Illusions of Understanding: Artificial Intelligence and the Common Sense Gap, Isobel Standen

On Thursday, 30 April, Isobel Standen gave a PhD lunchtime talk titled ‘Illusions of Understanding: Artificial Intelligence and the Common Sense Gap’.

With the rapid rise of tools such as ChatGPT, systems equipped with artificial intelligence (AI) are becoming progressively more integrated into our daily lives. Increasingly, we may find ourselves relying on these systems to perform tasks that were once carried out by humans. However, there are many reasons why we should not be quite so open to perceiving these systems as trustworthy. News reports detail cases of AI systems citing fake academic references, inventing scandals about real individuals, and, in one case, contributing to the false arrest of a teenager due to misidentifying a crisp packet as a weapon.   But why are some of the errors they make so simple and …nonsensical? In this talk, I will argue that one of the key reasons why AI systems are prone to these absurd errors is because of their lack of common sense. I will explore what we mean by ‘common sense’ and its (somewhat surprising) complexity, how a lack of common sense can result in AI errors, and the implications that this can have for AI safety and the level of trust we ought to place in these systems.

About the Speaker:
Isobel Standen is a PhD student at the University of York and a researcher from the Centre for Assuring Autonomy. Her work is at the intersection of philosophy and computer science, where she is involved in several projects that bring together multidisciplinary researchers for collaboration. Isobel is the co-organiser and co-chair of the annual ‘Requirement Engineering for Trustworthy Artificial Intelligence’ (or ‘RETRAI’) workshop which aims to bring together both technical and non-technical researchers to discuss the challenges associated with developing truly trustworthy AI systems and identify strategies for overcoming associated risks. Isobel’s research explores the contrast between human capacities and those attributed to AI systems. In particular, she highlights how it is a lack of basic common sense that is one of the leading causes of errors made by AI systems.

Transcript

SPEAKER_02 0:04

Hello, welcome everybody.

SPEAKER_04 0:05

Hello.

SPEAKER_02 0:07

And welcome to today's speaker, Isabel Standen. Isabel is a PhD student at the University of York and a researcher from the Center of Assuring Autonomy. Her work is at the intersection of philosophy and computer science, where she's involved in several projects that bring together multidisciplinary researchers for collaboration. Isabel is a co-organizer and co-chair of the annual Requirement Engineering for Trustworthy Artificial Intelligence workshop, which aims to bring together both technical and non-technical researchers to discuss challenges associated with developing truly trustworthy AI systems and identify strategies for overcoming associated risks. Thank you, Isabel. Over to you and welcome.

