Ethical AI development – what does it mean?
Discover what Ethical AI development is, the concept behind it and the principles it follows.
As Artificial Intelligence tools in general keep gaining popularity – think of the wildly popular ChatGPT (or, in a colleague’s words, The Revolutionary Bullshit Parrot), Midjourney, Bard and so on – AI has essentially gone… mainstream. Scientists have known and researched the topic for years, but opening it up to consumers for everyday use has pushed its development further still. It has boosted demand for every kind of AI solution even more, and by now most big companies have already bragged about adding AI to their products/services. AI has grown into an essential element of our ever more digital world, from running voice assistants to supporting sophisticated medical diagnostics. Yet as AI systems spread everywhere, worries about their ethical consequences have surfaced too. You have surely heard people saying that AI development leads nowhere but to the destruction of humanity. Sounds like a Black Mirror episode, doesn’t it?
Today I won’t speculate on where AI development will lead us. Let us look instead at a somewhat less apocalyptic vision, but one that is still important and can cause problems.
What is Ethical AI?
Even though the breakthroughs and value of Artificial Intelligence are evident, many AI-based systems raise doubts about the way they are designed and used. Of course, it was easy to foresee that AI would also become a tool for people with bad intentions. This ranges from spreading stereotypes or leaking sensitive data, through spreading disinformation and using AI for propaganda and manipulation, to deepening socio-economic inequalities and even building powerful autonomous weapons with AI. And that’s before mentioning the hard-to-foresee effects of AI’s rapid growth, including the possible emergence of Artificial Intelligence able to take over the world and wipe out humanity.
Yes, yes, I know the Daleks weren’t Artificial Intelligence, but whenever someone talks about extermination, they’re the first thing that comes to my mind ;)
Problems like these spark rather heated discussions about how to prevent such scenarios. In the end, the conversation frequently turns not only to how the AI tools are built, but to the doubtful quality of the data they learn from. Companies and all sorts of organizations have ever more data to extract valuable insights from, as the demand for automation and data-driven decisions keeps rising. And that is, naturally, only the start of AI use cases. However, as AI solutions are used more and more, the unexpected effects of flawed assumptions – stemming either from the data or from how the tool is configured – have resulted in ethical AI guidelines being drawn up.
Ethical AI is about building, deploying and using Artificial Intelligence systems in line with ethical principles and values, in a way that supports the well-being of individuals and of society at large. Such systems are meant to be transparent, accountable and equitable, putting the protection of human rights and dignity first. Ethical AI follows a set of values and rules, including, among others:
- Fairness: making sure AI systems are built and deployed so that they treat every individual and group reasonably and discriminate against nobody. Fairness in AI is, however, a complicated, context-dependent notion, and no single universal definition covers every AI application. Full fairness is hard to reach, but the goal is to identify and reduce fairness-related harms as far as possible.
- Transparency and explainability: building AI systems in a transparent, explainable way, so that users know which data is used, how decisions are reached and how those decisions affect them.
- Accountability: making sure responsibility for the decisions taken by AI systems is clear – the creators and anyone else responsible can be easily identified, and there are mechanisms that prevent and deal with any harm or negative effects that might occur.
- Privacy: AI systems should be built to guarantee individuals’ privacy and to protect personal data and information over the whole life cycle of the system.
- Safety and reliability: making sure the system is safe, meaning it works reliably and as intended and poses no risk to individuals or even to society.
- Contestability: since they can significantly affect the user, a particular group or even society as a whole, people should be able to dispute doubtful outcomes.
Promoting ethical AI therefore helps ensure the technology is used in line with commonly accepted values and for the common good of society. AI is, after all, meant to serve us, support us, make tasks easier, take over the ones that would take us too much time, and so on.
A central element of ethical AI is inclusiveness in AI datasets, that is, making sure the data used to train and build AI systems is varied and represents different populations. Such inclusiveness matters especially to keep AI systems from reinforcing or worsening existing social inequalities or biases.
