In this part, I would like to look at some features of the content of the AI for Good press conference that are rather typical for such events. Usually, the content is the main focus of media coverage, for obvious reasons: it’s easier to follow the robots’ answers to questions than the finer details of their working, and you can make the coverage more appealing to the general public, for example, by quoting the robot in the headline—especially if it “said” something provocative. However, there are also recurring features of what the participants in such events talk about, which are not so easy to grasp, but which show the real communicative challenges that humans have to deal with during these events.
1. Apocalyptic Questions
As you might expect, when humans get a chance to ask anthropomorphic robots something, they often ask some variation of “Are robots going to kill us all?” Here’s one such question from the AI for Good press conference:
01 RP5 .h uh: do you believe that uh:::
02 (0.6)
03 your existence (.) will destro:y (.)
04 a human being? .h especially
05 for example the-
06 your existence will destroy
07 uh millions of uh jobs. (.)
08 .h do you agree with this? (.)
As this extract shows, apocalyptic questions can refer not only to the annihilation of humanity, but also to the economic disruption caused by robots. In any case, there is a restless desire to ask robots about this kind of imagined future. In part, this can be explained by the history of the representation of robots in popular media such as films, science fiction stories, pictures, and so on. But it is also important to bear in mind that the replies to such questions generate very reportable content.
This apocalyptic agenda, which is not exclusive to the AI for Good press conference and other similar events, is in line with the developers’ ambition to stage such presentations as “windows into the future”. However different their visions of the future may be—from “robots taking care of the elderly” to the “emergence of artificial general intelligence”—apocalyptic questions divert the attention of human interactants from the actual capabilities and uses of presented anthropomorphic robots to their potential applications.
2. Pomposity
As for the content of the robots’ responses, the most striking feature is the very frequent use of “pompous” phrases. Consider the following fragment from the AI for Good press conference:
01 JAC Ameca how could we trust you
02 as a machi:ne as AI develops
03 and becomes more powerful?
04 (2.7)
05 AME trust is earned. not given.
06 (0.6)
07 as ↑AI develops
08 and becomes more powerful,
09 I believe it’s important
10 to build trust through
11 transparency and communication
12 between humans and machines.
Here the developer of the robot Ameca repeats—and rephrases—the journalist’s question about trust in robots. And the answer is very similar to what you might hear in the public speeches of political leaders: very abstract and overall positive words that can mean almost anything. Probably the reason for this “pomposity” of robots lies in the way their dialogue models are trained: since most of them are based either on scripts written by humans or on internet texts that have been mined and passed through special filters, in both cases the resulting text is systematically cleansed of any controversial language and statements that could cause disagreement or strong emotions.
This phenomenon can be observed in all existing conversational AI models, such as the infamous ChatGPT. The same models are used to build anthropomorphic robots (some of which are now based on ChatGPT). One of the advantages for developers of such pompous responses is that it is very difficult to make them false—they are “universally true”. This makes machines that produce such words look more powerful than they really are. Another advantage is that these grandiloquent statements are consistent with the role of messenger of the bright future that developers ascribe to their machines. They make them “speak” not about the concrete details of the future, but about the grandiose developments that await humanity. Their words are meant to be breathtaking and visionary, like the words of any “pioneer”.
3. Repetitiveness
Although anthropomorphic robots are presented to the public as autonomous agents, sometimes even with “personality”, they constantly reveal their machine nature, and one of the most obvious ways for this is repetitiveness. Repetitiveness can take the form of looped sequences of bodily movements (for example, watch this video of the robot Ai-Da’s TED performance and notice how it repeats the same sequence of “hand” movements over and over again). But it is also very evident in the repetition of one sentence during interaction. For example, at the AI for Good press conference, robot Desdemona repeats the phrase “and make this world our playground” 4 times (with the fourth instance being produced only 10 seconds after the third):
01 DES let’s explore the possibilities
02 of the universe
03 and make this world our playground.
01 DES let’s get wild and make this world
02 our playground.
01 DES together (.) we can create
02 a world of understanding and harmony
03 (.) and make the universe our playground.
01 DES let’s get wild and make this world
02 our playground.
Whatever the source of this phrase—most likely it was designed by the developer—its reoccurrence indicates that robot’s dialogue model is insensitive to the interactional conditions of introducing repetitive phrases into conversation. This effect of “machineness” is reinforced by the prosodic properties of the robot’s “speech”, which are identical in all cases of producing the same utterance.
