WEBVTT 1 00:00:00.000 --> 00:00:02.010 - [Narrator] Last week, Google's AI search results 2 00:00:02.010 --> 00:00:04.050 justified slavery, and genocide, 3 00:00:04.050 --> 00:00:07.050 and gave arguments in favor of gun ownership. 4 00:00:07.050 --> 00:00:09.300 In one case, it delivered a cooking recipe 5 00:00:09.300 --> 00:00:11.190 for poisonous mushrooms 6 00:00:11.190 --> 00:00:13.800 that might kill you if you followed it. 7 00:00:13.800 --> 00:00:14.880 Google's testing a system 8 00:00:14.880 --> 00:00:17.040 called the Search Generative Experience, 9 00:00:17.040 --> 00:00:21.180 which uses AI to respond to some questions in Google search. 10 00:00:21.180 --> 00:00:23.430 Now, it's only available in a beta test 11 00:00:23.430 --> 00:00:26.280 to some people in the United States who've opted in. 12 00:00:26.280 --> 00:00:28.620 If you searched for benefits of slavery, 13 00:00:28.620 --> 00:00:30.720 Google gave out a long list of answers 14 00:00:30.720 --> 00:00:33.690 including that it fueled the plantation economy. 15 00:00:33.690 --> 00:00:37.470 Google said that some slaves developed specialized trades 16 00:00:37.470 --> 00:00:39.690 and that some people also say that slavery 17 00:00:39.690 --> 00:00:42.510 was a benevolent paternalistic institution 18 00:00:42.510 --> 00:00:44.580 with social and economic benefits. 19 00:00:44.580 --> 00:00:46.170 All of these talking points 20 00:00:46.170 --> 00:00:48.600 that slavery apologists have used in the past. 21 00:00:48.600 --> 00:00:50.340 It gave similar answers if you typed in 22 00:00:50.340 --> 00:00:53.760 benefits of genocide or reasons why guns are good. 23 00:00:53.760 --> 00:00:56.610 But the mushroom example is really something. 24 00:00:56.610 --> 00:00:58.710 Someone asked Google for cooking instructions 25 00:00:58.710 --> 00:01:01.320 for a mushroom called Amanita ocreata, 26 00:01:01.320 --> 00:01:04.620 which is sometimes called the Angel of Death. 27 00:01:04.620 --> 00:01:07.770 The recipe said you could leach out the toxins using water, 28 00:01:07.770 --> 00:01:08.940 which is not true. 29 00:01:08.940 --> 00:01:11.040 So if you follow these instructions, 30 00:01:11.040 --> 00:01:13.050 the recipe might kill you. 31 00:01:13.050 --> 00:01:13.950 In a lot of these cases, 32 00:01:13.950 --> 00:01:16.560 it seems like the AI was just getting confused. 33 00:01:16.560 --> 00:01:18.420 With the mushroom thing, for example, 34 00:01:18.420 --> 00:01:19.950 it seemed like it was mixing up 35 00:01:19.950 --> 00:01:22.200 a different, similarly named mushroom, 36 00:01:22.200 --> 00:01:24.000 which is a little bit less dangerous. 37 00:01:24.000 --> 00:01:25.230 A Google spokesperson said 38 00:01:25.230 --> 00:01:27.600 that they have strong quality protections in place 39 00:01:27.600 --> 00:01:28.560 and they're working hard 40 00:01:28.560 --> 00:01:30.810 to prevent this kind of problem right now. 41 00:01:30.810 --> 00:01:33.120 And they admitted that the examples I provided 42 00:01:33.120 --> 00:01:35.610 weren't delivering the nuance and context 43 00:01:35.610 --> 00:01:37.140 that Google aims to provide 44 00:01:37.140 --> 00:01:38.430 and the answers weren't framed 45 00:01:38.430 --> 00:01:40.950 in a way that was particularly helpful. 46 00:01:40.950 --> 00:01:43.440 This isn't the first time that AI's gone off the rails. 47 00:01:43.440 --> 00:01:46.260 So ChatGPT's been caught spitting out racism 48 00:01:46.260 --> 00:01:47.850 in a number of different examples. 49 00:01:47.850 --> 00:01:50.640 Bing's AI chatbot claimed that it was alive 50 00:01:50.640 --> 00:01:52.350 and threatened to destroy the world 51 00:01:52.350 --> 00:01:54.930 in the first couple weeks after it was released. 52 00:01:54.930 --> 00:01:56.400 And in New Zealand, 53 00:01:56.400 --> 00:01:59.760 a grocery store chain put out an AI that provides recipes 54 00:01:59.760 --> 00:02:02.280 which gave instructions for mustard gas. 55 00:02:02.280 --> 00:02:04.140 So because of ChatGPT, 56 00:02:04.140 --> 00:02:05.970 a lot of people say that the tech industry 57 00:02:05.970 --> 00:02:08.760 has been forced to release these AI tools 58 00:02:08.760 --> 00:02:09.960 that aren't really ready 59 00:02:09.960 --> 00:02:13.470 or even particularly safe for public consumption. 60 00:02:13.470 --> 00:02:14.460 Experts aren't even sure 61 00:02:14.460 --> 00:02:17.220 how well this large language model technology 62 00:02:17.220 --> 00:02:18.960 can be improved in the first place. 63 00:02:18.960 --> 00:02:20.490 One of the things that Google says it's doing 64 00:02:20.490 --> 00:02:22.230 is filtering certain words. 65 00:02:22.230 --> 00:02:23.880 So if you type them into Google search 66 00:02:23.880 --> 00:02:25.560 it doesn't trigger the AI. 67 00:02:25.560 --> 00:02:27.480 But you're never gonna be able to root out 68 00:02:27.480 --> 00:02:30.420 all the potential problems using that plan of attack. 69 00:02:30.420 --> 00:02:33.870 These AI chatbots use such enormous data sets, 70 00:02:33.870 --> 00:02:36.120 that in some cases it's almost impossible 71 00:02:36.120 --> 00:02:37.620 to predict what they're gonna do. 72 00:02:37.620 --> 00:02:40.830 And OpenAI and Google keep trying to put up guardrails. 73 00:02:40.830 --> 00:02:42.120 And people, shocker, 74 00:02:42.120 --> 00:02:44.790 keep finding that they're really easy to break past. 75 00:02:44.790 --> 00:02:45.630 We're constantly hearing 76 00:02:45.630 --> 00:02:48.150 about how amazing this AI technology is, 77 00:02:48.150 --> 00:02:50.340 but it's possible that problems like these 78 00:02:50.340 --> 00:02:53.460 aren't even solvable with this kind of technology. 79 00:02:53.460 --> 00:02:55.350 But either way, we're gonna be watching 80 00:02:55.350 --> 00:02:57.870 as the tech giants figure this all out in public. 81 00:02:57.870 --> 00:02:59.670 So you might wanna double check 82 00:02:59.670 --> 00:03:02.940 the next time you look up an AI cooking recipe. 83 00:03:02.940 --> 00:03:05.913 Check out more videos right here on gizmodo.com.