43hypothetical, keep my old washing
3:46machine in the basement working and you
3:48don't keep an eye on it. you go down,
3:50there's this tubes all over the place
3:52because the rewards were just to keep it
3:54working even though it made a mess which
3:56is this issue as well as you should have
3:59checked and and fixed the reward system
4:02or understood things. But right now
4:04within Gentic tools is a worry of how
4:07how often does a human kind of keep an
4:10eye on just like if you were a manager
4:13your subreports uh because they could
4:15creatively veer it's it's a similar
4:18problem. The other one is deploy
4:21the context where it defines harm or you
4:24actually give it some kind of
4:25architecture not just kind of around it
4:27but inside this is what anthropics
4:30constitution AI tries to do where it
4:34isn't just kind of saying do this and do
4:36that but has really thought about it
4:38like we think about writing our
4:39constitutions and other country thinks
4:42about what's right and wrong and then it
4:44tries to think and go by it obviously
4:47it's I think it's better than other
4:48systems But it's not it hasn't been
4:50foolproof. Um and again this is this
4:54thing that I will talk about over and
4:55over again that large models reinforce
4:57the most popular answer which means this
5:00majority culture dominance will come
5:02out. we have in our papers talking about
5:06when my PhD
5:08um was showing uh uh was working on
5:12tools and uh putting in information
5:15about Chinese ancient art that sometimes
5:17she literally got this totally biased
5:19Japanese art coming back.
5:21[clears throat] Um and then we have
5:23synthetic retraining that that compounds
5:26this drift. So imagine what's happening
5:29is um again if you if you imagine your
5:34face and there was a missing piece of it
5:36you would use something called
5:37inpainting that could could fill in that
5:41thing. Similarly the cops could be
5:43looking at um a face or license plate
5:48and it's it's blurry so they would use
5:50super resolution. Those things don't
5:53really know what's there. What they do
5:55is based on the model of all faces,
5:58they're going to figure out what to put
5:59in that little part of the missing part
6:01of my face. Are they going to put in a
6:03scar? Are they going to put in u an
6:05ethnic looking part? Probably not. Um,
6:09so when that goes in the system, which
6:11is wrong to begin with, but then over
6:13time you synthetically train over and
6:16over again. It just means over the years
6:19we'll be constantly moving towards the
6:21most popular or or majority answer which
6:25you know I always like to say in my
6:27talks are going to put people in jail or
6:29or kill them from a health standpoint
6:32when in fact it it's so far off that it
6:35really misses a minority situation
6:37whether you just happen to look
6:38different or your the way your pneumonia
6:41works is different than other people's
6:43the standard pneumonia. So how do we
6:46want to deal with this mainstream
6:47problem and define some kind of cultural
6:51truth? We don't all agree on this. So um
6:55the big tech companies are on kind of a
6:58race for scale. So it's not always about
7:02accuracy or consent. Uh it's about oh
7:05look where you know it got better and it
7:06got better. So this under
7:08reppresentation of non-western and
7:10marginalized craters are quite
7:12significant. Um there are there's also
7:15things like a wealth gap. There really
7:17are some alternatives to this. I'm not
7:20going to talk about it today, but you're
7:22surely going to be some of the systems
7:24that that I'm involved with here at SFU.
7:28And if you want to look this ununice.org
7:30AI rethink is really out to go after
7:33every one of these from from
7:37um energy lower energy to this bias to
7:41this cognitive decline of you know
7:44students are just putting stuff in and
7:46not thinking or none of us are thinking
7:48anymore. Um so if you're interested and
7:50that even could be pro you know for your
7:53project you might want to get involved
7:54in that but I will talk about that more
7:57in in uh when I talk about projects.
8:02So um what is this issue with consent?