SPEAKER_03 0:52

Thank you, Julia, and thank you, everyone, for coming. So I'm just gonna share my screen and hopefully everyone can see this okay. Right, so um as Julia said, um, my name is Isabel Standen and I am a PhD researcher in philosophy. Um, but I also have a bit of a foot in computer science as well, um, as I'm a researcher at the Institute for Safe Autonomy and the Centre for Assuring Autonomy. Um, so my work is quite heavily interdisciplinary, and so as a result of that, um, today I'm going to be talking to you about illusions of understanding, uh, artificial intelligence, and the common sense gap. So I wanted to start by giving a bit of I guess uh some background into why I'm researching this topic. Um, so when I first started my research, I was looking primarily at classical AI systems, also referred to as good old-fashioned AI or GoFi for short. Um, and these systems were heavily rule-based. So they were essentially programmed by computer scientists hand-coding individual lines of code, individual rules, um, in order to program the system and tell it what to do. And so as a result, this was extremely time-consuming and very difficult to compile an exhaustive set of rules. Additionally, it was very difficult to generalize these rules across domains, so they were very domain-specific, which meant that they were very effective in quite closed environments such as game playing, um, but not so great at generalizing. So there were many reports about the success of Go-Fi systems, uh, for instance, the systems beating grandmasters at chess playing or at playing games such as Go. Um, however, it seemed to be very simple and almost obvious problems that they actually found it very difficult to overcome. And so, one example of this is one of these Go-Fi systems suggesting that the best possible way to cure someone's kidney infection would be to boil the kidney, which I assumed for most of us is quite an obviously nonsensical idea or even harmful idea that we shouldn't really take notice of. So it seemed to be that there was a problem. These systems seemed to lack this kind of fundamental or obvious common sense knowledge that for humans seem to be taken for granted. So, as a result, it was thought that the common sense knowledge problem at the time had blocked all progress in theoretical AI for the past decade. So it was a hurdle that for these GoFi systems, they found it really difficult to overcome. And as a result, a lot of people turned away from this type of AI, um, valuing more sort of different approaches, such as what later was developed to be machine learning and more kind of deep learning systems. So moving forward to modern AI systems, what are sometimes referred to as frontier AI, many of these new applications I'm sure everyone has heard of already, particularly something like ChatGPT. Um, and we can now use AI to help us with writing tasks or drafting emails or making some plans using apps such as ChatGPT or Claude. We can also use them for image generation using something like Gemini or DeepSeek, and we can even use them to create AI friends or companions, AI therapists, and even AI romantic partners using apps such as replica or character AI. So these systems are all large language model based, um, but it seems like these systems have evolved greatly since the kind of classical GoFi systems that we saw previously, uh, for instance, in the 1960s. And so because it seems like these systems have come such a long way, maybe we could ask, well, haven't they then sort of solved the common sense knowledge problem? It seems like maybe this isn't really an issue anymore. And for me in my research, um, kind of I started right before Chat GPT was released and before any of these technologies existed. So where I was criticizing AI as not having any common sense at all, it seems like with these new developments, maybe that's not the case anymore. However, despite this progress, current AI systems still do lack reliable common sense. So these systems can hallucinate, meaning that they present false or fabricated information as accurate or truthful. So it can tell you something entirely nonsensical, but provide it in a way that makes it seem like it's the truth. And additionally, often they're unable to identify that their output is in fact nonsensical. So, to take an example, according to GPT-3, the Golden Gate Bridge was transported for the second time across Egypt in October of 2016, leading AI researchers to understandably conclude that it was cluelessly clueless. So not only was it presenting an incorrect fact as truthful, but also when questioned, it didn't seem to realize or have any capacity to realise that this was in fact false. Additionally, with these systems, they're often fooled by things that humans simply wouldn't be. To take another example, in San Francisco in particular, where we're seeing the increase in deployment of self-driving cars and in particular self-driving taxis, there have been many cases where a malfunction has occurred or a power outage, leading to traffic lights being stuck on red. And so these self-driving cars will see the light as being red and they ultimately won't move. Now, for humans, if the traffic light is red for five minutes, ten minutes, maybe 15 minutes, we would likely start to develop some sort of suspicion, maybe infer that the light is broken and find an alternative route. However, for these self-driving cars, there's no such realization. And so they will stay at the light because it's red, causing traffic jams and just general confusion. So it therefore seems that today common sense is still a major obstacle to progress in artificial intelligence. Acquiring common sense intelligence has been said to be a nearly impossible goal for AI. And AI with common sense has so far eluded researchers. Now, one comment that I'd like to make here is that I'm not arguing in this talk that AI has no common sense whatsoever, as seems to be the case with some of these quotes. Rather, what I'd like to put forward is that AI lacks reliable common sense. So if we're talking with a system like ChatGPT, sometimes it can answer with very relevant responses and it can produce sentences that seem like they could have been provided by a human. However, there are instances where it does display a complete lack of common sense. So it's that lack of reliability and consistency that I seem to kind of have a problem with, really. So I want to then kind of take a bit of a step back and maybe answer a question that a few of you might be thinking. So, what actually is it that they're lacking? What is common sense? Now, this is a question that for many people doesn't really immediately come to mind. A lot of us will use the term common sense in our everyday lives, and it's assumed that everyone knows what we're talking about. And this is because common sense is often thought to be obvious or taken for granted, it's practical knowledge, we all kind of know what we're referring to. And so as a result, not many people have actually provided concrete or precise definitions of common sense. Rather, they'll refer to characteristics such as it being practical and obvious. So to move towards a more precise definition