This is the fourth and last episode of mini series "The dark side of AI".
I am your host Francesco and I’m with Chiara Tonini from London. The title of today’s episode is Bias in the machine

C: Francesco, today we are starting with an infuriating discussion. Are you ready to be angry?

F: yeah sure is this about brexit?
No, I don’t talk about that. In 1986 the New York City’s Rockefeller University conducted a study on breast and uterine cancers and their link to obesity. Like in all clinical trials up to that point, the subjects of the study were all men.
So Francesco, do you see a problem with this approach?

F: No problem at all, as long as those men had a perfectly healthy uterus.
In medicine, up to the end of the 20th century, medical studies and clinical trials were conducted on men, medicine dosage and therapy calculated on men (white men). The female body has historically been considered an exception, or variation, from a male body.

F: Like Eve coming from Adam’s rib. I thought we were past that...
When the female body has been under analysis, the focus was on the difference between it and the male body, the so-called “bikini approach”: the reproductive organs are different, therefore we study those, and those only. For a long time medicine assumed this was the only difference.

Oh good ...
This has led to a hugely harmful fallout across society. Because women had reproductive organs, they should reproduce, and all else about them was deemed uninteresting. Still today, they consider a woman without children somehow to have betrayed her biological destiny. This somehow does not apply to a man without children, who also has reproductive organs.

F: so this is an example of a very specific type of bias in medicine, regarding clinical trials and medical studies, that is not only harmful for the purposes of these studies, but has ripple effects in all of society
Only in the 2010 a serious conversation has started about the damage caused by not including women in clinical trials. There are many many examples (which we list in the references for this episode).

Give me one
Researchers consider cardiovascular disease a male disease - they even call it “the widower”. They conduct studies on male samples. But it turns out, the symptoms of a heart attack, especially the ones leading up to one, are different in women. This led to doctors not recognising or dismissing the early symptoms in women.

F: I was reading that women are also subject to chronic pain much more than men: for example migraines, and pain related to endometriosis. But there is extensive evidence now of doctors dismissing women’s pain, as either imaginary, or “inevitable”, like it is a normal state of being and does not need a cure at all.

The failure of the medical community as a whole to recognise this obvious bias up to the 21st century is an example of how insidious the problem of bias is.

There are 3 fundamental types of bias:

One: Stochastic drift: you train your model on a dataset, and you validate the model on a split of the training set. When you apply your model out in the world, you systematically add bias in the predictions due to the training data being too specific
Two: The bias in the model, introduced by your choice of the parameters of your model.
Three: The bias in your training sample: people put training samples together, and people have culture, experience, and prejudice. As we will see today, this is the most dangerous and subtle bias. Today we’ll talk about this bias.

Bias is a warping of our understanding of reality. We see reality through the lens of our experience and our culture. The origin of bias can date back to traditions going back centuries, and is so ingrained in our way of thinking, that we don’t even see it anymore.

F: And let me add, when it comes to machine learning, we see reality through the lens of data. Bias is everywhere, and we could spend hours and hours talking about it. It’s complicated.

It’s about to become more complicated.

F: of course, if I know you…
Let’s throw artificial intelligence in the mix.

F: You know, there was a happier time when this sentence didn’t fill me with a sense of dread...
ImageNet is an online database of over 14 million photos, compiled more than a decade ago at Stanford University. They used it to train machine learning algorithms for image recognition and computer vision, and played an important role in the rise of deep learning. We’ve all played with it, right? The cats and dogs classifier when learning Tensorflow? (I am a dog by the way. )

F: ImageNet has been a critical asset for computer-vision research. There was an annual international competition to create algorithms that could most accurately label subsets of images.
In 2012, a team from the University of Toronto used a Convolutional Neural Network to handily win the top prize. That moment is widely considered a turning point in the development of contemporary AI. The final year of the ImageNet competition was 2017, and accuracy in classifying objects in the limited subset had risen from 71% to 97%. But that subset did not include the “Person” category, where the accuracy was much lower...
ImageNet contained photos of thousands of people, with labels. This included straightforward tags like “teacher,” “dancer” and “plumber”, as well as highly charged labels like “failure, loser” and “slut, slovenly woman, trollop.”

