The usual suspects — the resurgence of Lombroso and other pseudo-sciences in automated facial recognition

In the second half of the XIX century, Cesare Lombroso became popular for his theories about natural born criminals and how to identify them. Mixing social darwinism, phrenology, and physiognomics, his contributions gained him the title of founder of criminology.

For example, did you know that criminals, being closer to animals, have a longer tailbone? Or that being left-handed is one of the marks of born criminals and ‘lunatics’? And beware, criminal women present an ‘atavistic’ face and a high number of moles!

Lombroso’s work was comprised of pseudo-sciences and racism, packaged neatly in a format that seems to follow the scientific method, and embellished with pearls of wisdom and aphorisms. This mix contributed to make him popular throughout the second half of the 19th and the first half of the 20th century.

At their core, Lombroso’s theories were deterministic. He affirmed that the physical and physiognomic characteristics of an individual would determine the likelihood of possible criminal behaviours.
Lombroso’s research was immediately rejected by some of his contemporaries. Napoleone Colajanni, a Sicilian criminologist and politician, harshly criticised Lombroso’s positivist approach and pointed out the terrible consequences that his theories could have at the social level (this should not come as a surprise, seeing how Lombroso, a Northern Italian, applied his theories first and foremost against Italians from the South). Charles Goring, an English criminologist, responded carrying out a study that showed little to no statistical evidence supporting Lombroso’s claims.

With time, the connection between physical features (shape of the forehead, physique, positioning of the eyes, presence of moles, shape of the nose, and so on) and certain criminal behaviours proved first weak, then non-existent. Research confirmed that criminality could be present in people without a prominent forehead and an ‘atavistic’ face.
Furthermore another determinist theory, focused more on the environmental and social causes that push individuals to behave criminally, became more and more popular and is still at the basis of some important works of today’s criminology. Throughout the second half of the 20th century, research showed that an individual’s aspect was not the symptom of an innate criminal personality, but the manifestation of environmental factors, such as poverty or malnutrition, that played a role in pushing individuals towards criminality.

Nowadays, Lombroso’s theories are considered by the academic community nothing more than pseudoscience and the product of a deeply problematic vision of human beings and of society that was popular in the decades leading to the two great wars.

Or so I thought.

Periodically, research pops up in various sectors that offers new lifeblood to the ghosts of past pseudo-science. From psychology to computer science, the list is unfortunately long.

Some readers might remember, for instance, the researchers that claimed to have programmed an algorithm capable of identifying with staggering precision the sexual orientation of individuals from their faces. That research hit the pages of the major news outlets and was publicised to the general public. In the first few months of 2020 only, two studies emerged, the first in a journal and the second in an announcement by Harrisburg University (later deleted after protest from the scientific community). Both had as their objects algorithms capable of identifying potential criminals using face recognition. The first of the papers even goes as far as affirming that the research was triggered by Lombroso’s works (and quotes among the sources the ’gay face AI’ paper).

Developing an algorithm or a research method based on the premise that specific features of the face could be used as generic traits to identify certain behaviours or personal preferences is severely problematic. Let’s see why.

First, it is bad science. It should be superfluous at this point, but repeating helps: no, an algorithm cannot discern sexual preferences or criminal tendencies based on facial features. The idea that there are generic facial traits connected to certain behaviours or personal preferences is in itself preposterous and disproved by proper science. Presenting these theories as reliable sources re-inserts into circulation ideas that were abandoned and disproved. It does not help science, it weakens it, bringing it back 150 years. Using it to develop an algorithm or any other technology perpetrates scientific falsities.

Second, it hurts people. Using these theories as a starting point, quoting authors like Lombroso in a theoretical framework, or simply referring to them as valid scientific sources infuses a study with a certain set of values. Claiming that a project was inspired by Lombroso’s work means that the algorithm developed with that research will operationalise the idea that people with certain facial features are natural born criminals. Discrimination and bias permeate such research from its very foundation, and its product, the criminal-detecting algorithm, will enhance and exacerbate harms to certain categories of people and vulnerable subjects.
Having established that the use of pseudo-scientific authors or theories in science is bad, it is inevitable to wonder: how is it possible that the editors and publishers allowed the abovementioned articles to see the light? Why didn’t they reject such papers, raising an objection to the use of such sources?

Publishing mechanisms make it easier for things like these to happen. The pressure put on researchers by the publish-or-peril academic mentality can translate into sloppiness, lack of time to properly build a theoretical framework or to investigate in depth the sources to use. This is even worse in certain disciplines such as computer science, where funding mechanisms exacerbate competition and push researchers to publish as much as possible and constantly put forward new ideas. Consequently, publishing times also have to be short, cutting as much as possible the time dedicated to reviewing and revisions. Publishers also have a greater economic incentive to publish a great number of papers, therefore possibly underestimating negative reviews. These circumstances can potentially translate into articles being written and published without enough time for appropriate research and revisions.

