July 16, 2020

Is ethical AI possible? Keeping SkyNet fictional.

Mike Khanna

From HAL in 2001 to Skynet in Terminator, SciFi has always painted a dystopian vision of AI. This has led many people to question the ethics of AI. Here, we address an important question. Is ethical AI possible?

Artificial intelligence refers to any computer system that is able to exhibit intelligence. Typically, this intelligence is quite narrow in scope, allowing a computer to excel in one specific area. This includes technologies like facial recognition, demand forecasting, and anomaly detection. However, computer scientists are increasingly pushing towards an era of general artificial intelligence. Typically, this relies on deep learning using complex neural networks designed to replicate the way human brains work.

General artificial intelligence allows machines to learn for themselves and to replicate human-like intelligence. This was most clearly shown in 2015 when DeepMind’s AlphaGo defeated the European Go champion, Fan Hui. Go is often described as the Chinese equivalent of chess, and is notoriously hard to play well. The striking thing about AlphaGo was that the computer had taught itself how to play. First, it learned the rules of Go, then it played repeated games against itself, learning new strategies each time.

The potential of deep learning

Deep learning is one of the most promising fields in computer science research. In 2014, Google bought DeepMind for over $3 billion. Since then, deep learning has become a mainstream concept, driving many of the latest advances in AI. Deep learning promises to deliver extraordinary capabilities that may change the world we know. Here are two applications that raise issues regarding ethical AI.

Medical AI

AI promises us amazing advances in medicine. However, this raises the question of ethical AI. There are two areas where deep learning and AI are proving to be revolutionary.

Diagnostics. In January 2020, DeepMind published an article in Nature evaluating their AI system for analyzing mammograms. The article, titled “International evaluation of an AI system for breast cancer screening”, revealed their AI system outperforms human experts. Specifically, it reduces false positives between 1.2% and 5.7% (lower numbers UK, higher USA). It also reduces false positives by 2.7% to 9.4%. These are significant reductions for both these measures.

Drug discovery. Deep learning is revolutionizing the process of drug discovery. This has become especially important during the current pandemic. Several projects are using AI to test whether existing drugs might also work to treat COVID-19. Other projects are seeking to design new treatments from scratch.

Profiling and decision making

AI systems are increasingly used to make decisions that affect our lives. Many of these raise serious doubts about ethical AI.

Human resources. AI is sometimes used to make decisions regarding hiring and firing of staff. The idea is that AI systems are able to make unbiased decisions and just use the facts. However, this leads to some really perverse decisions that really raise doubts about ethical AI.  For instance, in 2018, a man who was wrongly fired by an AI. However, that same AI wouldn’t reverse its decision, even though it was a mistake.

Credit rating. Infamously, in 2019 Apple became embroiled in controversy regarding their credit card. They were using an AI algorithm to set credit limits for people. However, women were routinely being given lower credit limits thanks to an inbuilt bias in the system. Steve Wozniak, the co-founder of Apple, reported that his wife received 10x lower rating than he did, despite them sharing all their finances. This issue with bias is one of the biggest problems for ethical AI.

Trustworthy and ethical AI

As we have seen above, AI can have a profound impact on our lives. As a result, more and more people are asking whether ethical AI is possible. In turn, this is leading to an erosion of trust in AI.

Recently, the European Union tackled this problem head-on. They released a set of Ethics Guidelines for Trustworthy AI. These were written by the EU’s High-Level Expert Group on Artificial Intelligence.

The three tenets of ethical AI

The EU guidelines identify three essential requirements for any ethical AI.

  1. Lawful. The system must follow all applicable laws. Essentially, the AI must be subservient to the law.
  2. Ethical. The system must be ethical. More specifically, this means it must: Respect human autonomy; Prevent harm; Be fair; Be explainable.
  3. Robust. The system needs to be robust. This means it must: Be secure; Allow for results to be reproduced; Be thoroughly tested; Fail safely

These requirements address the issue of trust in AI by making it clear that humans should have ultimate control.

Guidelines for ethical AI

The document goes on to list seven specific guidelines for ethical AI. These are:

  1. Human agency and oversight. Ethical AIs must respect human rights. Users should be able to make informed autonomous decisions regarding AI systems. In turn, we must ensure a human is ultimately in charge. This could be through human-in-the-loop, human-on-the-loop, or human-in-command.
  2. Technical robustness and safety. We must design AI systems to be robust and safe. This means focusing on risk prevention, failing safely, and being resilient to attacks.
  3. Privacy and data governance. Globally, data privacy is increasingly important. The CCPA recently came into force. Brazil has the LGPD (Lei Geral de Proteção de Dados). And the EU’s General Data Protection Regulation (GDPR) is two years old. An ethical AI must respect personal data and mustn’t become a way to reduce privacy.
  4. Transparency. One of the biggest issues for AI is a lack of transparency. People have never trusted what they can’t see or explain. So, this leads to a drive for explainable AI. Without transparency and explainability, no one will trust in any AI.
  5. Diversity, non-discrimination, and fairness. As we mentioned earlier, bias is one of the biggest issues for ethical AI. Unfortunately, bias is inherent in many AI models. This is because they are trained on data that is biased. For instance, women receive lower credit scores because, in our society, women are typically lower-paid.
  6. Environmental and societal wellbeing. This guideline is a little more obscure. But in essence, any ethical AI must try to improve the wellbeing of humanity as a whole.
  7. Accountability. The final guideline is one of the most important. We must put in place systems that hold any AI to account. This is essential to ensure AIs are fair and trustworthy.

Ethical AI in practice

Science Fiction has always liked to paint dystopian visions of our future. Where AI is concerned, this has often led to the idea of machines rising up to overthrow humanity. However, Douglas Adams gave another vision of AI. Marvin the Paranoid Android has “a brain the size of a planet”. However, rather than use this to conquer humanity, he is beset by that most human of ailments, depression. The lesson here is we must focus on humanizing AI if we want to win the trust of users. Only then will we be able to answer the question, is ethical AI possible?

White Paper

SAIBRE AI Ecosystem

End-to-end AI applications that solve any business problem