unthinking
photography

This text was commissioned as part of a research project supported by the Paul Mellon Centre for Studies in British Art that traces the transformation of photographic practices from the socially aware political contexts of the 1980s to the realms of computational, machinic and AI practices of today. Contributors were prompted by each being loosely assigned a chapter title of Victor Burgin’s ‘Thinking Photography’, leaving the types of response open: for this text, the prompt title was ‘⁠On the Invention of Photographic Meaning'.

The commissioned texts will be published fortnightly on Unthinking Photography, and a new exhibition and online commission by Planetary Portals is now on view.

“How about my interpersonal relationships? What does my face say about me?” I asked the face reader while subtly checking if my camera is rolling.

“The shape of your eyebrow… Your eyebrows are very distinct and imposing,” he replied. “That suggests you’re quite stubborn when it comes to work.”

“The shape of your eyebrows represents your life from 31 years old onward,” the face reader maintained firm eye contact with me. “You’ll get an opportunity during these years. Do you follow me? This opportunity will let you realise your capabilities and your potential. Do you understand?”

I’m getting a face reading on Temple Street, the heart of Hong Kong’s divination scene. Nestled beside the bustle of the city’s main road, this street is lined with dozens of tarot card readers and fortune tellers of all kinds. While lesser known in the West, face reading is a popular esoteric practice in East Asia that assesses one’s fortune and character by analysing facial features—palm reading for faces. In face reading, every part of your face represents something. Even moles can indicate whether you’re a glutton or a lustful lover. A face is an interface to access the unseen future.

The art of facial interpretation flows in and out of mainstream culture, sometimes retreating into the obscured realm of the occult, but it never truly fades. Instead, face reading updates and scales up alongside imaging technologies. Image-making and facial interpretation are inextricably intertwined.

I.

Before there were drawings and cameras, there were faces. A face is an image in a movable frame, one of the oldest we study closely. The desire to decipher a face is ingrained in human culture. The oldest record of face reading dates back to the Paleo-Babylonian period in Mesopotamia and the practice is referenced in scholarly text as early as 8 BC.

In the West, the practice of face reading is known by another name: physiognomy (from the Greek physis, meaning “nature,” and gnomon, meaning “interpreter”). The seminal treatises on physiognomy linking facial features to individual character were attributed to Aristotle (though scholars now agree they were likely written by a pseudo-Aristotelian author), and Pythagoras reportedly required his students to undergo physiognomic evaluation before accepting them as mentees. Late mediaeval physiognomists referred to physiognomy as nomos, a “rule” of nature. Instead of moles, Czech scholars interpret the lines on foreheads (also known as metoposcopy). From the seventeenth to the eighteenth century, a wave of thinkers sought to elevate physiognomy to the status of a natural science. Physiognomists like Giambattista della Porta and Johann Kaspar Lavater proposed systematic frameworks in which each facial feature corresponded to particular traits of character. They proposed that each part of the face corresponded to particular moral or psychological attributes. Their followers took this further, categorising these features into subgroups with names like "cruel eyes" or "deceitful ears," creating a pseudo-scientific lexicon for reading personality from appearance.

The credibility of physiognomy, much like face reading, has fluctuated over time. However, unlike face reading (sometimes referred to as Chinese physiognomy), physiognomists took advantage of the advent of photography and woodcut printing, which provided an effective mechanical means to reproduce ‘exact’ likenesses. Mechanical reproduction bestowed the burgeoning pseudoscience an aura of scientific objectivity through (often self-invented and self-proclaimed) accurate anthropometric measurements. Our obsession with faces fueled the development of photographic techniques. The desire to capture faces transformed into a commercial incentive, giving rise to popular products like the “carte de visite” and daguerreotype, both of which pushed forward innovations in photographic technology, accelerating the adoption of the technology and improving the efficiency of image capturing and developing.

The development of physiognomy coincides with massive intellectual and technological advancement in Europe. On the one hand, the Industrial Revolution and urbanisation expanded village communities into sprawling city populations. The unprecedented social condition created a (play)space for thinkers to experiment with new methods of governance. This is the period when we saw the invention of modern statistics and census data collection. On the other hand, Enlightenment thinkers advocating for rationalism and empiricism ignited the desire to unveil the law of nature through the apprehension of different phenomena as signs (data) and their subsequent interpretation. A perfect storm of photography, statistics and physiognomy gave rise to a new wave of facial interpretation practices, this time scaled up. Photography allows faces to be captured en masse, while semi automated mechanical reproduction facilitates measurement.

