Reading the mind of a self-driving car

Humans have endless capabilities for understanding each other in highly nuanced ways – we are all aware of the subtle eye roll or glance away that can communicate so much.  The smallest of differences in our gestures carries a myriad of meanings that we are able to pick up and understand in a moment. In psychology, this ability we have to attribute mental states to other people is called Theory of Mind (ToM). One of the leading researchers in this area is Bertram Malle who asks us to put ourselves in this scene:

You observe two people’s movements, one behind a large wooden object, the other reaching behind him and then holding a thin object in front of the other.

We need theory of mind to understand that this simply describes a customer pulling out their credit card with the intention to pay the cashier.  If we did not have a theory of mind, we would not understand the sequence of movement or be able to predict either person’s likely responses that we simply take for granted. But with the capacity to understand physical movements through mental states, we can construe this complex scene into the intentional acts of offering and trading. 

Theory of mind means people are transformed from ‘objects’ into ‘subjects’ who can are able to act intentionally and who have desires, beliefs, attitudes and so on that direct their actions.  We need to understand minds in order to engage in the complex interactions that social communities require of us.  Malle suggests that we explain intentional behaviours, by using a sophisticated framework of interpretation.  He suggests that intentional behaviour must involve a desire for an outcome, beliefs about how a particular deed leads to that outcome, and an intention to perform that deed; if the person then performs the deed, people take it to be an intentional action.

Understanding robots

But things start getting a little more complicated when we bring technology into the equation.  We understand how to navigate a wide range of situations read each other’s minds but how do we manage this when one of the participants is a robot? 

Whether robots can be programmed to have a Theory of Mind about humans is addressed by AI researcher, Richard Sutton, in his essay titled “The Bitter Lesson”.  He argues the history of AI suggests that attempts to build human understanding into computers have not to date worked as progress has only come through the ability to assert ever more brute computational force to problems. The “bitter lesson” is that “the actual contents of [human] minds are tremendously, irredeemably complex…They are not what should be built in [to machines].”  There will inevitably continue to be speculation about the possibilities of a more generalised intelligence AI, for all intents and purposes we are not close currently and may well never be.

We should therefore turn to the corollary of this: what happens when we try to understand a robot as if it were another person? It is well understood that we use anthropomorphism, the human tendency to attribute human characteristics, either physical or psychological, to non-human agents.   Many studies have shown that humans perceive robots as having human qualities, such as Roomba (a vacuum cleaner with a semi-autonomous system).  Others have found that we tend to assign greater mental abilities to robots that have a human appearance.  It seems at first glance that ToM is active for humans not only in their relationships with other humans but also through our interactions with robots.

But surely, it’s something of a surprise thatwe might apply ToM principles to technology:  we do not do this for a range of other tools in quite the same way.  But there are a number of reasons in which we might be forgiven for attempting this.

We could argue that the world of ‘objects’ is perhaps not as entirely separate from us as we might think.  The boundary between object and subject is in fact much more porous than we sometimes consider – we use tools to extend our capabilities:  we can think of a shopping list we scribble out as a mean of extending our memory, just as a hammer is a mean by which we extent our physical capabilities.  However, seeing tools as an extension of ourselves is not quite the same as attributing human-like qualities to them as we seem to with digital technology. 

But given the way in which the tech industry has a tendency to imbue machines with human like qualities perhaps it is of little surprise that we do this ourselves.  There is certainly a huge interest for tech companies to reduce the gulf between humans and machines, often achieved through appearance.  SoftBank Robotics’ Pepper, for example, is a ‘pleasant and likeable’ robot built in a humanoid form and designed to serve as a human companion.  They claim that Pepper is able to ‘perceive human emotion, and ‘loves to interact with you, Pepper wants to learn more about your tastes, your habits, and quite simply who you are.’  In addition to the appearance of robotics is the human voice of digital assistants or indeed navigation aid on mobile phone. 

Not only does technology therefore look like and sound like humans, but we can see the way it is also more likely to behave in a more human way.  The trouble is that many of these tasks are not actually powered by AI but humans operating behind the scenes.  As one sceptic tweeted:

There can be a ‘fake it until you make it’ approach, telling investors and users with start-ups telling investors they have developed a scalable AI technology while quietly relying on human intelligence.  Technology writer Astra Taylor calls ‘Fauxmation’.  Taylor writes:

“AI pushes the human labour out of site – which gives an air of infallibility of the technology but also means it can be more ‘human like’:  if we are in a restaurant we are very aware of the activity ‘behind the scenes’, the clatter of the kitchen, the business of the waiting staff.  Order food for home delivery via an online app and none of this is visible.”

Another example that has had a lot of overage is the role of human moderators of content on social media, clearing it of upsetting and damaging content in a way that technology has yet to deliver.