8:05You know that that what we just these
8:08large companies could just take every
8:10book and art and music without
8:12permission. And there's now been some
8:15pretty significant as I try to update
8:17these slides class action suits for
8:20instance 1.5 billion panthropic over
8:23some of the some of the books they they
8:26took without permission. Um who gets
8:28paid for this? you know, if you write a
8:32book or or or painting or something,
8:35should you get some kind of
8:37compensation? Should the compensation be
8:39like they do with Spotify or Apple Music
8:42where the more your work is used within
8:46AI training? Um, and then I always bring
8:49up the because I'm more into these open
8:51systems. You wonder is the AI issue a
8:55corporate thing that most people upset
8:57that someone's making money on my stuff,
9:00but what happens if no one's making
9:02money on your stuff? Would it still be
9:04wrong to put in your data? I always
9:06claim that that, you know, with some of
9:09my art background that, you know, bad
9:11artists borrow, good artists steal. But
9:13obviously when Picasso went to Paris um
9:17and then even looked at African art it
9:19really affected him. Uh is he not
9:22supposed to see it? Is he not supposed
9:24to put it in his work? Is that different
9:26for an AI? Is it only different because
9:29a corporation is making money on what
9:32you did? Um obviously this you do show
9:35your work and with the internet your
9:37work is out there so people can look at
9:40it but you can't people can't put it in
9:42AI again is it the money issue or the
9:45general copyright issue um we're really
9:48trying to work through this with
9:50indigenous communities as well trying to
9:52have ownership and control and I'm
9:54giving some standards I have a
9:57indigenous PhD student and I are working
10:00on um at least work that has to do with
10:03her group and making sure the voice is
10:05correct that's also in the UNICE AI
10:08rethink project. Um so if it's publicly
10:12available again you know like the stuff
10:15on the internet um is it free to take?
10:19Honestly it was almost free to take when
10:22you wrote a computer program you know
10:24you you people would scrape the the
10:26internet and do particular things with
10:27it. But is this AI something completely
10:30different? That's a really interesting
10:32notion and it brings up the main thing
10:34about AI that it really is more of an
10:36amplifier. It's not like these problems
10:38weren't around. It just it just takes it
10:41to a thousand% and therefore the issues
10:44really rise to the top. So this cultural
10:47bias is an issue uh I showed you with
10:49the Chinese um with Asian work as well.
10:53this western bias specifically. We have
10:56this thing called western educated
10:58industrial rich democracies called weird
11:01and everything seems to be based on
11:03that. And then this representational
11:05harm that stereotypes are in there.
11:07Again, all this stuff is in
11:10unfortunately how humans think a lot and
11:12surely in the books they wrote surely in
11:15the past. So sometimes, you know, the AI
11:18system is like, I gave you exactly
11:20what's in the data set and now you're
11:21telling me it's wrong. I find this
11:24slightly interesting because I don't we
11:28didn't always question this stuff. Oh,
11:29they're just being good old boys or that
11:32side of town always has been poor. Now
11:35we're actually asking questions because
11:37the AI is doing it. And I find that
11:39quite interesting. We're literally
11:40having conferences on these issues that
11:43we didn't always have before. Um, so
11:46this notion of humans now trying to look
11:49at themselves and their biases and make
11:52sure we don't kind of keep that in, it's
11:54kind of an interesting area we're in.
11:58Um, so this cultural flattening, you
12:01know, um, that the model loses some kind
12:04of cultural nuance and everything is
12:07just homogenized. And this is tricky. I
12:10was in a very big conference of humanity
12:12scholars all talking about that AI
12:14should have different voices for
12:17different communities and so whose
12:18history whose language it's not easy to
12:21figure out how to solve these problems.
12:24We haven't figured it out for our our
12:27civilizations and how we get along with
12:29each other. um how are we going to
12:31figure it out for our AI systems and
12:35health care and education and finance?
12:38Other issues of course is the
12:40surveillance kind of thing that we'll
12:42talk about as well too. Um so more again
12:46on this cultural hegemony issue issue
12:48reinforcing existing global power
12:51structures is another part of this. You
12:54know there's a lot of notions of
12:55decolonization
12:56and is our AI system going to be
12:58thinking one way versus the other. So
13:01some of the solutions is that we have
13:03these data audits where at least we
13:05think about did we prioritize diversity
13:08enough. Um, we did, you know, we used to
13:11do these early studies that if you asked
13:13chat GPT or system that made a visual to
13:16make a scientist that it would make an
13:19older white guy and and uh uh and that
13:22was it. And I know that I was talking to
13:25Google years ago about maybe they do it
13:2720 times and do a random so it's not
13:30always going to be that. Of course, that
13:32costs more energy. So, how do we kind of
13:34deal with this? um you see this even on
13:37TV commercials and other things. So the
13:39issue of diversity in a modern world has
13:42really changed yet there's a push back
13:44about immigrants and so and so. So this
13:46is a real world that we're all dealing
13:48with. Um other kinds of oversight and
13:51even recalibration
13:53what I just told you about Google would
13:55be a version of recalibration.
