Designing Ethical Algorithms
Ethical algorithm design is becoming a hot topic as machine learning becomes more widespread. But how do you make an algorithm ethical? Here are 5 suggestions to consider.
By Claire Whittaker, Artificially Intelligent Claire
Data ethics are increasingly important as we look to scale applications of machine learning.
But how do you make an algorithm ethical?
What are the key levers you can pull within your algorithmic design that will make it ethical?
Here are 5 areas to consider when designing an ethical algorithm.
What will this algorithm be used for?
Spend some time thinking about the end use of your algorithm. Do you think that this application of the machine learning tool is ethical?
Increasingly we are seeking machine learning scientists and developers speak out when they do not agree with how their work is being used.
Notably, people working at Microsoft recently spoke out about the company signing a contract with the US military to make AR headsets. They say that they did not sign up to be war profiteers and do not agree with the move.
When working on machine learning projects with broad application, ask yourself if you think that the work is ethical. Are you happy with the potential for the machine learning algorithm when applied in the world?
This is the first step in ethical algorithm design.
Is your training data biased?
Another process to ensure ethical algorithm design is by looking at your training data set.
How representative of the world as a whole is your data set? Is this an ethical data set to make an algorithm on or to predict characteristics about?
Being mindful of the data set you are working with from an ethical perspective is one of the best ways to prevent some of the ethical challenges that have recently surfaced with machine learning applications.
When Amazon announced that it was halting development of its machine learning algorithm for recruitment it was because of an unexpected bias identified in the training set. Though Amazon had tried to remove gender indicators from the system to prevent bias, the algorithm was able to detect them.
The unconscious biases we all possess can be magnified by machine learning.
Being aware of how the data set used impacts ethical algorithm design allows you to mitigate biases.
Are you considering the different weightings you apply?
Linked to the data set you are using, when developing, it is also important to consider how you weight different features of an algorithm.
When optimizing weightings, you may want to consider the output and if you think that this is ethical by design or not.
How will you define an ethical feature set?
The final point to consider specifically on ethical algorithm design is the features you are using to develop a machine learning algorithm.
Is the range of features you are using to understand the data the best to give an ethical output?
On evaluating the features identified by the model as higher importance, do you agree that this is a good representation or do you see any potential red flags?
Again, being mindful of ethical algorithm design will ensure that you are not caught out by unintended consequences later in the process.
The last thing we want is a surprise in prod!
The best way to make ethical algorithms? Diversity in development
The final, and potentially most important component of ethical machine learning, is diversity.
Lack of diversity of all types within tech is a known issue.
This lack of a representative sample developing AI has caused some significant challenges in the past.
It’s not just in tech that this is a challenge either.
How many scandals due to tone-deaf advertising could have been prevented with diverse decision makers?
It only takes one voice to spot a potential ethical issue that others could be blind to.
A report by McKinsey into how to control for bias in machine learning suggested implementation of standards and controls.
Having a diverse board of evaluators before commiting a machine learning algorithm to production could ensure ethical development.
When developing machine learning technology, I urge you to seek out that voice or risk wishing you had.
Be ethical and inclusive in developing technology, or you may regret it.
With that point, I conclude this article. We all have the opportunity to develop technology that will be a real force for good in society.
Ensuring your algorithms are ethical by design will allow this good to be realized.
Bio: Claire Whittaker is on a mission to help everyone with understanding big data concepts and how to use artificial intelligence. It is her passion! So much so that she wants to pass this love on to you! She helps inquisitive millennials who love to learn about tech and artificial intelligence by blogging about learning to code and innovations in AI. You can read more about her experiences with ethics within artificial intelligence on her blog Artificially Intelligent Claire.
- Ethics + Data Science: opinion by DJ Patil, former US Chief Data Scientist
- The Algorithms Aren’t Biased, We Are
- A Concise Explanation of Learning Algorithms with the Mitchell Paradigm