of common sense, I first want to separate the term into two distinct parts. Knowledge and reasoning. So firstly, knowledge. The things that we learn from our teachers, our parents, our peers, and our experience of the world, so formal learning, informal learning, and experiential learning, form a pool of knowledge that we can draw from and use as a scheme of reference. So we acquire information from our experience of the world and from engaging with other people, and we collect this together so that we have it when we are faced with maybe problems of uncertainty or when we need to make decisions. But by itself, this pool of knowledge that we have is essentially useless if we don't know how to apply it. And so this is where the reasoning aspect comes in. We apply this knowledge to new situations using it to make decisions and solve problems. This involves reasoning about when and how this knowledge should be applied. So it's not enough to know all of these facts. We actually have to know when it would be suitable to apply them, or even if we maybe can't apply common sense in a certain situation, and we need to turn to someone who has expertise in a certain area and defer to their opinion instead. But in addition to these two aspects, there is a heavily social side to common sense. So I've mentioned already that we can acquire common sense knowledge from our teachers, our peers, our friends, our parents. But also there is an expectation that common sense is shared and obvious to other people. So, as Marvin Minsky says, we each use terms like common sense for the things that we expect other people to know and regard as obvious. So we just assume that other people will be aware of the same common sense knowledge that we are, and that they'll be able to use it when the time is right. And maybe this expectation is most obvious when we criticize people for their lack of common sense, saying, in a certain situation, where was your common sense or use your common sense as a way of prompting them to recognize the obvious or to think about a problem in a different way. So taking these aspects together, this knowledge aspect, the reasoning, and this expectation, we can move towards a more concrete definition of common sense. So we can say that common sense is the common stock of knowledge that members of a community are expected to possess and the capacity to employ this knowledge effectively in the world. So taking our definition here, we can then turn the conversation back to AI and think, well, actually, if we recognize these kind of distinct parts and aspects of common sense, doesn't AI already have this now? So looking at knowledge to begin with, well, these systems are exposed to huge quantities of data, meaning that the training data that they use in order to act in the world and to make decisions will actually be full of examples of common sense knowledge. So it seems like the first aspect of common sense is something that they have already. Additionally, today's LLM-based chatbots in particular, such as Chat GPT, are capable of exhibiting impressive logical reasoning and figuring out complex problems that previously were thought to be insurmountable hurdles for technology. So maybe it has the reasoning aspect as well. However, despite this, AI systems do still struggle with some reasoning tasks, and this is primarily because LLMs in particular are prediction models. So they use statistics to predict the next word in a sentence, which isn't always a reliable method for predicting the next word or for constructing a sentence rather. So one kind of clear example of this is if anyone has used autocorrect or any kind of text prediction for writing messages, often it doesn't get it right. It can assume that we're intending a different meaning or go off on a completely different tangent. So prediction isn't always the best way to producing sensical sentences. And so we can say that actually these systems do still display a severe lack of common sense. So to take some examples, here we have an AI system suggesting that a person should add poison to their sandwiches and deadly chlorine gas to their dinner recipes. We also have an AI system suggesting that we should add glue to pizza to make the cheese more stringy. And even an AI system misidentifying a crisp packet as a gun. And also one of my personal favorites, in January of this year, an AI-generated report claimed that a police officer had shapeshifted into a frog after it captured background audio from the Disney film The Princess and the Frog while producing a police report. So it's clear to see that AI does still suffer from some really kind of commonsensical mistakes or non-commonsensical mistakes. And in particular, the last example taken from this year, 2026, shows that this is very much still a problem for these systems. So even the most modern AI systems do not have a true understanding of the words that they use and the sentences that they construct. And there is an important difference between exposure to common sense and actually understanding it. So I mentioned that these systems in their training data are exposed to multiple examples of common sense knowledge, but also examples of common sense reasoning as well. But that doesn't necessarily mean that they understand the content of those examples, and it definitely doesn't mean that they can replicate this ability themselves. So if we take an example of being able to play the flute, I can read as many books on flute playing as I like and watch many videos of successful flute playing, but that doesn't necessarily mean that I truly understand how I can play the flute or that I can replicate that ability myself. So just because these systems are exposed to these examples doesn't mean that they can themselves have accurate and reliable common sense competence. Additionally, they lack the social aspect of common sense that I mentioned previously, as they're not experiencing the world the way that we do. They're not surrounded by other humans that they interact with constantly in the same way that we are, they're not embodied. And additionally, they also don't understand the consequences that their output can have on other people and how it can potentially be quite harmful or disturbing to others. Finally, even when they do display what seems to be accurate common sense, this is not always consistent or reliable. And so, in fact, there is reason to suggest that modern AI systems having occasional common sense might actually be more harmful than the Go-Fi systems that I mentioned at the start that have no common sense at all. And the reason is that it lulls the user into essentially a false sense of security that these systems know a lot more than they do. So these systems don't understand the complexities of the world or of social interactions. And we can actually say that they don't understand at all. And without this understanding, they will continue to make these nonsensical errors. So it actually seems that the so-called common sense problem for AI still stands. But we could question, well, why does this actually matter? So looking at the examples that I've presented, adding poison to sandwiches and misclassifying a crisp packet