you can do to reduce bias in AI and how it can harm marketing data

Algorithms are the core of martech and marketing. Algorithms are used to analyze data, collect data, segment audience and many other purposes. They are the core of artificial intelligence that is built upon them. Marketers depend on AI systems to provide reliable, neutral data. It can cause misdirection in your marketing efforts if it doesn’t.

Algorithms are often viewed as a set of rules that can be applied without bias or intention. They are exactly that in themselves. They do not have opinions. These rules are based on the assumptions and values of their creator. This is one way bias can be introduced to AI. Another, and perhaps more important, way AI can learn from data is through its training.

Look deeper: ChatGPT and Bard will make your search experience more enjoyable

Facial recognition systems, for example, are trained using images of people with lighter skin to identify faces. They are known for not being able to recognize darker-skinned people. 28 members of Congress were wrongly matched with their mugshot photos. This failure to rectify the situation has led many companies, including Microsoft, to discontinue selling these systems to police departments.

ChatGPT, Google’s Bard, and other AI-powered chatbots use autoregressive languages models that use deep learning to create text. This learning is based on huge amounts of data, which could include everything that was posted online during a particular time period. It can also be contaminated with bias, disinformation, and error.

It is only as good as the data that it receives

Paul Roetzer is the founder and CEO at The Marketing AI Institute. It’s a mirror of humanity in many different ways.

This is something that the builders of these systems know.

” According to [ChatGPT creator] OpenAI, negative sentiment is more closely associated African American female names than with any other set of names within there,” Christopher Penn cofounder and chief data scientist at TrustInsights.ai. If you use any type of fully automated black-box sentiment modeling and you are judging people’s names, then you have a problem. These biases are reinforced.

OpenAI’s best practice documents also state that “From hallucinating incorrect information to offensive outputs to bias and much more, language model may not be appropriate for every use case without significant modifications.”

What does a marketer do?

Marketers who want to use the best data possible must mitigate bias. It will always be a moving target.

Christopher Penn says that marketers and martech companies need to ask themselves, “How can we apply this to the training data so that the model has fewer biases that we have later?” You don’t need to filter out garbage if you don’t want it in.

You can use tools to help you accomplish this. These are the top five most popular tools:

• Google’s What-If is an open source tool that can detect bias in a model. It allows you to manipulate data points, generate plots, and specify criteria to determine if any changes have an impact on the final result.
• AI Fairness 360 is an open-source toolkit that detects and eliminates bias in machine learning models.
• Fairlearn from Microsoft to assist in navigating the trade-offs between fairness, model performance, and fairness.
• Local Interpretable Modular-Agnostic Explanations (LIME), created and maintained by Marco Tulio Ribeiro lets users modify different parts of a model in order to understand the source of bias, if any.
• FairML from MIT’s Julius Adebayo provides an end-to-end toolbox to audit predictive models by quantifying their relative significance.

Penn says that they are best when you know exactly what you’re looking at. They are less effective if you don’t know what’s inside the box.

Judging inputs is the simple part

He explains that AI Fairness 360 can be used to give it loan decisions as well as a list with protected classes such age, gender, and race. It will detect any biases in training data and in the model, and alarm if it starts drifting in a biased direction.

Penn states that generation is a lot more difficult than copywriting or imagery. “The tools currently available are mostly meant for tabular rectangular data that has clear outcomes that you want to mitigate against.

ChatGPT and Bard are extremely computing-intensive systems that produce content. Additional safeguards against bias can have a significant effect on their performance. It will only make it more difficult to build them. Don’t expect any quick resolution.

Marketers can’t afford just to wait for models to work. Brand risk is a major concern. They must constantly ask what could go wrong with AI-generated content. This is the mitigation they should be doing. People who are most qualified to ask these questions are those involved in diversity, equity, and inclusion.

Penn says that organizations give little attention to DEI initiatives, but this is where DEI can really shine. Have the diversity team inspect the outputs of models and tell them if they are OK or not. Then have it be integrated into processes like DEI has approved.

These systems are a key indicator of a company’s culture.

Paul Roetzer says that each organization will have to come up with their own rules about how to use and develop this technology. “And I don’t know how it can be solved other than at the subjective level of ‘this is what we consider bias to be’ and we will, (or will not) use tools that enable this to occur.

How AI bias can harm marketing data and how you can fix it was originally published by MarTech.