Hello World

Hello World is a focused take on how algorithms permeate our world today. It doesn’t try to predict what’s ahead, instead taking an objective look on where algorithms are used, where they aren’t, the challenges they face and the ethical dilemmas they create. The book discusses six areas in which algorithms have been making an impact: data, justice, medicine, cars, crime and art.

I really enjoyed reading this book as it gave me an extensive overview on the impact made by algorithms today. It also triggered many insightful questions and perspectives on these algorithms and their future.

I picked out the most interesting conundrums and questions that I came across while reading the book and wrote them here. Hopefully they will inspire the same sense of wonder for you as the did for me.

Data

Targetted advertising

These algorithms could figure out if someone was pregnant through their spending habits. This allowed advertisers to target pregnant women, as they realised that the habits built during a pregnancy last. Unfortunately this has some consequences.

• What happens when the woman has a miscarriage? She would be constantly bombarded with reminders of your lost child.
• What if she’s just a teenager and is hiding her pregnancy from your parents? An algorithm would have successfully figured out that she’s pregnant even when her parents couldn’t.

Regulation of data usage

Thankfully, the law is catching up, with regulations like the GDPR in place now to prevent unauthorised usage of our data. But the author warns us not to let our guards down especially when using free algorithms — there’s no such thing as free lunch.

Justice

Balancing consistency and customisation

Human judges are extremely subjective. Research has shown that given the same case files, different judges came up with different sentences. They even contradicted their own decisions when given similar case files with basic details changed.

Using algorithms can make sentences consistent. Give out sentences too consistently however, and we lose some possibility of rehabilitation. When introducing algorithms, we need to be able to balance consistency and customisation.

Lack of feedback loop

Today, algorithms are used to calculate the risk of an individual re-offending. However, there is a major problem with these algorithms, in that they have no source of reliable feedback to improve upon. High-risk individuals would be denied bail, which means that the algorithm would never have a chance to see if these individuals were truly high-risk.

Biased algorithms

There is also the issue of bias. These algorithms are built upon information gathered from reality, which unfortunately, is biased. Individuals could be considered higher-risk simply because they belong to groups of people with higher crime rates.

We should instead design algorithms for the ideal world instead of furthering these biases.

Medicine

State of algorithms in medicine

When discussing algorithms in the context of medicine, we have to concede that there are some things provided by a human doctor an algorithm can never replace, like empathy. Other roles like pathology however, are something algorithms already excel in.

Today, a neural network can be fed MRI scans and produce diagnoses for breast cancer with a decent degree of accuracy. Interestingly, human pathologists typically make more false-negative mistakes, while the neural networks make more false-positive mistakes. It is by working together that human and algorithm can diagnose patients with astounding accuracy.

In another study, researchers were able to use the language ability of nuns (discovered through their writing) to predict whether they would get Alzheimer’s disease in the future.

Building high-level algorithms requires good data

Unfortunately, a major blocker in creating better medical algorithms is data. Medical data is a mess, often kept in physical records and not linked between hospitals. Services like 23andme are changing that. Despite the backlash and privacy issues from this, the medical data collected from these services could contribute to the development of algorithms that will save many lives in the future.

Who should we optimise medical algorithms for?

The author introduces an interesting conundrum here. Should a medical algorithm be built to serve an individual or the population? To put it in a different way, when making decisions on treatment, should the algorithm only use information about the individual, or consider the health of the population as a whole?

Cars

Constraints on today’s algorithms

Self-driving cars already exist today. However, many of them are limited in their area of operation. This is because there’s simply too much variation in terrain around the world that it’s nearly impossible to train a single model to fit all of them. One such study noted how machine learning algorithms used grass on the sides to detect roads, which made it unusable on desert terrain.

What we’ll probably end up with in a few years are vehicles that are autonomous but only up to a point as defined by the constraints we place on them. As Jack Stilgoe puts it, we will find ourselves in a situation where “things that look autonomous systems are actually systems in which the world is constrained to make them look autonomous.”.

Submissive machines?

Furthermore, an interesting contradiction comes in to play when we imagine the ideal autonomous vehicle. With safety being their biggest priority, they will avoid collisions wherever possible. Standing in front of an autonomous car would cause it to stop no matter what, and so would pulling out in front of one at the junction. Road dynamics would shift drastically into a system where cars are submissive, which could cause a whole host of issues like traffic holdups and annoying road-hogging pedestrians.

Detrimental improvements

Another problem with self-driving cars is that the more improvements we make to them, the less the human is required to actually drive. Over time, new generations of humans would not pick up the essential skill of driving. If we use humans as a failsafe as we try to improve cars, how can we expect them to rescue the car from a dangerous situation that it is unable to handle, if they haven’t driven in years?

Crime

Interference causing feedback loops

Unfortunately, augmenting crime-fighting capabilities with algorithms is not that easy. Cops-on-the-dots is a tactic where police run crime data through an algorithm to determine areas with high risk of crime, and deploy more patrols in those locations. The downside to this is that reducing the crime rates in those areas could risk a feedback loop, where new data contradicts old data making the whole algorithm produce inaccurate data.

Art

Can machines be creative?

There’s little challenge in terms of creating art with algorithms. Today, they can already generate paintings and pictures that look deceptively like real ones. However, the real question lies in whether humans would accept the pieces created by machines as real art.

If an algorithm and human somehow create the same painting, would they both count as art? If the human is then considered creative, can the machine be considered that as well?