F: Uh Oh.
Then “ImageNet Roulette” was created, by an artist called Trevor Paglen and a Microsoft researcher named Kate Crawford. It was a digital art project, where you could upload your photo and let the classifier identify you, based on the labels of the database. Imagine how well that went.

F: I bet it did’t work
Of course it didn’t work. Random people were classified as “orphans” or “non-smoker” or “alcoholic”. Somebody with glasses was a “nerd”. Tabong Kima, a 24-year old African American, was classified as “offender” and “wrongdoer”.

F: and there it is.
Quote from Trevor Paglen: “We want to show how layers of bias and racism and misogyny move from one system to the next. The point is to let people see the work that is being done behind the scenes, to see how we are being processed and categorized all the time.”

F: The ImageNet labels were applied by thousands of unknown people, most likely in the United States, hired by the team from Stanford, and working through the crowdsourcing service Amazon Mechanical Turk. They earned pennies for each photo they labeled, churning through hundreds of labels an hour. The labels were not verified in any way : if a labeler thought someone looks “shady”, this label is just a result of their prejudice, but has no basis in reality.
As they did, biases were baked into the database. Paglen quote again: “The way we classify images is a product of our worldview,” he said. “Any kind of classification system is always going to reflect the values of the person doing the classifying.” They defined what a “loser” looked like. And a “slut.” And a “wrongdoer.”

F: The labels originally came from another sprawling collection of data called WordNet, a kind of conceptual dictionary for machines built by researchers at Princeton University in the 1980s. But with these inflammatory labels included, the Stanford researchers may not have realized what they were doing.
What is happening here is the transferring of bias from one system to the next.

Tech jobs, in past decades but still today, predominantly go to white males from a narrow social class. Inevitably, they imprint the technology with their worldview. So their algorithms learn that a person of color is a criminal, and a woman with a certain look is a slut.

I’m not saying they do it on purpose, but the lack of diversity in the tech industry translates into a narrower world view, which has real consequences in the quality of AI systems.

F: Diversity in tech teams is often framed as an equality issue (which of course it is), but there are enormous advantages in it: it allows to create that cognitive diversity that will reflect into superior products or services.
I believe this is an ongoing problem. In recent months, researchers have shown that face-recognition services from companies like Amazon, Microsoft and IBM can be biased against women and people of color.

Crawford and Paglen argue this:
“In many narratives around AI it is assumed that ongoing technical improvements will resolve all problems and limitations.
But what if the opposite is true? What if the challenge of getting computers to “describe what they see” will always be a problem? The automated interpretation of images is an inherently social and political project, rather than a purely technical one. Understanding the politics within AI systems matters more than ever, as they are quickly moving into the architecture of social institutions: deciding whom to interview for a job, which students are paying attention in class, which suspects to arrest, and much else.”

F: You are using the words “interpretation of images” here, as opposed to “description” or “classification”. Certain images depict something concrete, with an objective reality. Like an apple. But other images… not so much?

ImageNet contain images only corresponding to nouns (not verbs for example). Noun categories such as “apple” are well defined.
But not all nouns are created equal. Linguist George Lakoff points out that the concept of an “apple” is more nouny than the concept of “light”, which in turn is more nouny than a concept such as “health.”
Nouns occupy various places on an axis from concrete to abstract, and from descriptive to judgmental. The images corresponding to these nouns become more and more ambiguous.
These gradients have been erased in the logic of ImageNet. Everything is flattened out and pinned to a label.
The results can be problematic, illogical, and cruel, especially when it comes to labels applied to people.

F: so when an image is interpreted as Drug Addict, Crazy, Hypocrite, Spinster, Schizophrenic, Mulatto, Red Neck… this is not an objective description of reality, it’s somebody’s worldview coming to the surface.
The selection of images for these categories skews the meaning in ways that are gendered, racialized, ableist, and ageist. ImageNet is an object lesson in what happens when people are categorized like objects.
And this practice has only become more common in recent years, often inside the big AI companies, where there is no way for outsiders to see how images are being ordered and classified.