Reviewers themselves are often very busy and, let’s not forget, carry out reviews completely for free on top of their already incredible workload. This makes them more prone to miss red flags. They might also not be aware of the issues connected to certain sources, authors, or theories. The praxis is to only use reviewers in the exact field of the paper. A reviewer coming from computer science might not be aware of the problems connected with sources from other disciplines, as could be the case for Lombroso.

As with the algorithm that allegedly detects sexual preferences from faces, including pseudo-science in research legitimizes these ideas not only with the scientific community, but also with the general public. When a newspaper or a magazine picks up a scientific article and ‘translates’ it for the general public, some nuances inevitably get lost. The public might be induced to think that if scientists use pseudo-scientific theories, then maybe they are not so wrong after all. This is a dangerous side-effect. Especially now, with the diffusion of conspiracies and anti-scientific movements affecting the very fabric of civil society, scientists must be responsible in the way they use their sources.

These reflections prompt some recommendations.

With regard to scientists, researchers, and authors, it is important to make them aware of the ethical and social issues connected to their job. Often times, algorithms like the ones mentioned above are developed as an ‘exercise in style’, just to test and show what an algorithm can do. In these cases, the research presented by scientists is like the test drive done by a car producer in a circuit, with a professional driver: they show off the features and the technicalities of the product, while at the same time advertising what the developers can do. With these software, computer scientists show things like accuracy, internal consistency, or low margins of error. They show what they can do. But just because you can, doesn’t mean you should.

Scientists and researchers need to remember that algorithms and other technologies are not mere exercises in style but concrete artifacts in a socio-technical system. They can legitimize discriminatory beliefs or actions. They might be employed in practice by governments, companies or the general public in ways that harm human beings.
With regard to editors and publishers, the pressure to publish cannot stand in the way of high quality, responsible science. It is pivotal to give the right weight to negative reviews and not underestimate red flags. And in order to spot the red flags in the first place, the more diverse the pool of reviewers, the better: why not entertain the idea of instituting ethics boards within journals, or a meaningful review given by experts of another discipline? The ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT) has already shown the way in January 2020, assigning to every manuscript a reviewer from a complementary discipline. In this way, authors of scientific papers obtained reviews also from social studies and humanities (and vice-versa), such as ethicists, sociologists, or legal scholars, offering different perspectives on their research. It is, however, still the only one to have done it, so far. Hopefully, in the future the ACM FAccT will no longer be alone, and this will become a standard praxis for journals and conferences of all sectors.

There is a vicious circle of pressures to publish quickly and abundantly, market incentives rewarding ‘easy’ technological fixes to complex issues, and computer scientists not being aware of the effects of their work on vulnerable people.
Scientists and researchers can break this vicious circle: promoting awareness of the ethical, social, and legal aspects of scientific research and demanding high quality, multidisciplinary, effective reviewing can prevent new pseudoscience-inspired articles seeing the light in the future.

Rather than asking themselves whether they can create an algorithm to identify and predict personality traits of individuals from their appearance, scientists need to ask themselves what the implicit biases are underlying such a concept. They need to reach for the rest of the academic and expert community, including social sciences and humanities, to test their concepts. This way, they can immediately turn away from ideas that are not only ethically unacceptable, but also demonstrably not science.
They need to look for more and diverse perspectives on their studies, in order to assess their own ideas and work. Reaching out to other sectors and studies can provide researchers with the knowledge necessary to create algorithms that are original, innovative, inclusive, and beneficial for society at large.

The development of new technologies needs a Copernican revolution: researchers, experts, and the general public must radically change their perspective in order to move away from pseudo-scientific assumptions and myths, and open the way to scientific research that promotes values of equality, social justice, and dignity.
The next generations of scientists and researchers must be given the right tools to carry out such a change. We hope this blog post can be a useful warning, and a constructive critique, to help show students and researchers that a different approach is possible.

As of June 2020 over 2400 experts and practitioners from different fields have signed an open letter addressed to Springer, the publisher of the abovementioned article quoting Lombroso. The letter asks Springer to block the publication of that article and goes beyond, demanding all publishers to stop publishing articles containing attempts at predicting criminal behaviour from face or other physical traits. The harms that this kind of research perpetrates are also clearly explained in the letter.
This initiative takes a step in the right direction and deserves everyone’s attention and support: click here to view the letter.

About the project

Places and populations that were previously digitally invisible are now part of a ‘data revolution’ that is being hailed as a transformative tool for human and economic development. Yet this unprecedented expansion of the power to digitally monitor, sort, and intervene is not well connected to the idea of social justice, nor is there a clear concept of how broader access to the benefits of data technologies can be achieved without amplifying misrepresentation, discrimination, and power asymmetries.

We therefore need a new framework for data justice integrating data privacy, non-discrimination, and non-use of data technologies into the same framework as positive freedoms such as representation and access to data. This project will research the lived experience of data technologies in high- and low-income countries worldwide, seeking to understand people’s basic needs with regard to these technologies. We will also seek the perspectives of civil society organisations, technology companies, and policymakers.