How to minimize Bias?

With ChatGPT and various other AI agents, systems, apps and tools, AI is all around us. ChatGPT is a well known tool, but there are many other AI-based systems, used for prediction in the consumer world, security assistance with face recognition, autonomous vehicles and so on. It is really all around us, invisible in most cases, but affecting our lives non the less.

We, mostly as consumers, but also as developers, should be aware that these systems could be biased (and many are), but that bias could be minimized if addressed correctly.

How you may ask?

Defining Bias

To minimize something, let's start by defining it - what is a "Bias"?

Bias is defined as:

1. Inclination or prejudice for or against one person or group, especially in a way considered to be unfair.
2. [Statistics] a systematic distortion of a statistical result due to a factor not allowed for in its derivation.

(Google, via Oxford Languages)

That inclination or distortion is what we are looking for. Will we be able to identify it? If we can't identify a problem, it won't be solved.

Examples

Let's look at a few short examples and see if we can identify the bias in them.

Example 1: A system that offers you shopping items.

For online shopping, a predictive system could save time by offering you what you need. This predictive system offered everyone to buy baby soothers and dummies. Is it biased?

Of course! Offering everyone baby products causes the system to be biased in favor of one group.

Example 2: Detect vehicles in the CBD for pollution control

The mayor wants to reduce the pollution in the highly populated CBD and uses the existing security cameras to detect the vehicles in that area. The system focuses on family cars, made in the last 10 years, since they are more common now. Is it biased?

Of course! It is obvious that trucks and older vehicles will be missed. And worse, the ones that are missed contribute a lot to the overall pollution.

In both cases, the bias is glaring, evident and extreme.

Why was it so easy to identify it straight away?

Key in identifying Bias - setting the Expectations

The answer is simple - because we knew the expected outcome! We knew that not all shoppers needed baby products and that not all cars were new, so in both cases the systems' outputs seemed completely incorrect.

When we develop a system, we know the project requirements, some are more evident, some are naturally assumed. With AI-based systems, expected output is of course the "ground truth" to which we compare models' outputs and measure their accuracies. When missing, how can we discuss bias or accuracy?

Similarly, as consumers, what are the expected outcomes from the systems around us? In the two examples above, the predictive system was meant to offer me products I'd want to buy, and with the second - lower the pollution. How they did it was a completely different matter, but those were the objectives of the systems. Without clear expectations, we might be unaware of biases, and it might be a long time before we are affected. To minimize bias, set your expectation, based on who you are and what your values are, and compare systems' outputs to those expectations. For example, News should mostly be about undistorted facts, otherwise they are "opinions".

Minimize the Bias

To summarize - minimizing bias is possible only if you:

• Are aware it might be there. Why would you look for it otherwise?
• Set expectations, i.e. the ground truth. Where does it come from?
• Actively correct it, either as a developer or as a consumer. How?

Below are a few suggestions addressing each step from the list above that can help minimizing Bias.

1. AwarenessAny statistical analysis could be faulty, any system could be faulty, any system could be compromised. If you accept it, you'll be more careful, double check your own work and be aware of the potential bias around you.
2. Define the expected outcomeWhether you are offered candidates by an AI-powered HR assistant, read about a conflict on the other side of the world, or build a great vision-based application - set your expectations. Form an opinion based on quality data from different sources and verify your own ground truth labels. Don't be afraid to question someone else's annotation quality, agenda or source of data.
3. Actively prevent itAs a developer you can never be too careful - testing the data(!) before testing the model, applying sanity checks along the way, and once released to production, collecting feedback and verifying the system performs as required by the business. As a consumer, you don't have access to the source code or the PRs, but you can report about any detected bias. If you don't (re-read example 2 above before continuing here), you'll end up in a polluted CBD, full or trucks and old cars.
4. Bonus points - Be practicalAs with any system or any engineering solution, 100% may not exist, but 99.5% is way better than 60%. Any bias should be addressed, but think if what you are solving is the highest priority.

Summary

Bias can be identified only after you set your expectations, be it ground truth for AI-based systems or derived from your core values. Once expectations are set, bias could be detected everywhere - sometimes easily fixed, and at times requires a new build.

Akridata 's Data Explorer will help you detect bias in your visual datasets - images or videos.