The boundaries between the digital world and the physical world have long since fallen. We live connected to our mobile. The walls of social networks are visited by more people than the municipal squares and our lives are influenced by decisions made by algorithms. For good – critical medical decisions -, for more dubious purposes – political influence – or simply to optimize certain processes, such as finding a taxi or stealing seconds of your attention.
Many of these algorithms feed on the data that each of us and our lives produce -the big data– that, together with the increase in computing capacity, have allowed the current explosion of artificial intelligence (AI). This term was proposed for the first time without much fanfare at the 1956 Dartmouth conference by John McCarthy, Marvin Minsky, Nathan Rochester, and Claude Shannon (probably the most undervalued scientist of the 20th century). Today the IA concept is presented in a nebula within the reach of the understanding of a few, even when it is in the mouths of scientists, businessmen, politicians and journalists. We are at a turning point on its use, and it is important to demystify it, to know what it really is and to decide as a society what we want from it.
- The mask factory of the AI
In recent years a family of algorithms called deep learning It has revolutionized the field of AI. It has broken all the records distinguishing photos of dogs and cats on the internet, translating languages or winning the ancestral board game Go. However the deep learning It is not magic. And there is little intelligence. The deep learning It's like a plaster mask factory. How does it work?
A million people stand in line at the factory. The factory craftsman puts a wet plaster mold on the face of the first of the row – a young girl. The plaster is molded according to its features. The second person arrives-this time a boy. The craftsman puts a second mold to make the boy's face. A third person arrives; it's a girl. The artisan takes the initial mold that is still wet and puts it on the new girl, so that the mask has features common to the first and third person -the girls-.
And so on. The craftsman is modeling two masks while the plaster is still wet, one man and one woman. Once the tail is finished, the craftsman lets the two average masks dry until they are ready for use. From the next day, an AI system made of masks will automatically open the door of some offices. Every time someone arrives at the offices, the masks will be tested and the door will only open if the visitor's face fits into one of the two that have been molded -for example, only women will be able to enter-.
This metaphor describes exactly how neural networks work deep learning. The AI for facial recognition is nothing more than a mold that literally keeps the common features of the pixels of the photos that we show you of a certain class: man or woman, white or black, dog or cat, or the facial features of the population from Oviedo, New York, Paris or Kampala.
In this process it is important to distinguish several concepts and their metaphorical correspondence:
- The people who put their face – their data– to create the mask. The more data, the better characterization the mask will have.
- The mask factory, which needs the craftsman. The tools of the craftsman are the algorithms. The interesting thing is that most of the tools are open source, they are available at no cost on the internet. However, only a few experts (artisans) in large companies and academia are able to use them. And not because it is very complicated, but because of the lack of a plan so that many more people can access this training.
- Masks – the Models– that are created with those algorithms. Once created, the masks can be used outside the factory – they do not require special machinery – in a very simple and almost free way: they can be used to open doors, unlock the mobile phone or charge without going through a supermarket box.
- Ethical algorithm factories
A few months ago, employees of an IA multinational sent a letter to their CEO where they argued that they should not delegate the ethical responsibilities of the technologies they developed. The factory can create masks that are coupled to a drone and transform it into a small missile guided by face recognition masks. A mask applied to a reading of your DNA can identify that you are likely to develop a lethal disease. Is it reasonable that, based on that reading, no insurer will make you health insurance?
The ethical aspect must be present in the whole chain of the AI, not only in the purpose of the use of the mask. Imagine that we want to create a mask to decide which curriculum to select for the interview from among thousands of candidates who want to work in our company. To create the mask, we use the data of the people that the company has hired and have triumphed in the past, so the mask will have the perfect employee's pattern. That employee would turn out to be a man, since in the company there are more men than women and the mask we have created has identified the masculine gender as part of the perfect employee pattern. The curriculum selection mask will discriminate by gender. The masks (the models) are the mirror of the data that is used to train the AI system. Only that.
A few months ago I shared a week with 1,000 AI doctoral students from around the world. Surprisingly -or not- the ethics and social impact of AI did not come up for debate at any time. Like the doctors or nurses, data engineers and scientists should subscribe a Hippocratic code with ethical principles. And being very practical, formations of ethics should be mandatory both in academia and in companies.
I propose that those whose actions are worth training models are compensated each time they are used "
Probably the greatest threat to democracy as we know it is the videos deep fake: videos generated by AI that seem totally real and are based on millions of archive videos accessible from the internet. This tool will allow very soon (before the next elections of any country in which you are thinking) to create, for example, a video with a totally real appearance where the chosen politician is saying exactly the words that the creator of the artifice wants. It will be the evolution of fake news who also uses the video: everything can be a lie.
- Artificial intelligence in the supermarket
The masks can be easily packaged, since they do not need the factory or artisans to be used. They could be bought and sold at the supermarket. A mask, once manufactured, has no costs to use or maintain it. So the business model on which you want to initially base IA companies is to charge for each time you use the mask.
Let's put 1,000 waiters in the factory queue. This time we are going to create a mask that captures the movements necessary to put a cane – a dynamic mask or robot -. After studying the 1,000 waiters we can put the mask to serve beers in a bar. In fact no longer need more waiters, since we have robots that do not have to pay. The knowledge of the waiters has been coded in the mask and now the waiters are dispensable.
The doctors of the Spanish health system spend many minutes entering data in forms that are also queues in the mask factory. And to whom will these masks belong? To the companies that collect this information – sometimes those that should simply give computer support – when they should be, in part, from the public system, doctors and patients.
The one who has amassed enough data to make a mask will have the power to replicate an action almost free of charge. This will give an unparalleled advantage to the first to arrive, something that companies and some governments are aware of. What some have already described like the new cold war.
- The royalties that you owe artificial intelligence
Should waiters receive compensation each time they wear the mask (put a beer) that they trained with their data? Should physicians receive compensation each time the AI detects a tumor based on the diagnoses they made in the past?
My proposal goes beyond the AI companies collect your data for free or pay for your data only once. I propose that those whose actions are worth training models (create masks) should be compensated EVERY TIME the mask is used. Each time the robot waiter puts a beer, part of the payment for that beer will go to those waiters whose data showed the robot. A kind of royalties, or copyright system similar to the one that exists in the music industry when we listen to an online song. But with any AI system based on neural network training. A mechanism to enhance creativity and human relevance, redistribute part of the productivity increase produced by AI and try to equalize an increasingly unequal world.
The race to get the masks without anything in return has already begun.
Miguel Luengo-Oroz is chief data scientist of UN Global Pulse, a United Nations think tank that uses the big data and emerging technologies to act in development contexts and humanitarian crises.