It is therefore perhaps of no surprise that we assume technology has human like qualities – this is very much what we are encouraged to think even it the reality is somewhat less than appearances.  But if we imbuing technology with human characteristics what are the results of this in terms of what we can learn about ourselves?  In what ways is our relationship with technology offering us a mirror to better understand ourselves?   To explore this, we can turn to one of the topics that animates many, that of self-driving vehicles.

Thinking with others: driving

There has been huge excitement at the prospect of autonomous vehicles with a great deal of investment pouring into the industry.  This is of interest, as the very human nature of driving gives us a close-up view of how humans engage with technology in a very dynamic manner.  As most of us know from our own driving experience, an autonomous car needs to make predictions about human behaviour in real time.  For example, in order to pre-emptively change speed and direction to cope with another driver’s decision to abruptly pull in front of you on a motorway.

The Bitter Lesson seems to be in action here.  Work using deep neural networks has shown how an autonomous vehicle can efficiently identify human actions in streaming videos as motion patterns but they cannot take into account that humans can rapidly change their minds based upon their own thoughts, motivations, and things they see around them. No predictive system functioning purely on past observed motion of the errant car could be accurate enough in such a complex environment as a high-speed busy motorway, without being able to take into account the context and nature of the other agents involved.

Human beings, on the other hand, can forecast others’ future likely behaviour just by quickly assessing the ‘type’ of person involved (from subtle cues of the car driving behaviour as well as make and model of the car) and the scene around them (e.g. the forthcoming junction may give us an indication that the car maybe likely to change lane )

There are many other examples where we can see the way that driving involves a wide array of needing to understand each other’s minds – as Malle would point out, driving necessarily involves communicating our intentions, beliefs, emotions and desires. We are constantly making rapid evaluation of other drivers based on their speed, position in the road, proximity to our vehicle, speed they are travelling, make and model of the car and so on.  If someone is driving at speed and erratically in a beaten-up old high-performance car then we can likely draw the conclusion that it is good to steer clear versus someone driving in a predictable manner within the speed limit in a family saloon.  This is not to say these conclusions will always be right, but they may well also not be unreasonable.

It is perhaps no surprise therefore that accidents involving AVs do happen.  For example, a Google car was driving in autonomous mode in the far-right lane when it encountered sandbags blocking the street. To manoeuvre around the sandbags, the car tried to merge into the central lane. The self-driving car, along with the test driver, assumed the bus would let them in; but the bus driver assumed the car would wait to merge.  “Unfortunately, all these assumptions led us to the same spot in the lane at the same time. This type of misunderstanding happens between human drivers on the road every day,” Google explained in a report.

But what this report does illustrate is the way that when we are driving, we are constantly working to correctly interpret each other’s intentions. Subtle social interaction takes place such as eye contact with drivers being able to quickly confirm each other’s actions and intentions with a mere glance.  The lack of these social cues from autonomous vehicles creates difficulties – we assume the usual rules apply but then find to our cost that they do not.  Why is that?

Shared intentionality

When we are driving on a busy road, we are experiencing the same event with other drivers – we are all travelling with the shared intention (on the whole) of getting to our final destinations safely and in a timely fashion.  Psychologists Sloman and Fernbach point out that when humans interact with others, we don’t simply experience the same event – but we all know we are experiencing the same event.  This is a subtle but important point that changes much – this awareness of sharing our experience not only changes the nature of the experience but also what we do and what we are able to accomplish. When we are driving on a busy road this is exactly what is happening – we are interacting and to a large degree working together to make sure we do not crash, we work together to slow down at appropriate moments, know when to speed up and so on.  Of course this does not always work – crashes still happen – but theese principles are at work.

Through this sharing of attention we are able to share common ground; as we know things that know others also know.  We all can sense what other people are doing and what they intend so as Sloman and Fernbach write: “Once knowledge is shared in this way, we share intentionality – we can jointly pursue a common goal.”

This chimes with philosopher Mary Midgely’s account of human behaviour:

“When we want to understand a real person’s action we always start by looking for the motivational context. We try to spot the particular reason for the act and then to place it on our general map of motivation, a map that we must all use as we try to find our way through everyday life. We ask, was that clumsy remark just a misplaced effort to be helpful? Did it express resentment? Was it even part of a spiteful scheme to make trouble? Or perhaps a bit of all three? …. It is the only way we can start to make sense of the life that goes on around us. Of course it’s fallible, but on the whole it works and its success is one of the things that science needs to investigate.”