The 19th and 20th centuries sees a massive influx of datasets and archives and the merge of scientific principles with photographic practices. This fusion is perhaps most apparent in statistician Francis Galton’s work, where he re-photographed multiple portraits on a single negative. Galton’s composite portraits—a visualisation of averages—were intended to represent a typical face of a certain ethnic race or economic class, which he later uses as proof for his racist pseudoscience Eugenics. French policeman Alphonse Bertillon developed an anthropometric catalogue, profiling suspects through meticulous measurements. Physiognomy provided the intellectual foundation for eugenics, social Darwinism, and biological determinism. The application of statistical principles to photography shifted the focus from individual photographs to extensive databases. As a result, “roughly between 1880 and 1910, the archive became the dominant institutional basis for photographic meaning.” By this time, faces encased in rectangular frames and stored alongside thousands of other portraits became the basis of modern facial interpretation.

II.

The motivations behind 21st century big data collection now extend beyond governance to commercial interests, particularly the capture of behavioural surplus. Social media platforms utilise statistics to optimise recommendation algorithms and predict user behaviour. The hunger for data intensifies as these platforms compete to anticipate user actions, valorising on the cognitive capital waiting to be extracted from datasets. Social media and smartphone photography facilitate the collection and capture of faces, providing the raw material for large-scale vision datasets like CelebA and Labelled Faces in the Wild (LFW), laying the foundation for the second major upscaling of facial interpretation.

Physiognomy has evolved into facial recognition with the advent of machine learning. The quest for facial recognition technology has existed since the early days of artificial intelligence in the 1960s, with pioneers like Woodrow Bledsoe aiming to develop computers capable of interpreting human faces, but it is machine learning that greatly improve the viability and computational efficiency of facial recognition. Machine learning enables neural networks to self-adjust their outputs by optimising parameters based on input data. In this context, historical photographic archives become crucial training data. Faces in images are reduced to segmentation maps and facial landmarks for computer readability, serving as standards to evaluate the performance of facial recognition algorithms. The dots and bounding boxes, often labelled by inexpensive human labour, become the ground truth from which computers extract meaning from faces. The insatiable demand for data in deep learning creates a market for specialised data-labeling platforms, such as Scale AI.

Today, instead of consulting a face reader, we turn to computers. We ask machines to predict emotions, ethnicity, gender, criminality or even sexuality. In retrospect, the meteoric rise of facial recognition comes as no surprise as it is a natural extension of physiognomy and eugenics, inheriting their logical fallacies and prejudices. Writers Luke Stark and Jevan Hutson aptly refer to facial recognition as “physiognomic AI”. It is the automation of the calliper, a racist pseudoscience repackaged in a shiny high-tech shell. The excess of portraits forms the fertile soil from which automated recognition grows. Against vehement critiques of racial bias and disregard of the complexity of identities, facial recognition technologies continue to rise. Scalability does not fret about complexity and nuance. It cares about growth and efficiency. It applies the same procedure to many indiscriminately through automation. As long as it is accurate enough, it is deployable, and when things go wrong, it is because there is not enough data.

III.

An autoencoder is a type of artificial neural network designed to create simplified representations (compressions) of complex data, effectively compressing complexity into a statistical latent space. It operates through two primary functions: the encoder, which analyses input data and extracts a set of key features encoded as numerical values, and the decoder, which uses these features to reconstruct the original input as closely as possible. The existence of facial recognition algorithms naturally suggests the possibility of facial reconstruction. History has proven this true, as DeepFake technology swiftly entered the public domain following the rise of facial recognition. Like photography before it, DeepFake and generative AI have triggered yet another massive wave of facial images, this time of persons that do not exist.

The surplus of this new genre of facial images won’t go to waste under the invisible hand of the market. Datagen, a synthetic data agency founded in 2018, was quick to adopt these faces in machine learning. According to their white paper, synthetic facial data offers multiple advantages over real data in training facial recognition algorithms. Firstly, computer-generated 3D faces are automatically labelled, reducing labour costs for data cleaning, which could potentially disrupt the dominance of data-labelling platforms like Scale AI. In other words, these faces come with their own "cheap" ground truth.

Secondly, synthetic fakes sidestep the consent and privacy concerns associated with large-scale vision datasets scraped from the internet. Because the people in these datasets don’t actually exist, Datagen claimed that it circumvents the ethical dilemmas surrounding surveillance and exploitation. Lastly, synthetic data offers full customisation. Datagen’s now-defunct website allowed users to control gender, ethnicity, facial expressions (like happy or angry), and even facial accessories (such as glasses). As far-fetched as this might sound, the tech scene has been convinced. According to Techcrunch, the company raised $50 million in 2022.