The bizarre thing about these systems is that they remind of early 20th century criminologists like Lombroso, or phrenologists (including Nazi scientists), and physiognomy in general. This was a discipline founded on the assumption that there is a relationship between an image of a person and the character of that person. If you are a murderer, or a Jew, the shape of your head for instance will tell.

F: In reaction to these ideas, Rene’ Magritte produced that famous painting of the pipe with the tag “This is not a pipe”.

You know that famous photograph of the soldier kissing the nurse at the end of the second world war? The nurse came public about it when she was like 90 years old, and told how this total stranger in the street had grabbed her and kissed her. This is a picture of sexual harassment. And knowing that, it does not seem romantic anymore.

F: not romantic at all indeed

Images do not describe themselves. This is a feature that artists have explored for centuries. We see those images differently when we see how they’re labeled. The correspondence between image, label, and referent is fluid. What’s more, those relations can change over time as the cultural context of an image shifts, and can mean different things depending on who looks, and where they are located. Images are open to interpretation and reinterpretation. Entire subfields of philosophy, art history, and media theory are dedicated to teasing out all the nuances of the unstable relationship between images and meanings.
The common mythos of AI and the data it draws on, is that they are objectively and scientifically classifying the world. But it’s not true, everywhere there is politics, ideology, prejudices, and all of the subjective stuff of history.

F: When we survey the most widely used training sets, we find that this is the rule rather than the exception.
Training sets are the foundation on which contemporary machine-learning systems are built. They are central to how AI systems recognize and interpret the world.
By looking at the construction of these training sets and their underlying structures, we discover many unquestioned assumptions that are shaky and skewed. These assumptions inform the way AI systems work—and fail—to this day.
And the impenetrability of the algorithms, the impossibility of reconstructing the decision-making of a NN, hides the bias further away from scrutiny. When an algorithm is a black box and you can’t look inside, you have no way of analysing its bias.

And the skewness and bias of these algorithms have real effects in society, the more you use AI in the judicial system, in medicine, the job market, in security systems based on facial recognition, the list goes on and on.

Last year Google unveiled BERT (Bidirectional Encoder Representations from Transformers). It’s an AI system that learns to talk: it’s a Natural Language Processing engine to generate written (or spoken) language.

F: we have an episode in which we explain all that

They trained it from lots and lots of digitized information, as varied as old books, Wikipedia entries and news articles. They baked decades and even centuries of biases — along with a few new ones — into all that material. So for instance BERT is extremely sexist: it associates with male almost all professions and positive attributes (except for “mom”).

BERT is widely used in industry and academia. For example it can interpret news headlines automatically. Even Google’s search engine use it.

Try googling “CEO”, and you get out a gallery of images of old white men.

F: such a pervasive and flawed AI system can propagate inequality at scale. And it’s super dangerous because it’s subtle. Especially in industry, query results will not be tested and examined for bias. AI is a black box and researchers take results at face value.

There are many cases of algorithm-based discrimination in the job market. Targeting candidates for tech jobs for instance, may be done by algorithms that will not recognise women as potential candidates. Therefore, they will not be exposed to as many job ads as men. Or, automated HR systems will rank them lower (for the same CV) and screen them out.

In the US, algorithms are used to calculate bail. The majority of the prison population in the US is composed of people of colour, as a result of a systemic bias that goes back centuries. An algorithm learns that a person of colour is more likely to commit a crime, is more likely to not be able to afford bail, is more likely to violate parole. Therefore, people of colour will receive harsher punishments for the same crime. This amplifies this inequality at scale.

Conclusion

Question everything, never take predictions of your models at face value. Always question how your training samples have been put together, who put them together, when and in what context. Always remember that your model produces an interpretation of reality, not a faithful depiction.
Treat reality responsibly.

Dolly Parton's America (with Jad Abumrad)