The economy is a highly complex system. It’s driven by billions of individuals, each with their own complicated motivations and actions. These combine together in such a way that they are much harder to understand than the sum of their parts. In order to be able to analyse this system, economists have made simplifying assumptions about agents in the economy, namely they make rational decisions based on all of the information that is available to them.
Rationality assumes that humans always make the best decision they can. It assumes they think about each economic action they take for as long as needed to arrive at the decision that maximises their preferences. For example, when choosing a meal, a truly rational agent would make the optimal choice by balancing the nutritional benefits against the taste and other factors such as its environmental impact. However, people frequently make decisions that appear not to be in their best interests, such as choosing an unhealthy diet or failing to save for retirement, even with or without all necessary information.
These assumptions led to the coining of the phrase homo economicus: the idealised human that makes optimal decisions based on perfect knowledge about every economic action they take. This species enabled breakthroughs in economic theory and understanding. However, there were many who doubted the existence of homo economicus.
Despite the success of the simplifying assumptions that the economy consists entirely of homo economicae, researchers have begun to question the extent to which humans are rational and blessed with the ability to perfectly evaluate the information they have available. It is argued that humans are not rational enough and lack some of the capabilities necessary to make informed decisions about how to obtain maximum utility.
A particularly ardent group of dissenters on the assumption of rationality are behavioural economists. They argue that instead of the unlimited thought power attributed to homo economicus, humans are better modelled by agents that use simple heuristics to make decisions that would be impossible otherwise. In their eyes, humans don’t possess unlimited rationality, but bounded rationality. A lot of these heuristics have been claimed to pose an evolutionary advantage: the ability to make instantaneous, slightly sub-optimal decisions is more likely to keep you alive when the going gets really tough than ponderously thinking through every avenue of possibility and arriving at the absolute best thing to do. A famous example of irrationality is humans’ imbalanced aversion to negative experiences: a bad experience, such as a financial loss, will be felt about twice as strongly as a gain of similar magnitude. Back to our meal example, our economic actor may stick with the meal they knew was ok last time, rather than try the special on the menu which is equally likely to better and worse. This loss aversion helps humans stay alive when they are at or below the subsistence level, but doesn’t help people make optimal economic decisions in developed societies.
The assumption of perfect human judgement does not need a new field of economics to refute it. When making a decision humans are clearly not in possession of all the facts. Even if we were presented with all the information relevant to make a decision, the limited capacity of the human brain means it is unable to absorb it. The limits of the assumptions in homo economicus mean that human economic agents are forced to behave irrationally in the face of complex decision-making. Understanding how and why humans make these sub-optimal decisions, and how this sub-optimality affects the economy, forms a large part of modern economic research.
Humans have long overcome their deficiencies by using tools and have developed machines to allow ourselves to inch closer to the idealised economically perfect man.
Just as humans have made tools to increase our physical ability, we also have tools that increase our mental capacity. A simple example of this that augments our intelligence is a calculation device, such as a computer. Humans can use this to automate some of the more mechanical aspects of rational thought. For example, when making decisions about what stocks to buy, humans previously relied on gut instinct to try to pick a winner. Nowadays, computers allow traders to derive mathematical signals that can be used to indicate future performance. As computers have become more powerful, the number and complexity of signals has increased, and it can be tricky for humans to differentiate between noise and true information.
Humans have also developed tools, such as the internet, to improve the completeness of information. This has greatly increased the amount of information we have available to base our actions on. When buying an item online, we can see what price is offered for the exact same item at multiple outlets, full technical descriptions of the item, and can easily see recommendations for items that are similar to what we are currently looking for. Interestingly, this information can be presented in such a way that biases our decision towards purchasing items the seller would prefer us to buy. For example, a retailer can prioritise the advertisement of items that generate a better profit margin for itself. There is also the issue that greater amounts of information can overload the human thought process, and lead paradoxically to less optimal decisions. Look again at our online shopping example: when presented with such a vast array of items, it is harder to scrutinize the specification of each product individually, and we are more likely to make snap decisions based on shallow understanding.
