3 Ways Companies Can Ethically and Effectively Develop Generative AI Models

President Biden is meeting with AI experts to consider the dangers of AI. Sam Altman and Elon Musk have publicly expressed their concerns. Consulting giant Accenture became the latest company to bet on AI, announcing plans to invest $3 billion in the technology and double its AI staff to 80,000. That’s more than any other consulting firm, with Microsoft, Alphabet and Nvidia also joining the fray.

Leading companies are not waiting for the bias problem to go away before embracing AI, so solving one of the biggest challenges facing all major generative AI models is even more important. It’s urgent. But AI regulation takes time.

Since all AI models are built by humans and trained on data collected by humans, it is impossible to completely eliminate bias. However, developers should try to minimize the amount of “real world” bias they reproduce in their models.

Real-world biases in AI

To understand real-world biases, imagine an AI model trained to determine who is eligible for a mortgage. Training that model on the decisions of individual human loan officers, some of whom may implicitly and unreasonably avoid lending to people of a particular race, religion, or gender. No, but there is a big risk that real-world biases will be replicated in the results.

The same is true for models intended to mimic the thought processes of doctors, lawyers, human resources managers, and countless other professionals.

>>Follow VentureBeat’s Coverage of Continuously Generated AI<

AI presents a unique opportunity to standardize these services in a way that avoids bias. Conversely, failure to limit model bias risks standardizing seriously flawed services to the benefit of some and at the expense of others.

Here are three key steps founders and developers can take to get it right.

1. Choose a suitable training method for your generative AI model

ChatGPT, for example, falls into the broader category of machine learning as a large scale language model (LLM). This means absorbing vast amounts of text data and inferring relationships between words in the text. On the user side, the LLM fills in the blanks with the statistically most likely words given the surrounding context when answering the question.

However, there are many ways to train machine learning models on data. For example, some medical technology models rely on big data to train AI using individual patient records or individual physician decisions. For founders building industry-specific models such as medical AI or HR AI, such a big data approach can be unnecessarily biased.

Imagine an AI chatbot trained to respond to patients and create clinical summaries of medical presentations for doctors. When built with the above approach, the chatbot references millions of other patient data (records in this case) to create its output.

Such models can produce accurate output at an alarming rate, but they also import the biases of millions of individual patient records. In that sense, big data AI models are a mixture of biases that are difficult to track, let alone correct.

An alternative to such machine learning techniques (especially for industry-specific AI) is to train the model on the gold standard of industry knowledge to avoid imparting bias. For medicine, it’s peer-reviewed medical literature. Laws can be national or state legal documents, and in the case of self-driving cars, actual traffic rules rather than individual human driver data.

Yes, even those texts are man-made and contain prejudices. But given that all doctors strive to master the medical literature, and that all lawyers spend countless hours studying legal documents, such literature is less biased AI. It could serve as a reasonable starting point to build on.

2. Balance literature and changing real-world data

While my field of medicine is full of human biases, it is also true that different ethnicities, ages, socioeconomic groups, locations, and genders have different levels of risk for certain diseases. More African Americans than whites suffer from high blood pressure, and Ashkenazi Jews are known to be more prone to certain diseases than other groups.

These are notable differences as they are taken into account in providing the best possible care to our patients. Nevertheless, it is important to understand the roots of these differences in the literature before injecting them into the model. Do prejudices against women result in doctors giving women a higher rate of certain medications, increasing their risk of certain diseases?

Once you understand the root of the prejudice, you will be much more prepared to correct it. Let’s go back to the mortgage example. Fannie Mae and Freddie Mac, which back most mortgages in the U.S., found that people of color were more likely to earn income from jobs in the gig economy, Business Insider reported last year. Even though many gig economy workers still have strong rent-paying histories, such income is perceived as precarious, unfairly hindering them from securing a mortgage.

To correct for that bias, Fannie Mae decided to add a rent payment history variable that is relevant to credit rating decisions. Founders need to build adaptable models that can balance formal, evidence-based industry literature with changing real-world facts in the field.

3. Build Transparency into Generative AI Models

Detecting and correcting for bias requires a window into how the model reaches its conclusions. Many AI models do not track the original source or explain their output.

Such models often produce responses confidently and with surprising accuracy. See ChatGPT’s miraculous success. But if not, it’s nearly impossible to determine what went wrong and how to prevent inaccurate or skewed output in the future.

Given that we’re building technology that transforms everything from work to commerce to medicine, it’s important that humans can find and fix flaws in that reasoning. It’s not enough to simply know that the answer was wrong. Only then will we be able to act responsibly for the achievements of such technologies.

One of AI’s most promising value propositions for humanity is the massive removal of human bias from healthcare, employment, lending, justice, and other industries. It will only be possible if we cultivate a culture among AI founders committed to finding effective solutions to minimize the human biases introduced into our models.

Michal Tzuchman-Katz, M.D., is Co-Founder, CEO and Chief Medical Officer. of Kahun medicine.