Ethical AI

We're living in a world where machines and algorithms are increasingly giving recommendations, tagging content, generating reviews and even taking decisions. A number of ethical questions arise. Can we trust machines to do this right thing? Was my loan application processed in a fair manner? Why was my application rejected?

Then there are questions about AI replacing humans in the workplace. With widespread use of personal data, we're worried about privacy and data theft. What happens to us if AI agents acquire human-like cognitive abilities?

Ethical AI addresses these issues by building systems that are safe, fair, transparent, accountable and auditable. To practice ethical AI, it's been said that,

We are not at the point where you can simply download a tool to do the job. Data ethics is still new and it requires critical thinking.

Discussion

Are there real-world examples that highlight the need for ethical AI?
In March 2016, Microsoft released a chatbot named Tay. Tay could interact and learn from real users on social platforms. Soon Tay was exposed to nasty tweets and ended up becoming mean and racist on her own.
A study by ProPublica found that software used to predict future criminals is biased against blacks. Risk scoring done by the software is seen by judges and directly influences sentences. White defendants were often mislabelled as low risk compared to black defendants. Only 20% of those predicted to commit violent crimes went on to do so.
In March 2018, an Uber self-driving vehicle killed a pedestrian. Who is responsible: the distracted driver, the pedestrian, Uber, developers who wrote the code, or a sensor manufacturer? It's unrealistic to expect AI systems to be perfect but determining liability isn't trivial.
More recently, it was found that Facebook's ad delivery algorithms discriminate based on race and gender even when ads are targeted to a broad audience. Influencing factors include user profile, past behaviour and even the ad content.
What are the common ethical concerns raised by AI?
Ethical AI has the following concerns:
- Bias: AI systems can be biased because they're designed to look for patterns in data and favour those patterns.
- Liability: AI systems can't be perfect. When mistakes are made, who's responsible?
- Security: As AI systems advance, how do we stop bad actors from weaponizing them? What happens if robots can fight and drones can attack?
- Human Interaction: There's already a decline in person-to-person interactions. Are we sacrificing humanity's social aspect?
- Employment: Repetitive, predictable jobs that can be automated will be automated. Those replaced have to retrain themselves in areas where robots can't come in easily, such as, creative or critical thinking.
- Wealth Inequality: Companies rich enough to invest in AI will get richer by reducing cost and being more efficient.
- Power & Control: Big companies that use AI can control and manipulate how society thinks and acts.
- Robot Rights: If AI systems develop consciousness and emotions, should we give them rights? Can we punish AI and make them suffer?
- Singularity: What happens if AI surpasses human intelligence? Will they turn against us to defend themselves?
How can AI be biased when it's based on trusted data?
Bias in AI systems. Source: Google 2017.
Even if data is trusted, it may not be a fair representation of the population. Dataset may be skewed in a number of ways including race, gender, education, or wealth. When algorithms are exposed to this data, they acquire the bias that's present in them.
Bias in AI comes from human biases since we are the ones building the algorithms and creating/selecting data. Minorities, often in low-income groups, lack access to technology. As a result, AI systems are not trained on such data. One infamous example is Google tagging photo of a black woman as "gorilla".
There are many ways in which human biases creep into AI systems. Selection bias, interaction bias and latent bias are some examples.
AI system that's trained primarily on American or European faces will not work well when applied on Asian or African faces. In one example, a Nikon S630 camera hinted to the user that perhaps she has blinked when in reality Chinese people have eyes that are squinty when they smile. In another example, a happy tribal woman was wrongly classified to have "disgusted" emotion.
What can we do to make AI more ethical?
AI ethics by design. Source: Atos 2019.
We need to design our systems to avoid biases. Check that proxy metrics don't introduce bias. We can minimize bias by having diversity in teams, data, and validation.
For transparency, data must be marked with metadata detailing the source, intended use, usage rights, etc. Data must be managed well leading to algorithms that are better traceable, reproducible and fixable. Context is key to transparency and explainability.
When AI systems fail, have a plan to fix them in production. We need to monitor them. Let the purpose be clearly defined and bounded. AI systems should not be allowed to explore unintended pathways.
What we need is a holistic approach. It's not just about technology and tools, but also about leaderships, standards and rules. Microsoft's Satya Nadella has said ethical AI also means how humans interact with AI. We must have empathy, education, creativity, judgement and accountability.
Accenture has suggested setting up AI advisory bodies. Have discussions. Publish guidelines. Engage with stakeholders. The AI Ethics Lab is looking into integrating ethics and AI right at the R&D phase.
Are there published guidelines or code of conduct for practising ethical AI?
The Institute for Ethical AI & ML has identified 8 ML principles for responsible AI: human augmentation, bias evaluation, explainability by justification, reproducible operations, displacement strategy, practical accuracy, trust by privacy, and security risks. They also identify a 4-phase strategy: by principle, by process, by standards, by regulation. They've also formed the Ethical ML Network.
The Future of Life Institute has created a interactive map addressing validation, security, control, foundations, verification, ethics and governance.
In 2018, GE Healthcare published a set of AI principles. Google published its own guidelines of what AI applications should and shouldn't do. Microsoft released ten guidelines for developing conversational AI. One of these guidelines says that developers are accountable until bots become truly autonomous.

Milestones

2014

Amazon starts using AI for filtering resumes and hiring top talent. But in 2015, it's discovered that the algorithm is biased towards male candidates since it was trained on resumes submitted to Amazon over a 10-year period. This dataset had mostly male applicants since they dominated the tech industry. The algorithm just reinforced the bias. The algorithm is found to have other problems as well and is discontinued by Amazon.

2016

This is the year when people start talking about ethics and AI. Interest in ethical AI grows through 2017 and 2018.

May
2018

Microsoft appoints Tim O'Brien to the role of AI Ethicist, a new full-time position. The role involves AI ethics advocacy and evangelism. In reality, O'Brien clarifies that the role includes IoT, analytics, and AR/VR, since real-world solutions are hybrids of multiple technologies.

Dec
2018

Maria de Kleijn-Lloyd of Elsevier finds from her research that only 0.4% of published work on AI deals with ethics. She comments that there's lot of discourse on ethical AI but not much in terms of research and rigorous inquiry.

Apr
2019

Google disbands Advanced Technology External Advisory Council (ATEAC), which was supposed to advise on ethical AI. This happens due to controversy over ATEAC membership. This example shows that avoiding human bias and discrimination is not easy.

Jan
2020

Mapping different AI principles to eight themes. Source: Fjeld et al. 2020, pp. 8-9.

Fjeld et al. note that many organizations working with AI have published their own guidelines. They study thirty-six such documents and identify eight main themes. Moreover, recent documents tend to cover all the eight themes.