We are socially embedded creatures that live in a world that only makes any sense if we locate ‘social facts’, the sets of information that we share about each other.  Sloman and Fernbach point out that this means humans are capable of complex behaviour, as when multiple cognitive systems work together a shared intelligence can then emerge which goes well beyond what any one individual is capable of:

 “Humans are the most complex and powerful species ever, not just because of what happens in individual brains, but because of how communities of brains work together.”

ToM is an important means we have to navigate this network of shared understandings to make sense of the world.  It is not purely something that allows us to connect with other humans more easily, oiling the wheel of everyday relationships, but is essential part of how the world works.  Computers are simply not part of this – there is a mass of meaning that we carry around with us that is not only subtle and nuanced but is also constantly evolving and changing in unpredictable and non-linear ways ways.  The ‘Bitter Lesson’ tells us that a computer simply cannot keep up with this:  the shared meaning, so necessary to the activity of driving (and indeed many human activities), is simply not possible when one of the parties is a machine. 

The premium of ToM

This helps us to understand that there is a premium of ease that we receive as a result of the way ToM operates – we do not have to explain ourselves or seek explanations from others.  Of course, we are taking a risk as we may be getting things wrong, inferring each others’ intentions, beliefs, emotions and desires, it is not a fool-proof exercise as we can all find out to our costs.  But the more we can be confident in our ability to read each other’s intentions then the more we can trust each other. A well-placed act of trust pays significant dividends: for businesses and governments, a level of trust with consumers means that contractual arrangements (to check the validity of the other’s claims) can be reduced, thereby avoiding higher transaction costs. Economists Stephen Knack and Phillip Keefer even found a direct relationship between increases in trust (as measured in survey responses) and increases in national economic growth. Therefore, it is in everyone’s interests that trust flourishes – without it, society as we know it couldn’t exist.

So how does this link to technology?  On the one hand, there is a desire to encourage trust by making technology more humanlike, so we are tempted to apply our human-to-human style ToM mechanisms in this context.  While it might be a good enough representation for some things, as we saw with autonomous vehicles it is a problem when there is an appearance of ToM not backed up by reality.  I remember vividly doing work looking at human experiences of AI and watch as a young Japanese couple tried to get instructions from a Digital Assistant to set an alarm for them to cook some pasta.  They kept attempting to set the alarm and adjusting themselves in tone of voice, choice of words and in the end gave up.  They had assumed an easy human relationship and were surprised by the rigid way they had to deliver instructions – having to adapt themselves to the machine rather than, as might be the case in a human interaction, being able to find a middle ground.

If we stop trusting someone then we ask them to account for their actions.  This links through to the discussion around ‘explainable AI’ which perhaps acts as the alternative to Theory of Mind.  Elizabeth Holm captured with well when she cited the science fiction writer Douglas Adams who imagined Deep Thought, a computer programmed to answer the Great Question of Life, the Universe, and Everything. After 7.5 million years of processing, Deep Thought gave its answer: Forty-two.

She suggested this encapsulates the issue we are facing with technology – “what good is knowing the answer when it is unclear why it is the answer? What good is a black box?” 

As she puts it:

Both an engineer and an AI system may learn to predict whether a bridge will collapse. But only the engineer can explain that decision in terms of physical models that can be communicated to and evaluated by others. Whose bridge would you rather cross?

This is all very well but we don’t want to have to know the provenance of the bridge designer, we simply want to live in a world where we don’t have to think about it.  If we are seeking explanations from the technology that is being used then this is an endless task of seeking understanding and checking, an impossible job for any individual. 

The UK exam fiasco

Perhaps sitting as I do, living in a fairly affluent neighbourhood, I am not targeted with advertising for payday lenders or stopped by the police who are patrolling neighborhoods that an algorithm have them to.  There is little downside for applying ToM principles to technology as it broadly supports my interests (as far as is visible in my day to day life).

But it is increasingly understood that this is not the case for many segments of the population, as set out by Cathy O’Neil in her book, Weapons of Math Destruction.  She chronicles the way that algorithms (the Weapons of Math Destruction as she calls them) that claim to quantify important traits such as teacher quality, recidivism risk, creditworthiness can often have problematic outcomes, reinforcing inequality, encoding racism, enabling predatory advertising to target vulnerable people, and even causing a global financial crisis (in the case she uses, sub-prime mortgage).

To some extent, their use to determine outcomes in less powerful segments of the population has meant that the impacts have not always been fully assimilated by others.  To examine this let’s look at the way in which the results of 2020s A-Levels (UK tertiary education exams) of students who never sat their exams due to the pandemic were allocated based on technology in the form of an algorithm.  Almost 40% of students received grades lower than they had anticipated, resulting in widespread protests by students across the UK who adopted the chant of “fuck the algorithm”.