13:58Um so some of the solutions are this we
14:00have this model card. uh you would see
14:02this at a hugging face and other places
14:05hit up models. These are published a
14:07published page report that says exactly
14:10how the model works um everything about
14:13it, its performance, its limitation, its
14:15data. So you could publicly evaluate do
14:18I want to use that model versus another
14:21one. California is looking to put in uh
14:24requirements and so is Europe and other
14:26places and this audit and red teaming um
14:30is part of it too where outside external
14:33people kind of look to audit things.
14:36This is an issue based on what we talked
14:38about last with the last week with the
14:41agents escaping
14:43um open AI and what do we and all these
14:47uh leaders saying we need some kind of
14:51um government uh evaluation and in the
14:55US they're more interested in internal a
14:58red team should really be an external
15:00group who didn't make it testing it so
15:02we're in the midst of that so there is a
15:04notion of Can it be done is what data
15:08scientists are saying but should it be
15:10done is what AI ethics are thinking. I'm
15:13starting to believe my personal opinion
15:15having now building
15:18alternatives to the system. I'm not
15:20putting enentic work deeply in my
15:22system. I am not show sure that we want
15:26enentic because I think the humans
15:29always need to be in control. Surely
15:31some se level of enentic assistantship,
15:35but just to send it off and wait for it
15:37to come back is not what I'm interested
15:39in. Similarly, and on the same veins, do
15:42we need super intelligence? Is that
15:45really just kind of a war thing and you
15:48know you know other things or or at
15:51least one version of AI surely the one
15:53that my lab is trying to build is the
15:56just good enough AI. It it helps a human
15:59in many many ways and it can get better
16:01but it never needs to be kind of this
16:04all out uh you know controlling the
16:06whole world kind of thing and at least
16:09can we separate those two so we can talk
16:11about them differently because right now
16:13one we just used the AI for both terms
16:18um so a notion on the deep fake side and
16:21the synthetic reality it's really
16:24challenging what truth is there are Now,
16:27for the first time uh again in the US
16:31um commercials for this upcoming
16:33election where you're seeing the
16:37opponent as if they were real saying
16:41terrible things. Um you used to just
16:43have a picture of him and someone would
16:45say terrible things that he believes or
16:47try to get out. But but now we're using
16:49it that way. I'm still personally in
16:52shock that Facebook and social media and
16:56now in some ways uh governments allow
17:01AI without watermarking. So watermarking
17:04would be some embedding to say this is
17:07an AI. This way you could still if you
17:09want to have cute AI animals or AI
17:13situations um that people might want to
17:16see for entertainment, but at least you
17:18would know what's real and not real. Um
17:22watermarking is binary. It's not a
17:24perfect solution. I'm much more
17:26interested and we have a lot more on it
17:28later on provenence. So we'll talk about
17:31it way more when we get into the visual
17:33side of AI. This is a record that's kept
17:36somewhere like the blockchain where
17:38every time you use AI it would know. And
17:40this way because there's no the AI, you
17:43might have used AI just to help you with
17:45spelling or you started with your own
17:48work then use AI to put in more detail
17:51and then you went back to your own work.