The technical innovations that augmented homo economicus nudged it towards being more like the prototypical perfect economic entity. However, problems of low cognitive bandwidth still dominate our economic decision-making and this is exacerbated by ever more information becoming available. Recent advances in technology may allow us to not just approach homo economicus-like ability, but to evolve into an entirely new species of economic agent: machina economicus.
In machina economicus the limitations of human economic agents to be like the idealised human economicus are mitigated by replacing human decision makers with intelligent machines. These machines are able to not only process more information and make better decisions, but they are also able to anticipate negative consequences and interact more effectively with other automated agents.
It is widely known that AI is poised to transform the economy, although the replacement of vast swathes of the workforce is just the tip of the iceberg. AI-powered automatic economic agents can represent humans to allow a fairer, more efficient economy, where humans are not only materially better off, but can concentrate on tasks more suited to the creativity of human intelligence.
AI was born in the 1950s, and its initial incarnation focused on logic and formal reasoning, termed symbolic reasoning. In symbolic reasoning computer scientists produce long lists of known facts about reality consisting of logical relationships (known as an ontology). They then use inference rules that can generalise these facts to new situations to answer questions about the world using perfect verifiable logical reasoning.
Ontological representation of a very simple world consisting of people, food and feelings, and the relationships between them
The above ontology represents a very simple situation. Circles represent concepts, and edges between them represent relations. Edges between concepts of the same colour implicitly have the relationship “is a”.
Given this ontology we could ask the question:
How bob stops hungry?
Then using the ontological rules:
bob has a apple
apple stops hungry
apple is food
bob is people
people eats food
Through ontological reasoning we can arrive at the answer:
bob eats apple
The ability to reason infallibly is a very powerful tool, and is a requirement of many definitions of rationality. It enables an agent in possession of all the necessary facts in their ontology to make perfect economic decisions on a human’s behalf. If an ontology contained a list of all restaurants in an area, all their meals and ingredients, and information about your personal preferences for such items, it’s easy to answer questions such as: “What cheap Indian restaurants are in my vicinity that make my favourite dishes?”. Your representative machina economicus would be able to reason about your preferences for taste, price, environmental impact and match them with all the meals in all the restaurants in the area. Ontological reasoning is already being used by Google to enable better fact-based searching, and by Amazon to drive Alexa’s ability to answer questions. However, its true value will only be realised when it can directly empower an individual’s economic actions, and is not used solely for the benefit of large monopolistic corporations.
Rule-based methods suffer from a problem of scalability. They only work in limited domains due to slow inference methods, and require human experts to manually enter the rules of the ontology. However, in small domains with existing rules it is hard to beat the rationality of symbolic reasoning, and the perfect economic decisions it enables.
More recent advances in AI have taken the form of machine learning, where instead of rules for intelligence being manually defined by expert human operators, intelligence is learned from data or by interacting with an environment.
The massive amounts of data that are generated by economic processes today are difficult for humans to interpret, but is perfect for deep learning algorithms. These methods learn a correspondence between a provided input signal and an output target. They have state-of-the-art performance to resolve problems such as labelling the content of images, recognising human speech, and translating one language to another. As more economic entities generate more data, the potential benefit to including deep learning in direct economic decision-making grows. A company may currently have a stock control team whose job it is to monitor resources in the supply chain. They have to amass data from across the business to decide when to order a particular stock, and how much they require. These humans are not a perfect team of homo economicae, but instead may make incorrect, slow decisions based on only a portion of the information available. This team could be replaced by a single machina economicus that takes in all information from across the business. By learning from previously correct choices, the machina economicus would then be able to order the correct amount of stock at the optimal time, and subsequently make instantaneous and greatly improved decisions on resource acquisition.
Deep learning enables an agent to learn to predict what chemicals are needed when, directly from industrial processes.
