When Mary Shelley wrote her novel ‘Frankenstein’ at the age of 20, she struck a chord with the times. The phenomenon of electricity, which caused dead frogs’ legs to twitch wildly, had stirred people’s imaginations. In retrospect, it seems little wonder that imaginative poets would take up the phenomenon and create a character in a novel that would be given life by this mysterious force.
Every new technology tempts us to describe natural systems or technology by other technology. When mechanical machines entered our world, we explained the orbits of planets and stars in mechanical terms. With electricity, we compared the workings of the brain. The Internet was described by a U.S. senator in terms of how pipes and wires work in a drain installation.
With artificial intelligence, another technology is now coming our way that laypeople, like electricity in Mary Shelley’s time, understand only a little about and to which almost mythical powers are attributed. Just as the monster from Frankenstein’s Creation, which was actually misunderstood, represented the embodiment of electricity, so the menacing Terminator symbolizes AI.
At the same time, the fundamentals of artificial intelligence, such as machine learning, lend themselves to describing other processes. Such analogies occasionally allow us to take a different look at a supposedly well known fact. In the present case, we can subject democracy to such an analogy with AI. And as it turns out, some of the similarities are striking. But let’s analyze the individual components of AI for a moment.
In the current state of machine learning (ML), large amounts of data are required as information, and algorithms in the individual nodes of the neural network, arranged in multiple layers and in parallel, filter information from the data, transform it, and pass it on to other nodes for processing. Once through all nodes, the machine learning system spits out a result, which is then evaluated by human testers. Each evaluation of the result by humans serves as feedback to the system, which then adjusts the parameters in the algorithms of the individual nodes by processing this and new data again, thus making the result better step by step.
With this approach, large amounts of data (information) are needed today to adjust the system appropriately, or as it is called, to train it. It is therefore well suited where large amounts of data are already available, such as large amounts of cat images on the Internet, provided that one wants to train the system to recognize cats, or can generate them, such as by driving autonomous test vehicles on test tracks or public roads in regular traffic.
For example, the recognition rate of objects in images is used to evaluate how well the machine learning system is trained. In what percentage of cases did the AI correctly recognize the cat? And how does this recognition rate compare to human comparison values? The AI experts have to avoid several pitfalls when training the system.
This starts with the data, which must be appropriately selected and curated to recognize all contingencies at a similarly high rate. For example, if there are a disproportionate number of light-skinned faces in the training data for face recognition, then the recognition rate for dark-skinned faces is usually below average. In the industry, this is referred to as ‘bias’. The term is not chosen by chance, as it is used in psychology and behavioral research when referring to various types of cognitive bias. Our own faulty thinking, perception, memory and judgment are automatically and for many unconsciously incorporated into the data with which AI systems are trained.
The bad news is that we will never get bias out of an AI system. We can reduce it, but we always run the risk of reinforcing it elsewhere.
Other difficulties arise in how much the machine learning system is allowed to adjust the parameters in the algorithms of the individual nodes. If it is allowed to change too little, then the system becomes rigid and can be thrown off kilter by even slight changes in the information. The system has been ‘overfitted’, i.e. highly optimized for certain data. The opposite case occurs when parameters are allowed to adjust too much automatically. Then the range of behavior for the same or very similar data can be too wide, and the system no longer reacts predictably and too surprisingly.
The Similarities with Democracy
Machine learning and democracy thus have certain similarities. Both have a large number of actors who pass on and influence information. Decisions happen through the collaboration of the many participants. In machine learning, these are the nodes; in democracy, they are the voters and decision makers, as well as institutions and other organizations. Both systems are relatively less hierarchical, and they suffer from the same weaknesses. How the information is passed on, and how flexibly or rigidly they respond to it, make them more or less effective. The adjustment of parameters in the neural nodes corresponds to the natural outflow and inflow in democratic institutions. If there is no exchange, no adjustment, then the system becomes too rigid and can only deal with new and divergent information in a limited way effectively. If there are too many outlets and inlets, then the institutional knowledge and network is lost, which again leads to inefficiencies.