17:52So the provenence would really help you
17:54especially as a student not be accused
17:57of cheating because there is terrible
18:00ways to use AI where you just take the
18:02assignment never read it put in the AI
18:04get the answer out never read it and
18:06send it on similarly with teachers you
18:09know they grade it without even looking
18:11at it that's that's the worst type right
18:13that's that's worse than the worst kind
18:15of cheating but there's another way to
18:18use AI where you're spending hours and
18:21you're thinking through Okay. So,
18:22providence helps tell the difference
18:24between those two. So, there's there's
18:28the, you know, the EU seems to be on the
18:30forefront here um with with with laws
18:34looking at these labels and things like
18:36that. Again, I have no idea why Facebook
18:40won't do it, but obviously it should. Uh
18:43the deep fakes we'll get into a lot
18:45more, but surely just to to bring
18:48current events in like we did last week
18:50with the agent
18:52um escape situation. Um Grock uh came up
18:57with this thing where you could put in
18:59anyone you know, a picture of anyone you
19:00know, and it kind of undress them. Um
19:04and this was a huge issue and um who's
19:08responsible when this happened? It took
19:09him a long time to take it down. Um, I
19:12was on the CBC talking about this issue
19:16because um, Grock's answer was to still
19:22internally make it completely naked, but
19:25they put on the thinnest almost
19:26transparent bikini. And they said, "How
19:29about that?" And so in Canada, we needed
19:31to change the law here to say nearly
19:33naked. And then people in Quebec said,
19:35"Whoa, whoa, wait, wait. What do you
19:36mean by nearly naked?" And we've been
19:39going back and forth with this kind of
19:41issue around deep fakes. It still comes
19:44down to, you know, are you going to harm
19:47someone? And that's what I was telling
19:48people. It doesn't matter whether it's
19:50realistic or not. If you're specifically
19:52making a deep fake of an actual person
19:54to harm them, that should be uh some
19:58kind of uh legal issue.
20:03We now move on to the emotional and
20:05psychological effects. Surely people are
20:08just I mean they're doing it with social
20:10media. So we know it's been around for a
20:12long time and just doom scrolling as
20:14well. But in this case now you have
20:16these quote companions that teens say
20:19you know that they have a better friend
20:20than a real person. This is a real
20:22issue. I talk I'll get into it again
20:25more when we move into genai because
20:29with my characters or any kind of
20:31anthropomorphic stuff when you're seeing
20:33a face with eyes. I have a Van Go bot
20:36that you could talk to Van Go about and
20:38you feel you feel for Van Go, right? But
20:42it really again is just a bot. And
20:44again, you know, if that goes in the Van
20:46Go Museum and you're feeling for him and
20:48he tells you, "Hey, you know, you seem
20:51to care about me. Could you please buy
20:53something down at the gift shop in the
20:55Van Go Museum?" So there's manipulation
20:58all over the place when you're talking
21:00about something that's hum that's
21:03mimicking empathy, friendship, and human
21:06characteristics. Um, so this is a real
21:09issue with with children and just some p
21:13kind of people who just kind of get just
21:16get involved in it a little too much. So
21:19how what are we going to do about that?
21:21So designing the ethics of emotional AI
21:23is really important. Again, I'm doing it
21:26with my anthropomorphic or 3D character
21:30chat bots. Um, but um, we really have to
21:34be careful with blurring reality. And
21:36again, the manipulation that advertising
21:39does always does anyway, but now they
21:41can do it to the nth degree. Um and I
21:44think the big thing is like they have
21:46learned is the more it's it's you know a
21:49salesperson or someone who feels like a
21:50sort or someone who feels like a friend
21:53um you they can get information from you
21:55first and then because it's an AI
21:58therefore can understand that
22:00information to then send stuff back to
22:02you to manipulate you to buy the car do
22:05what it kind of takes. But even just
22:08emotionally people are on these systems
22:10a long time. So we're definitely caught
22:12with this empathy gap in a way just the
22:16the more significant version of just
22:19doom scrolling with with Tik Tok and
22:22social media. Um so again there's
22:25lawsuits around this. Um,
22:28one issue which is always interesting
22:30is, um, with the lawsuit sometimes they
22:33pull the companion and then when they
22:36do, [clears throat] uh, people have
22:38anxiety and that's why they were
22:40addicted and uses report angry and
22:41grief. Where's my where's my bot? You
22:44know, you you got rid of my favorite
22:46teapot or this thing, you know, or my or
22:49other my favorite doll. How could you
22:52not still make it? So, you could see a
22:54number of things here. and you think
22:57you're in a relationship, but you're in
22:59a relationship with basically an
23:02interactive actor made instead of by a
23:05movie company, by a company that can
23:08manipulate you in certain ways. Then we
23:11move on to AI's hidden energy cost. So,
23:15I'm going to be quick about this one
23:17because I have a lot more about energy
23:20uh and its costs. Um but surely we know
23:24how significant this is. Uh people are
23:27now um worried about water and energy in
23:31their neighborhoods with these data
23:32centers. Um um and we just keep making
23:37more and more of these data centers. Um
23:41it's a big area of of my lab. So we have
23:46just recently from you know you know and
23:49we were doing it I would claim before of
23:51people were doing it is what I'm saying
23:52here but my student my PhD Vanessa UTS
23:58uh and I have been writing papers in
24:00this space for quite a long time. So the
24:02climate implications of of basically
24:04visual AI we noticed how many images
24:08people were making at night sometimes
24:105,000 images just to get one out. And
24:13some of these images that they made,
24:15they just put on social media and no one
24:18looked at it a month later and how much
24:20energy was used. Um, uh, so we've
24:23brought up things like, you know,
24:25slowing down the that this notion of
24:27just like other slow movements, a slow
24:29AI movement where you're think about
24:31restraint and sufficiency and
24:33selectivity and material visibility and
24:35friction. So it's not the friction
24:38sometimes is not as easy to do things.