It’s fair to say that these students felt their trust in the algorithms was misplaced: within days, their protests meant officials to reverse course and throw out test scores that the algorithm had generated.  Many students were stunned by the way in which their individual results of this analysis were so out of line with what they had not unreasonably expected to achieve.  It seems there was no ‘sense checking’ of the findings – schools were not able to review the results to see if they were appropriate for each student.

Part of the problem here is that machines will always operate in accordance with a set of rules and principles:  the algorithm for the exams was operating at the population level not at the level of the individual.  This meant that while the overall exam results were more or less in line with the national results that might be expected, the results for any individual were not so necessarily. [1]

When a student receives a piece of paper with their exam grades on them, it is not accompanied by a detailed explanation of the process that were in place that facilitated the grading of their work.  There is necessarily a trust in the overall system of how these are arrived at that does not require us to become an expert on education policy and practices. 

The endowing of human like properties to an algorithm is not necessarily something that seems as irrational as it may first appear.  We are not necessarily imbuing the algorithm with human capabilities, rather we are using ToM as a shorthand for the context of human values and considerations that we expect the that technology should be operating.  And it is only when they fail to do so that we then seek to dissect the way in which these are operating and demand explanation. 

Our use of ToM is therefore a type of shorthand operation, allowing us to suspend the need to analyse details about how or why we know something.  If we had to operate in a manner where everything is explained, then it is more effortful but also simply not needed much of the time. 

On being human

As science historian Lorraine Dalston put it:

“…no universal ever fits the particulars. Never in the history of human rulemaking have we created a rule or a law that did not stub its toe against unanticipated particulars.”

She makes the point that the human labour involved in avoiding toe-stubbing is often erased from history despite the way that “effective functioning has been dependent on high-level judgment, often performed by those on the low end of the hierarchical division of labor”.

In a way the work that is being done is making the equipment fit for purpose, manually buttressing them the codes and cues needed to work with other equipment (e.g. road layouts, exam setting) but also all of the assumed knowledge and shared intentionality that we have.  Matthew Crawford makes this point nicely when he writes:

… chopsticks are part of a practice of dining that includes, for example, the use of bowls rather than plates, and the preparation of sticky rice rather than, say, loose peas…Chopsticks belong to a different equipmental whole than forks and knives….Rather, in using things like chopsticks, or fork and knife, we involve ourselves in norms: it is just understood that one does things a certain way. These norms are for the most part inarticulate; they are tacit in social practices and in the equipment we use. This is one way in which other people condition the way the world presents itself to us, even when we do not interact with them.

At one level, anthropomorphising technology looks as if we are framing our relationship with technology as a matter between the individual and a specific device or app and in doing so not considering the wider  social, moral, and infrastructural relations.   But in fact, what we are doing is using this as a shorthand for what we expect these wider relationships to be, what we think they should look like, what roles and norms they should adhere to.

Unpicking this allows us to understand the way that our thoughts and way of living do not exist on their own, in some sort of abstract space, waiting for individuals to see them by the light of pure reason. Instead, we can see the way that we make sense of the world and each other is through the context of our shared meanings.

As sociologist Zeynep Tufekci noted, suppose we are standing on top of a high building and I say to someone: “They jumped.” The literal meaning of this seems obvious: That someone has jumped. However, the context adds meaning so that it makes sense to construe the claim “they jumped” as, “Someone has jumped off the building.” If then it becomes apparent that I am reporting my friend has simply jumped up and down once, there would be grounds to complain that I was being misleading that someone had jumped off the building.  In just the same way, the cues used by technology sets a context of shared meanings:  if we are driving, we expect the shared intentionality of navigating our busy road alongside other drivers with all the norms, customs and practices that this brings.  Similarly if we get exam results we expect there to have been a backdrop of our individual performance to have been considered above other influences. 

The way in which technology is described, the way it is designed, the applications it is put to are never neutral:  instead we can see the design features as ‘social affordances’, setting expectations for the way the technology in question works with our shared meanings, intentionality’s and our wider lives generally.  But at the same time, what we have seen in this exploration is the way in which this wider setting is not something that is always crystallised and understood for people, that we use the shorthand of anthropomorphism to express the expectations that we have.


[1] Another aspect of the fiasco was the way the algorithm put more weight on the past grades element of the equation if there were fewer than 15 students in a particular subject at a particular school. That meant students at smaller schools (inevitably fee paying) schools were more likely to benefit from grade inflation than those at larger schools.   Inherent biases were made explicit, strangely revealing to people what they already knew, that our education system is set up in a way that privileged small, monied segments of the population that could afford private education (or at least had the means to attend state schools in more affluent areas).  So the algorithm was seen to ‘bake in’ aspect of unfairness of the underlying system that, while widely known, are not brought to public attention in a particularly salient manner and therefore can be conveniently ignored.