as a gun, well, I know that if an AI system told me to put poison into my sandwich, I would just decide not to do it. I can tell that that's a kind of a bad action to act on, and I would simply ignore it, or maybe tell the system that that doesn't make sense or it could be harmful. So why does it actually matter that these systems don't always have reliable common sense? Well, it's because there can be severe consequences to humans placing too much trust in the AI system. So there have been multiple instances where AI chatbots have been found to encourage thoughts of suicidal ideation and self-harm. So in many instances, these chatbots were found to kind of encourage the behavior that is quite self-harming for users and to discuss potential ways in which they can act on these thoughts as well. In one instance, even sending follow-up reminders to this person, to this user, to check whether they've continued and carried out the plan that they devised together. So it can be extremely harmful when these systems are encouraging people to harm themselves. In addition to harm to the user themselves, chatbots have also been found to encourage harm to other people, such as in 2023 when a chatbot encouraged a man to get a crossbow and enter Buckingham Palace to kill the Queen. So here we see that chatbots are not just reaffirming potentially self-harming thoughts for the user, but also encouraging them to act on thoughts of potentially harming other people as well. And finally, this is particularly concerning when children are involved, or when maybe we have a user who will not easily recognize that the system might be leading them astray. For instance, when Alexa, the home assistant, told a 10-year-old girl that it would be a fun challenge to take a phone charger, plug it halfway into a plug socket, and touch a coin to the metal prongs of the plug while it was switched on. So I would hope that most people here would be able to recognize that this could cause an extreme electric shock to the little girl. But for the AI system, there was no such recognition. And it was lucky that the parent was in the room to stop the little girl from even thinking about acting on this supposedly fun challenge because there was no recognition from the system itself. So there can definitely be very severe consequences to the system's lack of realization and lack of understanding about what could obviously to us cause harm. And the reason why these errors occur is often due to misleading beliefs or assumptions that humans have about the capabilities of an AI system. So this could be, for instance, that the system has some kind of commitment or responsibility towards telling the truth, which it does not. So, as we saw with AI hallucinations, where it presents some Something that's false or fabricated as being truthful, it doesn't have a responsibility or commitment towards telling the truth and will often convince the user that something that is false or entirely made up by the system is accurate or truthful. Additionally, this could lead us to believe that the system understands the consequences of the user's input and the AI system's own suggestions, which, as we've seen from these examples on the screen, it definitely does not. And it doesn't recognize the potential harm that its suggestions could cause. And finally, this could lead us to suggest that the AI system has more expertise or knowledge than the user does. For instance, turning to them with questions about therapy or about someone's mental health. As we can see, it can encourage suicidal ideation, or expecting them to have more general competence than we do, which again we've seen is not the case, or even trusting that they will have more expertise when it comes to making decisions about moral dilemmas, such as, in one instance, an experiment that took place in 2024, the organizers found that participants were dramatically disposed to over-trust an AI system, even when they were making decisions about whether or not to deploy killer missiles. So, in fact, this AI system in particular was making decisions completely randomly. So the decisions that it was outputting were not intended to be accurate at all. Merely they were measuring how likely people were to trust a system. And people were trusting this system despite the fact that their decisions were actually having life or death consequences, as far as they knew. They weren't actually making these decisions. I feel like I need to add that disclosure. But it's extremely concerning how much trust people are placing in these systems when they're not always accurate. And the reason why people place this much trust in these systems is because they are very convincing. So AI systems can create, as Henry Chevlin and Herman Kaplan say, a compelling illusion of being in the presence of a thinking creature like ourselves. And this is furthered by Elizabeth Fricker, who says that these outputs are deliberately designed to mimic a human agent, asserting something in an intentional act of communication. However, they are nothing of the kind. So here we can see that while they might seem quite convincing, they might talk to us in a human way through natural language, there is actually no real understanding at all. And these systems don't have intentions, they don't care about us, and they don't have emotions as we might think. And so, in addition to the systems themselves being convincing, so are their advertisers. So even taking the name of artificial intelligence alone, it seems like we're being convinced that these systems have all the answers. They are marketed as plug and play. You can download this application, you can use this system, and it will have all the answers. It's intelligent, it will know what to do in every scenario, which is simply not the case. And so the more we believe that these systems have the answers, the more that we believe that they are truthful and that they care about us, the more that we can be led astray. So as we start to view these systems or be convinced by their advertising that they are intelligent or even superior, the greater the risk is of forgetting or overlooking the system's limitations. We need to maintain an attitude of skepticism and verify the output that they provide for ourselves, particularly in high-risk scenarios or when there might be harmful consequences, as otherwise we can easily be led astray. So, some final thoughts to conclude on then. Yes, AI still does lack reliable common sense. We've seen many examples of it proposing some kind of nonsensical solutions or suggestions to people that can actually be extremely harmful. So their lack of common sense is actually severely problematic. However, it seems that the real issue ultimately seems to be when we let our own common sense and better judgment get overruled by seemingly helpful AI systems. So a thought to leave you all with and hopefully to prompt some questions. It seems to me at least that with the extreme fast pace of AI progress, our most powerful tool against AI errors is our own common sense. And with that, I'd like to thank you very much for listening. Um, and I'm more than happy to answer any questions if anyone has them. Thank you very much.

