AI Ethics: What is Algorithm Bias, and How Is It Prevented?
#Artificial Intelligence (AI) is a remarkable tool that can process large amounts of data way faster and more accurately than any human, exceeding all of our analytical capabilities. AI will revolutionize the healthcare industry, but even the most innovative solutions are never without flaws. The biggest obstacle for #AI is ethicality. To function in healthcare, we must hold AI to the same ethical standards to which we hold providers, and in order to remain credible, it must not be biased.
Bias can be integrated and displayed in various ways and can cause many issues. An AI system, or more specifically, a Large Language Model (#LLM), with bias, can lead to data being pulled incorrectly, used incorrectly, taught incorrectly, and ultimately will generate an incorrect diagnosis.
To prevent bias, it's essential to know how #AI makes decisions.
#AI is powered by advanced algorithms that try to replicate the way humans behave and learn. AI inputs massive datasets into the algorithms, which in turn allows the model to analyze the data and recognize patterns. If the data is incorrect, incomplete, or flawed in any way, the results will be as well.
How do we prevent this?
There are three places where bias can be found: inside the data, the algorithm uses, within the algorithm's function, and through human influence. When the data is biased, that means the data is incorrect, not diverse enough, or incomplete, and can cause hallucinations.
AI can cause bias based on:
The model's data access quantity
Data Accuracy
Data Completeness
Integrated human bias
If an LLM is trained on 100 patients, and 99 of the profiles are those of men and one is a woman, the LLM would have more knowledge about men's health than women's. The model would have an easier time correctly recognizing issues and creating solutions for men, and as a result, its accuracy would fall short when diagnosing women. This is a simple but straightforward example that shows why data diversity is essential.
Not only are women slighted in this scenario, but the men are as well. A 100-patient profile data set may not contain enough variety to accurately diagnose anyone, which is why having a diverse and vast data set is paramount. However, none of this is relevant if the data is incorrect.
The way to prevent incorrect information is by controlling the LLM's exposure.
This is why I don't trust using GPT models as an assistant in any field where correct data is pertinent.
GPT1 was trained on BookCorpus data, a database of unpublished books from random people, including all of the free books on Smashwords. If you have ever been on that site, you would understand why this is concerning.
That's 7000 unpublished books of random genres without fact-checking any:
GPT2 was trained on Webtext, which is essentially every website posted on Reddit by any user with a few upvotes. That's 40 GB of text and 40 million web pages.
GPT3 was trained on CommonCrawl, a collection of Wikipedia articles, books, and Webtext.
Structurally, GPT2 is essentially GPT1 but more refined, and GPT3 is the same as GPT2 but with larger scaling.
GPT4 claims to be trained on all of the assets mentioned previously: Amazon reviews, random books, and articles.
This data is suitable for AI to observe an accurate depiction of our society and how we speak. This is a terrible foundation for anything that relies on accuracy or facts.
Because of its exposure, AI has already learned some very over-the-top racial stereotypes and is perpetuating these stereotypes in videos completely unprompted. This brain, as it is, can not be used in healthcare because its lack of ethics will restrict accuracy in diagnosing patients fairly and promptly.
Here are a few things to keep in mind when trying to prevent bias
What is the model being trained on?
Is the data set biased?
Is the data set accurate?
Is there bias created by humans integrated into the data?
Is the data sample big enough to represent a majority?
Is a human overseeing the process?
What is the history and pathway of how the algorithm was created?
Is it transparent?
Is the model learning from its mistakes?
Is the system being audited?
What happens if an AI model is biased?
Just like biased humans, AI bias can perpetuate social inequalities and hinder adequate progress and treatment. AI can discriminate and misprocess against entities it has never seen before. If a model only sees medical records from men, then it won't know how to deal with women's health issues. In fact, it might not even recognize them.
Having a bias can lead to enormous regulatory and legal consequences. An algorithm has the potential to latch onto parameters like race and gender and use them for calculations based on the data it has or hasn't been exposed to, which can create unfair practices.
This is why model transparency is crucial; if an error occurs, the human oversight team can identify and correct it promptly.
How to reverse bias?
Data-Centric solution:
Replace the incorrect data and continue training the model with the correct information
Add more information from underrepresented parties.
Algorithm-centric solution:
Make sure the algorithms are adjusted and transparent.
Use available frameworks that detect bias.
Human oversight solution:
Regularly audit the system using more than one testing method and make the necessary adjustments.
AI models, just like humans, are constantly learning and can relearn.
Make your ethical guidelines clear to help the model and the human oversight team maintain goals and clarity.
There are numerous complex and intricate aspects to dealing with AI models. It's essential to monitor and recognize errors or flaws and learn from them to create a tool that's ethical and worth using in healthcare. Once we master this, we open up unlimited possibilities with life-changing technology.
