Citizen-Centered Best practices for AI/ML/DL

It is reassuring that many government agencies today operate on data-driven decisions, relying on numerous analytical systems and smart algorithms that assist in those decisions. These data-driven decisions vary from budgeting for bread and butter activities, or unexpected emergencies or distribution logistics, or financial risk detection such as anti-money laundering (AML) or fraud waste and abuse (FWA). 

With the influx of rule-based expert systems, data-driven applications rose a long time ago, and today it sits well into the AI realm. As we propel into the next generation of AI, the current generation of data-driven systems shall expand and employ neural networks and deep learning concepts for decision making. However, when implementing AI-driven decision-making systems, Government agencies must act carefully and thoughtfully because choices made by such intricate AI systems shall influence the decisions that have the potential to disrupt our daily lives. When harvesting choices from AI-driven systems, it is critical that every people-serving organization recognizes citizen-centric principles when implementing them.

What is a Citizen-Centered AI? The citizen-centered AI provides best practices that foster an inclusive AI ecosystem that serves citizens and communities. It eliminates bias in the system. 

Why do we need citizen-centric best practices? For example, agency leadership must understand AI bias. The government must ensure its citizens that decisions made by machine learning and neural networks are unbiased. A machine learning or deep learning network without citizen-centric principles could potentially lead to false conclusions, resolutions, and outcomes, and in the end, hurt the citizens. The fundamental principles surrounding a citizen-centric design is to embrace an AI ecosystem that drives more reliable business decisions while at the same time eliminating biases and skewed outcomes.

What is AI-bias? Our human cognitive thinking and reasoning operate on data that we pull from our worldly struggles, choices, pleasure, and other experiences. Our brain continually receives various types of data, and the data determines or influences our predilections, expectations, beliefs, and judgments. Bad data yields bad perceptions and biased opinions. Likewise, if AI training data constitutes implicit racial, gender, or ideological biases, the outcome ends up biased. 

For example, when we designed an AI recommendation engine to find the most suitable candidate for a position, our data science team created various models using data from past resumes and other HR resources. Now, when searched to find the most suitable network engineer, the engine recommended mostly males for the role. When it comes to network jobs, it is a male-dominated domain. We trained the models using training data generated from historical records that reflected the domination; it was biassed, and it influenced the AI engine to assume men are more suited for the networking positions, which is blatantly false, misleading. The bottom line - biased data exhibit skewed outcomes, and if we fail to address the data bias, something as simple as a job search could make social or economic disparities worse. Hence, Government agencies must foster an AI ecosystem that is fundamentally citizen-centered, that is inclusive, and fully eliminating situations where knowingly or unknowingly some race or some communities benefit. 

The following are fundamental best practices for citizen-centric best practices for government agencies to consider before any AI implementation:

1. Explainability/Explainable-AI: The realm of Explainable-AI has two facets to consider. a) The Artificial Intelligence realm is still a black box, and machine learning or deep learning applications rely on training data sets. When an outcome transpires - we cannot precisely assess the rationale for such a decision. The data sets might have hidden bias factors. Besides, Pinpointing the training data set(s) that led to the decision is complicated or not possible in many cases. AI application realm requires abilities to track the decision-making process and the rationale behind such decisions or predictions. Explainable-AI provides AI & ML systems the ability to provide the rationale behind its decisions, and understanding their future behavior and communicate it to the users. b) Employees within the agency and external stakeholders, and the citizen-consumers should have information on how any the AI system arrives at its contextual decisions. The focus here is beyond the explanation of AI system conclusions. As explained above, a detail explanation of decisions made by deep learning systems is not possible. However, agencies must disclose information about the algorithms and data sets employed and how human decision-makers use the algorithm's conclusion - Being Transparent.
2. Establish systematic methods for Unlearning: Our brain provides cognitive abilities to learn and unlearn - temporarily or permanently. Unlearning is the capability to choose an alternative mental model or paradigm. When we learn, we add knowledge to what we already know, but when we unlearn, we step outside from the mental model that we developed to choose a different model. For example, assume you are an established opera artist, and then you shift to pop music, the vocal and verbal nuances are totally different in these genres. It would take some time for the opera singer to unlearn the nuances of opera to master the pop genre. Until then, you are going to find the opera bias seeping into the artist's pop renditions. Our brain does have a system to unlearn, but it has to be a systematic effort to unlearn with facts. Why do we need it? We need unlearning with facts because it helps us let go of our false beliefs, bias, and assumptions that formerly governed us. In the AI context, unlearning is an ability to unlearn specific knowledge or data, to protect against biases in data sets. Government agencies must ensure that AI-systems have systematic processes and tools for unlearning. For example, by creating an unlearning training data set to offset bias. Such methods require cross-functional expertise and support. 
3. Create Data Stewards and Data councils: To reduce the prospect of inadequately trained AI networks, government organizations must nominate data stewards and promote data councils. Data Stewards and councils establish and ensure the data sets are appropriate and match the business case. Besides, they also ensure the training data is adequately diverse, delivering optimal outputs, eliminating bias. Bad data can pose risks for human workers and favor human biases. Proactively data council apparatus helps Government agencies to leverage additional human oversight and catches errors or flaws in data sets early in the process. 
4. Smart Data Fencing employing digital methods and processes: Explicitly build a smart fence around data features or data points when considering datasets for AI networks. For example, AI decision systems should not have data that produces race bias and control any data that can foster such issues. Implementing a process-based fence and a digital fence provides a shield for what AI systems can do or cannot do with the data it generates or acquires. Besides, the rules governing the fence must be verified continuously or audited by data stewards councils and external governing agencies. Unlearning systems help expand the data bounds. 
5. Methodical Data Monitoring: Establish human processes, automated tools, and technologies for methodical monitoring of test data sets for biases. Create trust level buckets to categorize data sets, for example, Trusted data, the potential to be trusted, and unreliable. Pre-board the data sets in these bucks and tag them ahead of time. Agencies must assess manually and mechanically the trust level buckets and try expanding trusted data. By this, we allow the AI systems to improve and get smarter while monitoring and protecting against inaccurate data.
6. finally, Being Transparent: Government Agencies implementing AI shall make high-level conceptual and implementation details of every AI project accessible to everyone. Furnish the associated details on the machine learning and deep-learning process, and the types of data set used in the algorithms and outline how machine learning/deep learning decisions can improve or affect or impact society. Practice caution when it comes to the Data sets that need privacy protection. Create external and internal groups consisting of management, security, and policy stakeholders with appropriate clearance to monitor compliance.

To conclude, AI has the potential to change our life for good. Government agencies must continue protecting their citizens from data biases by thoughtfully managing the data that power machine learning and artificial neural networks. Agencies should adopt citizen-centered principals and best practices. Such adoption leads us toward optimally trained AI algorithms, and unbiased foresight that makes the playground equal for everyone. I firmly believe that technology equalizes access for everyone. Unbiased data is essential for implementing successful AI programs to avoid significant risks in the future. 

Good luck.