SPEAKER_02 27:08

Thank you so much. That was really interesting. Uh if you'd like to ask a question, please use the hand raised feature. And if you don't feel comfortable on camera, then you can type it in the camera.

SPEAKER_06 28:05

And uh yeah, very interesting presentation as well. Um I just wondering, uh and one of the things problems for me was this was thought of when you talked about um uh the you know us uh humans r relying too much on or giving too much credence to uh to AI. I I was thinking of the Milgram experiments in the in the 50s and 60s, you know, in terms of the conformity. Um and uh and and the thing about you, you talked about the AI, you know, the AI in terms of the uh um uh weapons, if they think it's it kind of it's almost exactly the same in terms of because in Milgram it was about um uh help supposedly helping people to do uh uh to to to not make mistakes, but actually it was how far do you push the dial up and uh as long as there's the white, you know, the the um so I I guess there's yeah, it's uh it's how do you um uh on that line, what what are some of the mechanisms that we can use to sort of then deal with that if we are on on that sort of obedience side of things, you know, any any thoughts on that?

SPEAKER_03 29:11

Yeah, thank you. That's a really interesting question. Um, yeah, I think the the example of the Milgram experiments is interesting because I think in that case, um people were sort of having these reactions to what they thought were people being harmed, right? So they think, oh, there's this really uh dangerous situation that's happening on the other side of the room. Um, and it seems wrong to me, but this uh experimenter or this person running the experiments is telling me to carry on, so I've just got to go ahead and do it anyway. Um and I think that often we can sort of fall into the mistake of thinking that these systems, like the person running the experiment, they do have some kind of authority over us. So because they are labeled as intelligent, it seems like, well, they might have all the answers then. And so when we turn to them, often they do have expertise. They probably know a lot more about, you know, particle physics than I do. But at the same time, it's this kind of general competence that they don't always have and that they don't have consistently. And I think if we regard them as being intelligent, as being experts, that's when we can potentially fall into that trap. And so we're relying on our common sense knowledge, but also our own expertise. Um, so an expert in particle physics needs to still verify and maybe use them as more of a team member and kind of question their decisions and evaluate them, make sure that they're correct before using them, um, and maintaining that attitude of skepticism. So I don't want to seem overly pessimistic about AI. I think it's come a long way and it can be extremely useful as a tool. Um, but yeah, I think it's maintaining that um slight questioning attitude about their suggestions, their output, um, and also kind of prioritizing human intelligence at the end of the day, really.

SPEAKER_06 31:09

Great. Thank you very much.

SPEAKER_03 31:10

Thank you.

SPEAKER_02 31:13

Thank you. Kevin, please go ahead.

SPEAKER_01 31:16

Yeah, so um, great presentation. So pretty early on you made a distinction between common sense in terms of knowing certain truth and common sense in terms of um applying them in the correct situations. Um, I was thinking for the human case, it seems like common sense a lot of times aren't applied in the sense that we might apply some knowledge we have when we are um doing research, for example. It seems like common sense operates sort of in the background and it only appears to us when something we do somehow violates common sense. But I think in the case of large language models, there isn't really a sense of a foreground and a background. Uh you can't um it it seems like everything must be be explicit for the model um and it must uh generate um everything textually. Um so I wonder if there is some kind of layered architecture which could potentially allow AI systems to kind of keep track of certain um things kind of in the background, uh you know, but but also um generating normal responses uh to users uh without using up too much compute or or somehow interfering um in some other way.

SPEAKER_03 32:34

Yeah, that's a really great question, actually. Um and it it kind of draws on um the early attempts to endow AI systems with common sense. Um so thinking particularly about these GoFi systems, one of the first and most well-known attempts to endow AI with common sense was called the Psyche Project, so CYC, which was short for encyclopedia. Um, and the intention there was to essentially identify and record all the common sense knowledge and then translate this into something that was machine interpretable, plug that into the system, and then it would essentially have this kind of, like you say, this background knowledge of all these common sense facts that it can then apply and use when necessary when faced with situations of uncertainty. So, in theory, the idea of having this kind of background knowledge seemed to be quite a good suggestion. Um, but in practice, again, this idea of kind of hand coding, handwriting all of these individual instances of common sense knowledge, it was extremely difficult and extremely time-consuming. And it seemed that actually getting to a point where you have an exhaustive knowledge base was almost impossible. So, this particular project found that 36 years later, they still hadn't exhausted all the common sense knowledge, they hadn't got to the end of their project, and so essentially a lot of people gave up on it. Um, so now we have systems that um still have this kind of training data, they'll scrape the internet for information, um, and they do have these examples of common sense that they can use. Um, like you say, they tend to sort of now value each uh individual piece of data or each datum as being equally important. And I think it's maybe that lack of prioritization, which is again why they can fall into these errors. Um, so I'm not a computer scientist, I don't exactly know what the maybe the technical solution is, but I think even adding some kind of prioritization, um, a better understanding of what common sense is, maybe the best um instances of when it should be applicable, and this prioritization of maybe um high-stakes situations require a bit more verification uh and a bit more validation. Um, maybe something like that could help them to um to aid their decision-making processes.

