Enumerating AI principles to be used as a guide or a questionnaire before launching an AI system or during the development process.


SwissCognitive Guest Blogger: Shivi Gupta – “AI Ethics and Unconscious Bias”


Data science has been the buzz word since almost a decade now. We still don’t fully understand how Instagram knows the reels that we will like or YouTube, the videos we will be watching. Their recommendation engines have reached a stage where it is almost asking us to be lazy and just watch what they have to offer. They are taking away the choice and we are so mesmerised by the tech that we are comfortable this way.

Now let us talk about these recommendation engines and how they are knowing us so well. One of the 5 stages of data science is data collection. Here a firm collects data from public or private sources and “cleans” it as per its business acumen or requirement. This process is also called data mining. If this process isn’t monitored for transparency, diversity, privacy and integrity, it could have serious implications for the system that is in the market or about to be launched.

The need of ethics in AI is the need of the hour. We can’t let systems or automated scripts dictate us without any accountability. An AI system which hasn’t gone through the ethics or principles can give birth to bias or reiterate the unconscious bias within the society.

The stereotypes that exist in all machine learning data are like a nurse has to be a female, homemaker has to be a woman, a driver is always a man and so on. To make sure the Ai system is not upsetting a group, religion or a sect with its responses, a system has to follow all these following principles:

  • Integrity – ability to register complaint, admit and rectify the mistakes made.
  • Diversity – data set should be diverse enough to break sexist/racist/discriminatory stereotypes.
  • Robustness – should be able to identify any unconscious bias in the user ‘s prompt as well as the output it is producing.
  • Accountability – if an output is harmful, the creator or the company must be accountable and responsible for the output and its consequences.
  • Fairness – how fairly an algorithm was used, was the data biased against a particular group or community?
  • Transparency – an AI system must be able to clarify how it predicted a particular set of output from the prompt
  • Explainable – an AI should be able to explain how an output was reached and what techniques or algorithm were used. This can be tricky since too many technical details wouldn’t make sense to everyone.
  • Data privacy – the data used, has not disobeyed any data privacy laws of any government or country.

To remember all the principles better, I came up with the mnemonic I_DRAFTED which will cover all the principles. This is just a reference guide, any organization can amalgamate a few of the principles together and create their own mnemonic. This can help Data scientists ask the right questions before or during the process of developing an AI system. Intelligent questions is what separates a smart phone or responsible adult from an ordinary one.

Thank you for reading this post, don't forget to subscribe to our AI NAVIGATOR!






About the Author:

Shivi GuptaShivi Gupta is a  passionate data scientist and full-stack developer, working in the industry for over a decade. An ardent Researcher and Innovator, who loves to automate things. Hobbies include travelling, eating, and playing sports (indoor and outdoor).