These are the more technical aspects in which machine learning and democracy are similar. Where they clash is the moment when artificial intelligence interferes with democratic processes. AI systems with bias – which, as mentioned earlier, can never be completely eradicated – can disadvantage voters or prevent them from participating in the democratic process. For example, if minorities by color, gender, or ethnicity are more criminalized or otherwise disadvantaged by such systems, then that negatively impacts the democratic process and their representation, and thus the treatment of such systemic disadvantages through bias.
Dictatorships, which are mostly hierarchical, can process information and make decisions very efficiently in stable environments. Not all groups need to be involved in the decision-making process. This type corresponds to deterministic AI as it was developed in the early days until two decades ago. But it was abandoned because the decision trees became too complex and too prone to crises. Every decision had been pre-thought by the AI experts. Information that had not been pre-thought in the tree led to a problem.
Dictatorships quickly become inefficient when confronted with uncertainty, as the Soviet Union’s five-year plans showed, for example. New technologies and demands could not be anticipated and led to a shortage of supply that cascaded to all areas of the system. Hierarchical systems such as dictatorships tend to be very interested in information – after all, they deliberately spy on everything and everyone and build up a huge apparatus for this purpose – but disagreeable information that does not fit into the ideology is ignored or the supplier of such information is destroyed.
The advantage of AI as we experience it today is also its greatest disadvantage. In areas where sufficient amounts of data do not exist or cannot be created, it functions suboptimally. Also, we humans tend to look at information and patterns from the information we recognize as being outside of our expectations or experiences from a different angle through new thinking frameworks or models. We question the data by trying to identify causalities, the reasons for their deviation. Today’s AI is not able to recognize such causalities, to develop new thought frames, or to create such analogies itself, as even small children can do playfully in pretend play. A piece of wood quickly becomes a car, a celestial phenomenon a UFO with aliens. An AI that can only recognize cats doesn’t even know what a cat is and what it is for. Let alone does it know about dogs, leaves, the sun or the wind.
The frames of thought and the analogies that AI systems thus use are provided by humans. And humans feed them with data generated and selected by humans. Democratic systems, on the other hand, thanks to human actors, can make precisely these adjustments to the thinking frameworks and models, and bring in new information when the existing data and results can no longer meet the challenges.
Democracies are particularly good at this, at least better than dictatorships. Dissenting opinions, experiences and information are suppressed in dictatorships, and in democracies, if not encouraged, then at least tolerated. This preserves important perspectives and experiences for the system, which are flushed from the margins to the center when there are changes and uncertainties in the surrounding conditions and, thanks to their characteristics, can contain approaches to the new challenges. A democracy receives this kind of information and frameworks of thought, albeit often rather accidentally; dictatorships, on the other hand, actively seek to destroy them. AI itself doesn’t even know about these concepts.
The Optimization of Democracy
Now, mathematicians in particular are trying to create approaches that, if not eradicate the weaknesses of democratic systems through appropriate algorithms, will at least optimize them. At Harvard University, for example, such a course has been offered since this year by the School of Engineering and Applied Sciences. If already in the first democracies in Athens male and adult citizens of the city were admitted to elections, who also had to be present, and women or slaves could not decide, then also the US democracy was limited first times to men. And they received then according to the number of slaves in their property and/or in their state, appropriate election shares.
However, democracy is not an optimization problem like in mathematics. Many democratic systems have been deliberately constructed with certain biases for various reasons. The electoral college system in the U.S., as in Great Britain, gives a majority party a larger share of the vote, with the desire to create a functional government more quickly. In the individual constituencies, the winners then receive all the mandates. In other countries, there is a stronger representation on votes, which should lead more often to coalitions and thus to a need for compromise. Thus. Other possibilities of democratic bias are conceivable and used. For example, certain regions or metropolitan areas are favored more than others because of their economic strength or non-voting population shares, such as a large number of children.
As we can see, machine learning and AI allow us to take a fresh look at how democratic systems work. The interaction of democratic actors resembles that of neural networks at some points. The methods that ML/AI system improve could form an approach for democracies. And, thanks to the shortcomings of today’s AI, they sharpen our focus on the real strengths of democracies and, conversely, what AI and ML can learn from democracies.
Theoretical physicist Sean Carroll talked about this topic with his guest Henry Farrell on the Mindscape podcast and inspired this post.