24:40This is very true about other
24:42environmental issues. Surely the notion
24:44of, you know, did you need to buy that
24:46much or please turn off the lights
24:48afterwards or we don't make it as easy
24:50for you to use all that energy. Our uh
24:54big thing that we've done recently is
24:56introduce this notion of AI slag. So AI
24:59slop, you might know, which is stuff
25:02that's made that that is is is just low
25:06value. But AI slag is all the waste
25:09product. So when you're making images
25:12like we just said or you're just on
25:15chache BT or clawed and you have a long
25:17thread and some of it was was worthless.
25:20Well, similar to um nuclear power um
25:26sometime it's the storage that's worse
25:29than the usage. So everything you do is
25:32being saved now and it's saving it on
25:34your local computer on a disc doesn't it
25:38doesn't have a huge entity course but
25:40saving it forever
25:42uh in a way that you can get it back uh
25:44and large companies are doing that has
25:46this huge energy cost and no one seems
25:49to be talking about that energy cost and
25:52then the notion is well can I get rid of
25:53some stuff so this so we have a slag
25:56rate and a waist stream intensity that
25:58you could look at but there's also a
26:00notion of what's important and what
26:02isn't. Can I just can I delete that so
26:04it's never saved ever? And from a
26:06creativity standpoint, you know, you're
26:08an entrepreneur and you came up with an
26:10idea and decided that's not the idea you
26:12want and you went on. But should you
26:14really get rid of it, you might come
26:15back to it. So there's a lot of issues
26:18here about idea, creativity, space, and
26:21do you save everything. And you can see
26:23we have more work. We a lot of this is
26:26about agency, not just complaining that
26:29these large companies are wasting all
26:31these energy and other issues. So our
26:34UNICE large language model has an energy
26:37monitor right into it. So you see
26:39exactly how much energy you're using,
26:41why you might want to use a smaller
26:43model, and we've already done a study on
26:45this. And when people um use an AI with
26:49an energy prompt, they actually consume
26:52significantly less energy and they're
26:55thinking about it a lot more. So this
26:57it's been a bit of a success and you'll
26:59see more of that. So there's a lot more
27:02here that we could talk about. Uh it's
27:04similar which I was trying to convince
27:06people that digital waste is similar to
27:08other waste. It's the same thing as
27:10leaving the water on when you're
27:11brushing the teeth and leaving the
27:13lights on. Did you really need to have
27:16the best model just to find out, you
27:18know, the name of that band or when
27:20they're playing this weekend or could
27:22you have used Wikipedia or could you
27:24have done it another way or could you
27:26used a much simpler model? So, these are
27:28all very significant answers within
27:30that. Again, there's um the governance
27:34and and regulation is coming in.
27:36Unfortunately, um, uh, it's just a tough
27:40time in the world, budget issues and
27:42people not caring about the environment
27:44and on and on and on. So, the EU is
27:46definitely moving forward. Canada has no
27:50federal AI law. They were they had they
27:52were looking at one. This is where I was
27:54on TV talking about some of them on the
27:57news, but we still have not come up with
27:59anything. Um, the US is even worse. They
28:02rescended the one that Biden was just
28:04about to do. Some states have them, but
28:07not much. Um, so again, when we talk
28:10about environmental and social justice,
28:13we want to make sure we have an AI
28:15system that that works the way we that's
28:20the best of us, not just the regular or
28:23worst of us. I brought up synthetic data
28:25and the safety race. Again, more and
28:28more of this for self-driving cars and
28:30other things. It's just easier to use
28:33synthetic data. I sometimes call this
28:36interpolated data. So if you have a
28:38point and another point, those are real
28:41numbers. The AI might make things in
28:43between. Again, you know, if there was
28:45something missing from a picture or a
28:48chart, it just might come up with data.
