Tableau Conference highlights: The principles of responsible AI
Reflecting on the 2024 Tableau Conference session, learn how to implement ethical AI solutions with trust, security, safety, and privacy at the core.
By Laurie Rugemer and Jake Kough
The announcements at Tableau Conference this year centered around the future of analytics with AI and Tableau’s advancements in this space with Tableau Pulse and Tableau AI built on Salesforce Einstein 1. Our conference session, “Responsible AI,” focused on considerations and action items for businesses as they make plans to ramp up AI solutions ethically.
Why care about the principles of responsible AI?
AI is already all around us with search engines, personalized feeds on social media, chatbots, personalized recommendations on our favorite streaming systems, et cetera. At its core, it is the science of teaching a machine how to perform tasks and make decisions that normally require human intelligence. The introduction of GenAI models has brought AI front and center and has been the catalyst for an AI revolution, but it is not the AI to end all AI. It is important to understand where GenAI sits in the landscape of AI and machine learning and to recognize that the principles of responsible AI apply across all use cases.
Generative AI is a subset of machine learning and LLMs, in particular, combining deep learning with natural language processing to create original text. As is the case with all machine learning models and any AI solution, the model results and predictions are only as good as the data they are trained on.
With that idea in mind, it is easy to understand that a discussion of responsible AI principles starts with questions about data quality.
Good AI is good data
As Benioff says above, the cornerstone of moving forward with AI is being committed to ethical AI solutions that hold trust, security, safety, and privacy as core tenets. This is two-pronged — it is both important in the way we build AI solutions (the data we use, the processes to ensure accuracy, how results will be used for decision-making), and in how we ensure a robust data foundation before embarking on any AI solution to make sure that our data is as accurate and consistent as possible.
Salesforce’s AI Principles (above) underpin the importance of first thinking about why AI is being implemented, how the data is being collected and safeguarded, and how and to whom we’re applying results could be affected both positively and negatively. These principles underscore the importance of starting with a strong foundation of high-quality data to ensure that models built are as accurate as possible and are representative of the population they’re being applied to. In the end, good AI is good data.
Data infrastructure and management
Good data begins with strong processes, principles, and rules. A company’s data governance creates the foundation for strong AI by defining how, when, and why data is created, collected, and used. Data governance has a direct correlation to data quality, consistency, accuracy, and privacy. Data quality can be improved through validation rules, required fields, standardized field definitions across user bases, and avoiding redundant or underpopulated fields within the table. Implementing these steps is the beginning of ensuring quality data for AI training datasets. Additionally, data governance will increase the company’s data security and privacy through auditing access and adhering to regulations. Lapses in security can undermine a user’s trust in AI and their data. Companies should also avoid sharing data with third-party models and process unless they have been thoroughly vetted by security as third parties may also pose security threats.
Master data management, the practice of creating a single master record for each person, place, or thing across internal and external data sources, is another tool leveraged by data scientists for quality data. De-duplicated, reconciled and enriched records allow for consistent and reliable source data. MDM serves as a trusted view to promote accurate reporting, reduce data errors, and remove redundancy. Therefore, companies may have confidence data points are consistently represented no matter who or what is accessing the data and may be better interpreted by AI without introducing skews or misleading data.
Mitigating bias
Bias can creep into AI solutions in many stages of the process. If we don’t have data governance in place, the data being used to train models can introduce duplicate information, underpopulated fields that are not null at random, or generally not match our business process and therefore not be representative or give meaningful results.
Bias can also be introduced through the data we chose to train a model.
For example, if I am building a product recommendation system and mark toy data as “girl” or “boy,” I am likely introducing gender stereotypes into the training data (Association Bias). Furthermore, if I use that data to recommend products to shoppers and they purchase based on those recommendations, I may be bringing in a second layer of bias, Confirmation Bias, because I won’t know if they purchased those products because they really belong in those categories or because that’s what they were shown from the AI system.
To reduce bias and begin to incorporate the principles of AI, below are three overarching questions that should be answered, with action-oriented sub questions to help.
1. Why AI?
- What is the problem you’re trying to solve, and is the use case/question appropriate for AI?
- How explainable are the outcomes and predictions based on the chosen AI solution?
- How will I seek feedback from my stakeholders to ensure the output makes sense?
2. Do I have the right data?
- Do I have MDM and data governance in place?
- Do I have a diverse dataset that represents the population I want my AI system to solve for (and represents it across all factors and segments)?
- Does my data make sense from a data science solutioning perspective, for example, sample size, skew, model assumptions?
3. Do I have the right AI process?
- Am I involving a diverse team beyond data scientists (product, UX, business) in the model-building process?
- How will I monitor the output to ensure it’s performing well for all segments of the population?
- How will I share results with my stakeholders, potentially through Tableau dashboards to help understand the “why” behind predictions?
Responsible AI in Tableau
Visualizing data in a platform like Tableau before embarking on the build-out of any AI solution is imperative for this process. Exploratory data analysis to understand if we have appropriate data to answer important business questions is an important first step to knowing if we’re ready for AI.
After building an AI model, like a customer churn model for example, it’s also valuable to create model monitoring and explainability dashboards that show the why behind predictions (what causes some to be higher or lower, how can we explain differences based on the features in the model).
Beyond dashboards, Tableau Pulse has reimagined the analytics experience to include generative AI synopses of key metrics that our users need to see and digest quickly.
Tableau AI enhancements will continue with the release of Einstein 1 and the ability to easily visualize and share model results with the people who need to understand and take action on them.
Ramping up on AI means investing in a solid foundation of high-quality, accurate data that can be used to train these models, and incorporating the principles of responsible AI at the onset of this process is key to user adoption and trust.
Slalom is a next-generation professional services company creating value at the intersection of business, technology, and humanity. Learn more about our partnership with Tableau and reach out to discuss how to prepare for Tableau Pulse.