The previous example relied on prior existing data provided by human experts, similar to how rules were needed for symbolic reasoning, which may or may not be available. In order to allow machina economicus to not just supercede humans in existing domains, but instead to push into domains impossible for humans, we need another form of machine learning, called reinforcement learning. Using this approach, an agent learns to maximise its rewards by receiving feedback on how its actions affect an environment. Put another way, an agent generates its own data to train itself by interacting all on its own. This is the type of AI that DeepMind used to beat humans at Go, where the reward was defined as a win in a game and the actions were the individual moves. To apply this form of learning to economic decisions we can define the agents reward as economic utility, which could be the profit of a company, or the satisfaction rating of a human who has purchased a product. The agent could then try economic actions, of increasing complexity, and receive feedback on economic utility, teaching itself how to behave best for the entity it represents. This technology has the potential to allow machina economicae to be deployed into the economy, and for them to learn automatically how best to represent their owners.
Both of the above types of machine learning require a lot of information: training data for deep learning, and feedback for reinforcement learning. If they come across a situation that they haven’t encountered before, they can be overly confident about making a decision, despite lacking the understanding necessary to fully address the problem. Bayesian machine learning offers a solution By introducing prior beliefs about a system, we gain two benefits: better learning efficiency and a measure of uncertainty. A measure of uncertainty allows an agent to describe how certain it is about decisions that it makes. It can use this uncertainty to assess the risk of economic decisions, and if it deems a decision to be too risky it can give decision-making responsibility to another, more suited automated agent, or even ask a human to make the decision on its behalf. This allows machina economicus to be make decisions with small amounts of information and thereby adapt its behaviour accordingly This contrasts to the loss aversion heuristics that humans use: rather than have to err on the side of caution, machina economicus can more effectively balance risks against benefits.
The red shaded section shows where there is limited information. The prediction has much higher error bars to match, showing that it is aware of its uncertainty.
AI isn’t just changing the way that individual agents can make economic decisions, it is also transforming how they can act together to form economies, coining the phrase econima machinus. Multi-agent systems (MAS) is the field of AI concerning how agents communicate, behave and learn in the presence of each other. MAS looks at problems where agents either have to collaborate to solve a common task, or compete to reach an individual goal. With the speed that computers are able to communicate, this allows individual machina economicus to collaborate and compete within an econima machinus at speeds that human agents in the traditional economy are unable to match.
Humans perform best when information and decision-making power is centralised, for example in libraries, and in hierarchies with leaders/managers. This concentration is necessary for humans due to their limited ability to communicate disparate information effectively. By contrast, automated agents can seamlessly share masses of information, so responsibility for economic decisions can be delegated as needed. This makes traditional centralised architectures obsolete. As different information is needed for certain decisions, the structure of information and decision-making power can be dynamically altered by the participating agents to become more efficient.
Distributed ledger technology, such as blockchain, enables economica machinus by allowing trustless decentralised economic transactions to be stored as a complete and immutable record of all past decisions. This decentralised mechanism for financial exchange provides the essential missing component that will enable multi-agent systems to be deployed in a wide range of settings that were not previously considered possible, such as in smart grids, supply chains or in healthcare. Smart contract technology also enables autonomous agents to commit to actions in the future, and thereby participate in complex economic activity such as auctions and decentralised exchanges.
Humans can behave sub-optimally as a result of our cognitive limitations and the increased information volume of the modern world. We continue to try to augment our abilities with technology, but this will not turn us into agents that make the best economic decisions. Various strands of classical and modern AI allow a new species of economic agent, machina economicus, to emerge and greater autonomy of economic agents will allow greater automated representation.
Previous analyses of the economic effects of AI have focused on the negative consequences on people’s jobs. Machina economicus will allow beneficial effects on the side of consumption: acting on people’s behalf to make optimal economic decisions, with the potential to free people from the mundane aspects of economic choice, and grant them more free time to spend on enjoyable choices and actions.
As the number of decisions being automatically performed by machina economicus increases, the emergence of a new type of economy, economica machina, will yield ever greater efficiency and fairness. Such an advanced construct requires more advanced underlying technology architecture, which blockchain and distributed ledger technologies can fulfil.