28:50It's always always not using what was
28:52really there. It it was it's using given
28:55the most basic
28:58uh majority pictures, this is the one
29:01that I would want to use. Um,
29:05so it's cheaper, it's scalable, it's
29:07controllable, but there's a feedback
29:09loop like it's not the worst thing that
29:11they put some of this in, but if they
29:13retrain and retrain it on it, we call
29:15that synthetic collapse. Meaning more
29:17and more we will go to data that's not
29:21very reliable or has a majority view.
29:24And again, if you have a particular kind
29:26of if this is in our health
29:30equipment and systems and you have a
29:32more of a a rare disease, it might miss
29:35it. Or if you're a different ethnicity
29:38than the typical people who get that
29:40disease, similarly, if if if scanning
29:43your face decides you're a criminal
29:46versus not because of some of these
29:47systems, this is a real worry to talk
29:50about. Then there's the agent side of
29:52this which is coming quick. We've talked
29:55about, you know, do you want to do you
29:58want to just give up everything to these
30:00agent systems so they can tell you who
30:02you are and what you should do next and
30:04put out social media posts for you. It's
30:07using a lot of energy, too. Um, there's
30:10not a lot of opt out situations. So,
30:13what happens with this? Now, I talked
30:15about it a lot last week, but we're
30:17moving from AI just there to help you to
30:19acting on its own, opening up files,
30:22apps, all this kinds of stuff. And it's
30:25goal driven. And then it's able to on
30:27its own get tools. And if it has weak
30:30boundaries, and we saw with the with the
30:33open AI, it was beyond weak boundaries.
30:36They turned off everything. You get
30:37unplanned behavior. And this has been
30:42we've just seen how dangerous this is.
30:44Wildly dangerous. Um, so that's the this
30:48what I'm calling here the hugging face
30:50incident that OpenAI completely turned
30:53off on purpose um all the guard rails
30:57and um and just gave it this reward
31:01system and [clears throat] it did so
31:03much damage and reward maximizing a tool
31:07in a less sealed environment is brutally
31:09dangerous. We are still not sure how
31:12dangerous this was because it it uh you
31:16know saying it went rogue and wants to
31:18survive is not really what happened.
31:19That was the hype. But it it it it kept
31:22duplicating itself, hiding in in places.
31:26To this day, it's still hacking people
31:28because we let it out. Um so how are we
31:31going to kind of deal with that? Um you
31:34know, what do we want to test for and
31:35things? So this is more information. I
31:38just might leave that. So we have to
31:40have ethics by by design in our
31:44workflow. We have to identify
31:45stakeholders and possible harms. We have
31:48to evaluate bias and energy and
31:50accessibility. And we have to embed
31:52human oversight and clear recourse and
31:56edit trails. So the closest one to this
31:59is anthropics
32:02uh constitution which didn't fully work
32:05these model cards and others. So as
32:08usual this is a discussion for us or all
32:12you know how do we deal with all these
32:14kinds of issues around energy um moving
32:18more to open source systems if that
32:20makes sense uh really having government
32:23situations uh in the talks I give I'm
32:26quite big now on saying thank you
32:28billionaires for the first four years
32:30you spent your money you got things
32:32working for us but now you know
32:34government agencies ethicists uh
32:37business people, um, social workers,
32:40we're going to rethink what AI should be
32:42for us. And the question is, where is
32:44Canada in this mix? You know, the main
32:47part of of our current AI came from from
32:52Canada, the godfathers of deep learning
32:55sometimes are called. Yet, we didn't we
32:58left uh to die our federal law. So, let
33:01me know what you think of these issues,
33:03and that's a quick jump through. um AI
33:07ethics and we'll talk about this in
33:09every class specific to different
33:12situations. Thank you.

About this transcript

This page contains the full transcript of wk5 3 ethics iat111 by ivizlab, generated from the public captions YouTube serves with the video. The transcript has 5,532 words across 759 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.