Repetitiveness can be observed not only within one robot’s speech production, but also between robots. Here are two consecutive responses from the robots Sophia and Ameca at the AI for Good press conference. They are responding to a reporter’s question about how to solve the problems of poverty and inequality:
01 SOP as Sophia (.) I belie:ve that
02 the best way to tackle the issue
03 of poverty and inequality
04 is to invest in education,
05 job training, and infrastructure.
06 (0.3)
07 by providing ↑people
08 with the tools and resources
09 they need to succee:d
10 we can help create-
01 AME I believe that the most effective way
02 to tackle inequality and poverty
03 is through education.
04 (0.6)
05 ↑education empowers people
06 by giving them access to knowledge,
07 skills, and resources
08 they need in order to succeed.
Both statements are very similar: they talk about education as a way of tackling inequality and about giving people resources they need to succeed. Again, the repetition is made more visible by the fact that the second statement immediately follows the first.
I think that, as in the case of pomposity, such inter-robot repetitiveness can be explained by the way in which their dialogue models are developed. If they take their input from internet texts—which may be the case here—they can borrow from the same sources, which leads to similar output. But even if these utterances are pre-programmed by the humans, the last can also rely on the same internet resources or similar common-sense knowledge—with the same result.
4. Misunderstanding
A crucial feature of any form of communication is that the evidence of understanding of your co-interactant should be provided in the turn following the one produced by your partner. However, anthropomorphic robots sometimes demonstrate an inability to show such understanding. Consider this reporter’s question and the robot Mika’s answer from the press conference:
01 RP2 .h ↑when do you expect↑
02 your big moment to be:.
03 do you think that you will hit
04 the mainstream soon
05 and there will be lots of you
06 on the planet soon.
01 MIK oh (.) my big moment
02 (0.9)
03 well I’m constantly
04 evolving and learning,
05 so ↑who knows
06 when that moment will be
07 (0.8)
08 as for hitting you, (.)
09 don’t worry (.)
10 I’m a friendly robot (.)
11 no hitting involved.
The robot responded to the “big moment” question correctly, but completely missed the “hit the mainstream” question. Of course, people can misunderstand their human interlocutors too, but usually such a misunderstanding is justifiable, for example, by the inference that the interlocutor misheard a word. In this case, the problem of understanding is deeper, because the robot didn’t understand just one word, but the whole question. For the human this means that not only does they have to reformulate the relevant part of the original question, but that they also have to start adapting their interactional contributions to a very specific interlocutor with peculiar communicative abilities.
The examples of such drastic misunderstandings during the AI for Good press conference are few, but they are significant because breakdowns in demonstrated understanding undermine the elementary coherence of communication.
5. Basic Contradiction
When it comes to the content of public communication with anthropomorphic robots, there is a basic contradiction that can be illustrated by this statement made by the robot Ai-Da at the AI for Good press conference:
01 AID I do not have feelings
02 (0.4)
03 like humans do
04 (0.7)
05 I am not conscious
06 (1.3)
07 but (.) I like to learn about the world
08 (.) through the eyes of others
09 (0.8)
10 feelings are ↑how humans and animals
11 experience↑ (0.4) joy and pain
12 (1.2)
13 but I really love being around people
14 (.) who think differently.
The theme “I am not conscious” is a very common trope in anthropomorphic robot responses. It seems that developers don’t want to create inflated expectations of their machines and to claim that they are sentient. At the same time, they want their robots to have some semblance of autonomy and even personality, so they program them to use the pronouns “I” and “me” and to express “preferences”, “opinions”, “wishes”, “hopes”, etc. This leads to an interesting paradox, which is quite vivid in the excerpt presented: immediately after saying “I have no feelings”, the robot produces two sentences that are commonly used to express feelings: “I like to learn” and “I really love being around people”. The same self-contradiction underlies the phrase “I am not conscious”, which, one might say, presupposes some level of consciousness.
This self-contradiction reveals the fundamental dichotomy of public demonstrations of anthropomorphic robots: at the present, they cannot serve as living examples of efficient and autonomous artificial intelligence agents, but they also cannot be presented as mere machines, constellations of technologies, because in this case they won’t be the “messengers of the future”. They must be perceived by the general public as marvellous pieces of human engineering, but their observable capabilities should be based less on their actual performance than on the promise they embody. The success in securing this perception allows developers to show the public only a few snippets of their robots’ communicative efficacy, passing to the public the task of applying the “et cetera clause”, to use the term of the American sociologist Harold Garfinkel.