Taming an unborn Beast

Artificial Intelligence (AI) is pervasive and is permeating all jobs and all walks of life in every industry. It comes with huge power to automate mundane jobs and create new conveniences and opportunities but it comes with its limitations impacting lives when it is incorrect in its predictions. But the AI that we work with is constantly learning and adapting to recreate itself as it interacts with people and new situations.

AI is everywhere and it works by prediction. AI is trained with data we feed it. Then when we engage it to work, it makes a prediction. Every single time.

Ask AI for a financial forecast, whether we should expect an upcoming storm, or about how much power will be generated by wind for sustainability. It will readily give us the answer. All of them are predictions.

Critical answers from an AI such as whether someone has Covid or Alzimers is a prediction.

What appears as silly things with auto correct on our texts or search engines are all predictions.

Ask Google search 'On Thursday I will' and it will say I will get paid. Ask Google 'On Thursday you would' and it will say you will get pregnant. That is a prediction too.

AI is unleashing great power impacting lives and creating a new future for the world impacting everyone and everything.

AI is being used to create synthetic media or deep fakes. This distorts the idea of what is real news and creates fake news that propogates misinformation campaigns in the wrong hands. AI can create artificial human images and distort real videos such that we cannot see what is real or doctored. The challenge with AI is that there are positive and negative applications to AI. Listen to Manon De Dunne’s talk to understand the power and challenges of synthetic media using artificial intelligence below.

Taming the AI, making it work is like trying to work with the future that has not yet been created. It is about predicting the future before it is built.

To understand how we can manage AI we need to focus on three concepts that are core to how AI is trained, runs predictions and learns continuously.

The Confusion Matrix of AI

Everytime an AI engages with us it makes a prediction which can be true or false while the truth might be true or false. This leads to four possible situations about the prediction of the AI. The result predicted can be a ‘True Positive’, ‘True Negative’, ‘False Positive’ or ‘False Negative.’

The data scientist who builds AI tracks two metrics to track the AI confusion matrix so they can get visibility into how accurately the AI can make predictions. These are ‘Recall’ and 'Precision.’

Recall is how often the AI says Yes.

Trustworthy: When AI makes a decision, with the recall and precision metrix, we can know how often to rely on the prediction made by the AI. These metrics inform us about how much to trust the AI.

To paraphrase privacy expert Jessica Groopman, she says “We are taking AI’s advice as reference to make decisions but it is more of an inference that the AI makes with a certain level of accuracy with room for inaccurate predictions.”

Transparency: Whenever an AI engages with a patient, or customer or that critical system making life or death decisions, nobody knows how it made that decision, what factors did it consider in making that prediction.

Fairness: When the AI is built with a biased dataset that favors one group of people over the other, which is typically the case today, unfortunately, the AI gives more right answers to one group of people and more wrong answers to other groups.

For example, Amazon used a hiring algorithm that allowed for less women to be selected in the hiring process than men. As the Gender Shades project from Joy Buolamwini and Timnit Gebru showed, darker skin faces are less recognized than lighter skin faces in all facial recognition systems used in the US. AI predicts diseases correctly in the population it was trained with and makes inaccurate predictions for the population that were not well represented in its training data.

There are several kinds of bias that creeps into the AI at every stage of the AI lifecycle right from conception, data collection, labeling, model selection all the way to engagement with users. See image below from AI Ethicists Susanna Raj that shows the many possible types of bias that creeps into AI.

Image: Susanna Raj session about data bias at Career Pivot to AI Ethics class at Business School of AI

AI is constantly Learning. AI learns continuously and retrains itself to recreate itself with new data and new customers engaging with it. Each time an AI interacts with a user, it gets new training data from the interaction and this is used to re-train the AI to make a better version of itself.

This means that the AI that you will interact with tomorrow is not the same as the AI that you are interacting with today. This is the power of AI to adapt. Therein lies the challenge of managing the AI, taming the beast that is not yet born.

We need to tame the AI to make it work for everyone without bias. We need to improve AI from making incorrect interpretations that affect lives and take away agency from any group of people.

As much as AI is recreating itself by retraining itself, AI is creating a new world every time you interact with it. It is predicting its results based on who asks to open the door to be let into a building because of the bias in its training data. This makes the AI give unpredictable results in the future.

The good news is that we can make AI make a better world by focusing on the training data that trains the AI. This training data is what sets the precision and recall to create the confusion matrix. This training data is what contributes to the bias in the AI to act unfairly towards some groups of people. This training data is what allows the AI to retrain to become the AI that you will interact with in the future.

So the best way forward is to focus on the training data and ensure AI ethical principles of trust, transparency and fairness are incorporated into the data and incorporate active bias mitigation at every stage of the AI development lifecycle.

Listen to AI ethicist Diya Wynn from AWS’s talk about how to ensure that your customers build AI ethically in all industries.

Sudha Jamthe is a Technology futurist who teaches professional learners to create the connected, autonomous and ethical future with Artificial Intelligence at Stanford Continuing Studies and Business school of AI.

Events	Want Ads
Our Sponsors	Resources