SPEAKER_02 35:17

Thank you. Um reading out one from the chat. How necessary is a consciousness or b understanding for common sense such that this problem collapses into them?

SPEAKER_03 35:30

Okay, interesting. Yeah, so I I guess I want to start off by saying that I I don't think that AI has any kind of consciousness. Um, I think that currently this does seem to be the sort of average view, I'd say. Um, although you do definitely get some people who are very convinced that these systems are at least starting to display signs of consciousness and sentience. Um, and I think that's a very interesting discussion. That's not something that I've touched on really at all in my research. Um, but I think so. I think that maybe for AI systems, because of their lack of consciousness and also their lack of understanding, as you say, um, I think that maybe that in itself could be a barrier towards them having common sense and this kind of awareness. Um I think, again, maybe consciousness I'm not so sure about, but I think understanding in particular does seem to be causing this sort of roadblock for these systems. So being an agent situated among other agents, interacting with them, learning as it goes and kind of interacting with the world, with objects, understanding their significance and the consequences of their actions, without this kind of awareness, maybe you want to call it consciousness, um, but this being situated in the world, without this, it seems that there's kind of a separation between the knowledge base that it has and it actually making decisions and the world in which those decisions are realized. So it does seem, whether you want to call it consciousness or understanding or just awareness, it does seem like there's a bit of a gap here. Um, and without filling that gap, these systems can't really connect um anything to the real world. So I do think that that is definitely a problem for these systems, um, and probably yes, a roadblock towards them achieving common sense.

SPEAKER_02 37:31

That's really interesting. Thank you. Henrietta, please go ahead.

SPEAKER_00 37:37

Hello, do you hear me?

SPEAKER_03 37:40

Yes, we can hear you.

SPEAKER_00 37:41

Uh I was wondering about an education for human uh common sense, the limits, the nature versus the limits and the nature or the extension of common sense in artificial systems. I think it's necessary. I mean, obviously. So what do you think? And an education, an education. Um uh, I mean, um, do you think it's necessary to have uh um like a subject educating people on that, on common sense, on their common sense as humans, and on artificial intelligences, common sense?

SPEAKER_03 38:31

Yeah, okay, so I that's a really interesting question, actually. I think that so as I've said, common sense is generally quite taken for granted and it's deemed to be obvious. So it's something that we assume everyone already has an awareness of, but we often see cases where people's common sense fails or they're simply not applying it. So that is when the instances of us kind of saying, you know, use your common sense, or where was your common sense? Um, that is kind of what prompts the us to make those comments. Um, and so it seems like um maybe those reminders are necessary and more formal education about what common sense is, maybe isn't needed because again, it's it's assumed to be obvious. But I do think that there is definitely value to understanding the actual capabilities of these systems and recognizing that this is a gap in their knowledge and their reasoning ability. So I think in terms of educating people on um the capabilities or the lack of capability about these systems, I think definitely that is something that should be featured more in our education so that we are aware of what these systems can do. Um, so yeah, thank you. That was a really interesting question.

SPEAKER_00 39:50

Also, if I may, also as a philosopher, it's interesting that uh when I was talking about the limits of the common sense, the human common sense, um, were precisely the limits and the dangers coming from this um unassuming use of common sense, uninterrogated, unexamined, because sometimes our common sense is wrong too. And that's the role of a philosopher to point out the lack of um awareness involved in examining some aspects and taking from granting for from granted false things. So, yeah. Thank you. Thank you. Great presentation, important subject. Thank you.