Responsible AI
The two ideas are closely related. Generally, though, ethics is mostly concerned with the good of individuals, groups and society. Responsible AI covers putting ethical principles for AI into practice, thus broadening the topic of Ethical AI. Its goal is to make sure AI is built and used with the possible effects on different stakeholders in mind – users, employees and the wider community.
So Ethical AI forms a vital part of responsible AI. Responsible AI, however, reaches beyond ethics and adds a broader set of principles and practices meant to ensure that AI is built and used in a responsible and sustainable manner.
source: Reddit
Best Practices for Developing Ethical AI
Legally speaking, the future of AI development is still not settled, as it is a fairly new technology trend. Some steps have nonetheless been taken already, for instance by the Parliament of the European Union, which is preparing the AI Act that defines, among other things, prohibited practices and the obligations of providers and users depending on how much risk artificial intelligence may pose. In addition, in November 2021 UNESCO released its Recommendation on the Ethics of Artificial Intelligence, which (in theory) considers the protection of human rights, the environment, minorities, health and many other areas. It is a collection of recommendations, mostly about policy, meant as guidance for how AI should be developed and handled. Its focus is on promoting core values so that systems are built to benefit society, the environment and humanity as a whole.
As for more technical practices for engineers, companies such as Microsoft, IBM or Google publish concrete recommendations for building ethical AI systems on their websites. They offer useful advice; here are a few examples:
Human-centered design approach
This means that the user experience is key to evaluating how AI systems affect people. Developers should therefore build features with clear disclosures and give users control. It is also essential to provide one or several options depending on the context, to expect negative feedback and to iterate while testing. To get diverse and unbiased results, system creators should involve a varied group of users and use their feedback to improve outcomes.
Multiple metrics to assess training and monitoring
To fully grasp the trade-offs and experiences, it’s advisable to rely on several metrics instead of just one. Different metrics, such as user surveys or overall system performance metrics, can be useful. It’s essential to examine false positive and false negative rates across subgroups and to make sure the chosen metrics fit the system’s context and goals.
Examining raw data
Looking at raw data, whenever that’s possible, helps you understand it better and ensures accuracy and fairness. Whatever data is fed into the system directly shapes its results. That’s why it’s essential to look for errors in the data, make sure it is representative, and remove bias for training as well as unnecessary features.
Testing
AI systems also benefit from following software engineering best practices for testing and quality assurance. This means running thorough unit tests, actively monitoring data and concept drift, relying on a dependable test dataset, including iterative user testing, building quality checks into the system and running integration tests to grasp how parts of the system interact and where feedback loops might arise.
Collecting and handling data responsibly
To begin with, AI engineers must establish whether an ML model can be trained without sensitive data, or using publicly available datasets. If sensitive training data cannot be avoided, it should be used as little as possible and treated with the greatest care, in line with legal requirements and standards. Following best practices is strongly recommended, for example encrypting data both in transit and at rest and complying with privacy principles (naturally, Google recommends its own).
Safeguarding the privacy of ML models
ML models’ may expose the data they were trained on. This can happen via internal parameters or through observable behaviour of the model. To check whether sensitive data gets memorised or exposed unintentionally, specialists use dedicated tests and quantitative research that examine how machine learning models leak information about the individual data records they were trained on. Further good practices include optimising models by trying out data minimisation approaches, mathematically guaranteed techniques and procedures borrowed from cryptographic software development.
Identifying potential threats to the system
As you’ve probably noticed by now, proper planning of an AI system’s architecture includes security. To do it well, you also have to consider how the system might be misused, how vulnerabilities might be exploited (and, most importantly, how to prevent that), and what the consequences of such actions would be. When evaluating potential threats, their likelihood and the extent of possible damage, it can help to create a risk model that checks the assets and analyses every exposure and hazard.
Final thoughts
Artificial Intelligence brings enormous benefits, but the danger of it being used for harmful ends is nearly as great. That explains the rising interest in, and continuing work on, dedicated frameworks for those who build and use such tools.