Artificial Intelligence Explained: What Is Ethical AI?

There is a lot of controversy in business circles as to whether companies are using artificial intelligence (AI) technology for unethical purposes, or unknowingly doing so. This post isn’t about that - it's about what ethical AI means for model development.

One of the most common misperceptions I hear about bias is, “If I don’t use age, gender or race, or similar factors in my model, it’s not biased.” Unfortunately, that’s not true.

From a data scientist’s point of view, ethical AI  is achieved by taking precautions to expose what the underlying machine learning (ML) model has learned, and if it could impute bias. At first glance, the precautions often taken to isolate the input data fields used by models may seem sufficient. However, latent features of the model, which combine the inputs, are difficult to interpret as to whether or not they inject bias. Upon deeper inspection, the model often produces outcomes that are biased toward a particular class. (Here I am referring to data class, not socioeconomic.)

Bias and Confounding Variables

Machine learning learns relationships between data to fit a particular objective function (or goal). It will often form proxies for avoided inputs, and these proxies show bias. Bias is exposed when “confounding variables” cause these proxies to be more activated by one data class versus another, driving the model to produce biased results.  

For example, if a model includes the brand and version of an individual’s mobile phone, that data can be related to the ability to afford an expensive cell phone — a characteristic that can impute income. If income is not a desirable factor to use directly in the decision, imputing that information from data such as the type of phone, or the value of purchases the individual makes, introduces bias into the model. This is because, on average, affluent customers can afford more high-end, expensive phones than a non-affluent group.

Research into the effects of smoking provides another example of confounding variables. In decades past, research was produced that essentially made the reassuring correlation, “If you smoke, your probability of dying in the next four years is fairly low. Therefore, smoking is OK.” The confounding variable in this assumption was the distribution of age of smokers; in the past, the smoking population contained many younger smokers whose cancer would develop later in life. Many older smokers were already deceased and therefore their contribution minimized in reaching this finding. Thus, the analytic models representative of the “smoking is OK” conclusion contained overwhelming bias driven by a higher density of younger smokers, thus creating a biased perception about the safety of smoking.

Today, similar bias could be produced by a model concluding that, since far fewer young people smoke cigarettes than 50 years ago, nicotine addiction levels are down, too. However, youth use of e-cigarettes jumped 78% between 2017 and 2018 — to one out of every five high-school students. E-cigarettes are potent nicotine delivery devices, fostering rapid nicotine addiction and simply diverting nicotine use to a new delivery vehicle.  Without reflecting this nicotine delivery method, we would have an errant view of nicotine addiction among youth.

Finding Hidden Bias

The challenge of delivering truly ethical AI requires closely examining each data class separately, with respect to the relationships in the data that drive outcomes: the latent features. As data scientists, we must demonstrate to ourselves, and the world, that AI and machine learning technologies are not subjecting specific populations to bias and search for confounding variables. To reach that goal, the relationships learned need to be exposed using explainable latent feature technologies rather than complex webs of interconnected variables. The latter contain relationships that need to be tested but can’t be extracted from the machine learning models.

Three Recommendations for Ethical AI

The ethical issues concerning AI is a thorny topic, but there are steps you can take both during model development and during model execution. Here are three recommendations.

1. Use Blockchain as a Single Source of Truth in AI Development

This is something we practice at FICO in developing models using machine learning. We use blockchain technology as the single source of truth, allowing audit trails of data usage in models, particularly in data permission rights. Our patented application (which won an award from Banking Tech Awards  last year) uses blockchain to ensure that all the decisions made about a machine learning model are recorded and are auditable. These include the model’s variables, model design, training and test data utilized, selection of features, the ability to view the model’s raw latent features, test for bias among latent features and show which are removed, and recording to the blockchain all scientists who built different portions of the variable sets, and participated in model creation and model testing. You can read more about this important form of AI governance in my post, How to Use Blockchain to Build Responsible AI: An Award-Winning Approach, and in my recent article for the Harvard Business Review.

2. Use Focused Language Models

While large language models (LLM) used in generative AI have grabbed the spotlight, focused language models involve using domain-specific data. This approach is distinctly different from commercially available LLMs and SLMs, which offer no control of the data used to build the model , as this factor is crucial for preventing hallucinations and harm.

In a focused language model, you define the vocabulary of the problem domain, ensuring that data is aggressively filtered. Such a model in financial services would only include, for instance, credit risk, financial inclusion or payment card fraud.  The same applies for any problem domain where the utmost accuracy, reduced hallucinations and detailed data design are all required for trusted use.

A focused language model enables GenAI to be used responsibly and ethically because:

• It affords transparency into how a core domain-focused language model is built.
• On top of industry domain-focused language models, users can create task-specific focused language models with tight vocabulary and training contexts for the task at hand.
• Most importantly, the resulting FLM is accompanied by a trust score with every response, giving users a yardstick by which to measure its accuracy and, therefore, reliability. The trust score is a secondary analytic model that provides a trust score from 1 to 999, reflecting the probability that the key contexts (such as product documentation) that the task-specific FLM was trained on are used to provide the answer. Users can use the score to decide if they will trust the FLM’s answer.

You can read more about this in my Forbes article, Responsible GenAI: Earning Trust May Be Easier Than You Think.

3. Use "Humble AI" to Improve Accuracy and Ethical Use

The humble AI approach recognizes that AI models are fallible, and builds in a failsafe mechanism that reduces the risk of erratic or erroneous decisions. In this approach, companies have hot backups for their AI deployments, to tier down to safer tech when an audit indicates AI decisioning is not trustworthy.

Deciding when to move to Humble AI is done through real-time monitoring of the latent features of machine learning models, as they must behave in consistent ways defined by how the model was built and observing their behaviors. Latent features will have their distributions of activating, and groups may frequently fire together; one can measure deviations of firing distributions of individual latent features or groups firing together. These deviations may indicate that the data domain is changing in ways that invalidate the suitability of the machine learning model, and this is the trigger to move to the humble AI alternative model.

Are You Following Ethical AI Best Practices?

Here are some questions you can ask of your AI use:

• How is your company achieving Ethical AI?
• Which AI technologies are allowed for use in your organization, and how will they be tested to ensure their appropriateness for the market and unbiased representation?
• Is there monitoring in place today for each AI model and, if so, what’s being monitored?
• What are the monitoring thresholds preset to indicate when a AI model should no longer be used?  And is the humble AI model alternative ready to hot swap?
• Is your organization uniformly ethical with its AI?
• Is your company placing some models under the Responsible AI umbrella (due to being regulated and therefore high risk) while others are simply not built to the Responsible AI standard? How are those dividing lines set?
• Is it ever OK to not to be responsible in the development of AI? If so, when?

There are two more blogs in my AI explainer series on the three Es of AI: explainable AI and efficient AI. 

Explore FICO and Responsible AI

• Learn more at FICO® Responsible AI
• Download our AI Playbook: A Step-by-Step Guide for Achieving Responsible AI
• Follow me on Twitter @ScottZoldiFollow me on Twitter @ScottZoldi to stay in touch.

Note: A version of this post was published in IOT Agenda. This is an update of a post from 2019.