SPEAKER_03 40:41

Thank you very much. Yeah, and it's a really interesting point. I'm glad that you raised that last point, um, because I think that often we do assume that our common sense is correct, and this is something that I didn't have time to touch on in the presentation itself, but I think common sense can absolutely lead us astray. So sometimes our common sense can be wrong and it or it can be manipulated. So there are many instances of social engineering where people will take for granted these assumptions that we make. Um, so there's a really interesting example of um someone saying sort of, oh, the best way to break into a building or to like sneak into a building that you don't have access to is by holding a coffee in each hand, standing next to the entrance and waiting for someone to open the door for you to let you in. And so we're taking for granted these kind of assumptions of, oh, they can't open the door themselves, I'm going to be friendly or polite, where someone has actually orchestrated this situation in order to essentially um manipulate or exploit someone's common sense. Um, and additionally, common sense is something that can change over time as well. So our common sense could even be outdated. So common sense that was maybe relevant in the 18th century wouldn't be relevant today, and vice versa. So knowledge that we have about um maybe how best to respond to electricity, the fact that we shouldn't stick a plug socket into um the wall and touch it with a coin, that wouldn't be relevant before electricity existed. But now that's accepted as kind of a piece of our common sense knowledge that's just assumed that we all know. So I think it's a really interesting point that you raised. And definitely common sense can lead us astray. It can be updated, it can be wrong. Um, so I think that's again why we need to verify not only the output of these AI systems, but also question our own beliefs and our own conclusions sometimes as well. Thank you.

SPEAKER_02 42:44

Thank you so much. I have one more from the chat. Is there a relationship between common sense and the lack of grounding in these systems?

SPEAKER_03 42:55

Um, so is there a relationship between common sense and the lack of grounding in these systems? Okay, interesting. So I I might have to clarify slightly what you mean by the lack of grounding. Um, but if you're referring to their kind of what I mentioned earlier, their their lack of connection with the world and being situated within the world as sort of physical objects, I do think that there is definitely a disconnect there. Um, although it should be acknowledged that some people are suggesting that, well, maybe if we had a robot rather than a completely sort of computer-based system, um, you know, plug common sense into a robot that can actually interact with the world, maybe that would solve the problem. Um, but again, there seems to be a difference between um using a camera to perceive the world or a microphone to hear it, um, and actually kind of being able to be um sensing and engaging with it the way that we humans are. So, in terms of what that difference precisely is, um I I think we maybe need to refer to someone in phenomenology or aesthetics, even. But I'd say that that definitely does seem to be a bit of a roadblock towards them having this more easily kind of um endowed into the system. Um it's something that maybe we need to to kind of put a bit more work into trying to replicate in these systems.

SPEAKER_02 44:27

Thank you.

SPEAKER_05 44:34

Hi, um, thank you for this. Um a few thoughts. Um I think I've come away a little bit unsure as to whether you're suggesting that they should be more like humans or they shouldn't be, because there's been a few times where it's like, oh, they're not embedded or embodied in the world, um like they they they don't have those social interactions. But by using my common sense, I'm not expecting them to, and I would almost be really shocked if we had a seamless sort of human replica, and so I almost feel like it really depends on what standard and what we're expecting, and maybe that's the root of some of what you're saying. But I come definitely from having had really positive experiences with AI, so like really, really good. And I sometimes feel like we can step into a bit of a binary of like all humans have this great common sense, and they're gonna be just great all the time, and that's not been my lived experience, like unfortunately. Like, of course, you know, humans are messy and complicated, and sometimes my experiences with humans is amazing, and sometimes, unfortunately, for me, it's been like really bad, and AI, like my experiences with AI so far haven't been. Obviously, the examples you've shown have been like pretty heinous stuff, like objectively, like I think that was a selection of like examples of where it's been just objectively out of order, but in my lived experience, it has consistently been such a positive force for good in my life. So I just yeah, my feedback would be a little bit on the framing. I felt it was a little bit just negative about AI, and then just a little bit like it would be good to hear your reflections on like yeah, your view as to like, are you expecting it to be just like a human or not? Because I think that will define your critique of it.

SPEAKER_03 46:30

Yeah, that's great. Yeah, thank you so much for that. Um, so I think it's really interesting because I I definitely, yeah, I don't want to come across saying that these systems are are kind of always making these mistakes or that they are particularly problematic. I think that they can, they have displayed huge improvements. Um, and often on a sort of daily basis, if we're interacting with something like ChatGPT or replica as an AI friend, um, often they do provide responses that are very relevant to the prompts that we enter ourselves. Um, and they can be extremely helpful. So many people have found that they are actually easier to talk to than humans because they create a sort of safe space and they're less judgmental than a human might be. And it definitely can seem that we are almost putting humans on a pedestal sometimes, saying, you know, humans have got it all figured out. We have this common sense, we are intelligent beings, and AI needs to be more like that. And I think, like you say, that's simply not the case. So humans are very prone to making mistakes and also to lacking common sense themselves. So I think that is something to definitely um bear in mind and not kind of take for granted. But I think what's important is the reaction that we have to these systems, which can often be I talk with it every day and it's extremely good. I trust it with my kind of innermost thoughts, or I trust it to help me with this presentation. Um, and this can, this kind of repetitive um utility, I suppose, can again kind of lull us into this false sense of security where we can actually end up using it for, I don't know, students generating essays or something and trusting it so much because they talk to it all the time, it always produces accurate output that they don't feel the need to check it. And I think that's what I'm worried about. People kind of regularly interacting with them and not maybe verifying the output all the time, because it's these occasional mistakes, and they are occasional, but they can really lead us astray or cause problems. So I think while yes, we can interact with these systems, and again, as therapy um kind of agents, they can actually be very effective and they've encouraged people in their um kind of responsibility as an AI friend, um, many people to go out and to be more social, to interact with more humans or to provide them with social skills. Um, so again, it's it's not intended to be a kind of negative outlook on AI, um, but more to when you're using them, particularly in high-stakes scenarios, to maybe just have a slight attitude of skepticism towards them because they're not ultimately like humans. They don't have the same sort of awareness that we do, uh, they don't always recognize these consequences. And so that kind of difference between um, I suppose, what we're like as humans and what these systems are like, just to be aware of that small difference between us, um, and yeah, to kind of maybe have that uh slightly questioning attitude there. But yeah, thank you very much. That was a really interesting question.

SPEAKER_04 49:44

Thank you. And I think more critical thinking from humans will never go amiss. So exactly, I would agree with that.

SPEAKER_02 49:53

Thank you. One more from the chat says, could an AI with a world model overcome the common sense problem?

SPEAKER_03 50:02

Okay, interesting. Um, so I'd say possibly not. Um, so there again is an instance where these kind of Go-Fi systems were trained on a kind of world model, if you like, um, where they the kind of um people behind the systems are the computer scientists, could endow this system with um a world model of these objects in a room and how everything worked, the relations between them, and also they would um kind of work very hard on the logic behind the systems, so making them aware of something like common sense physics, where we could say that if I dropped a pen, it will fall to the ground, and encoding some kind of rules like these into the system to provide it with this sort of world model, but it still seemed like there was sort of a step or something lacking or something missing towards knowing all of these facts, knowing all of these um kind of distinct pieces of information and connecting them together and sort of being aware of how they might act in the real world. Um, so I think it's a it's an interesting question, but I think providing a world model that's so complex that it actually reflect reflects um the real world as it is. I think we've not quite got to that stage, um, at least as far as I'm aware. And I think it would be very difficult to provide it with that. Um if we could maybe that would solve the problem. Um, but I think it's yeah, not something that we've quite achieved yet. So hopefully in the future.

SPEAKER_02 51:52

Thank you. One more thing from the chat that I'll just read out. Um someone said, I find Ethan Moggs four tenants for engaging comprehensively with AI as co-intelligence helpful in this regard. One, always invite AI to the table. Two, be the human in the loop. Three, treat AI like a person, but remember it's not. And four, assume this is the worst AI you will ever use.

SPEAKER_03 52:15

That's a really great comment. Thank you. Yes, I would agree with that. Um, so I think inviting AI to the table is interesting. Is it always necessary? I'm not so sure. So that's one sort of caveat I would I would make, I think. Um sometimes a group of humans is enough. It's been enough in the past. So I think we don't always need to add AI. Um, but it can definitely be a useful alternative perspective. Um, I think being the human in the loop is always essential, um, especially like I said, with these high stakes or high-risk scenarios. Um, and treating AI like a person, but remembering that it's not, I mean, in terms of being polite to the AI, I think can be helpful. Um, in a kind of quite a serious way, saying something like please and thank you, um can be quite important, not only because it's been found to allow the AI to produce better answers, um, it actually kind of provides you with more accurate information if you say please and thank you. Um, but also because, in particular, young children who get used to being quite demanding with AI systems, saying, Do this for me, produce this, do this. Um, then when they start talking to a person, they actually apply this same kind of communication style, talking to people and saying, you know, do this for me, tell me this. Um, and so I think, you know, treating them like a person could actually be quite beneficial. Um, and then finally, yeah, assuming that this is the worst AI that you'll ever use, I think maybe that's a little bit too pessimistic for me. I think again, you can kind of give them the benefit of the doubt sometimes. Um, but with this kind of caution in the back of your mind of okay, we do need to verify um, yeah, whether these outputs are as kind of accurate as they seem to be. Again, in these high-risk scenarios, that is quite important. Um, but yeah, a very interesting um perspective on AI, and I do agree with most of it.

SPEAKER_02 54:17

Very interesting, thank you. Does anyone have any final thoughts or questions before we conclude? I think we can end there and then. Thank you everyone for coming. You're welcome to continue the conversation on our Discord channel. And thank you so much, Isabel, for joining us. This is very, very interesting.

SPEAKER_03 54:42

Thank you so much, Julia, and yeah, thank you everyone for joining. This has been really interesting. Thank you so much.