Tableau AI and Trust
It’s important that your data stays safe while you innovate with new technology. With Tableau AI, we keep Trust as the #1 value, and we strive to make sure your data is secure while also creating experiences that are accurate and safe.
Tableau AI and your data
To keep your data secure, Salesforce has agreements in place with Large Language Model (LLM) providers, like OpenAI. Organizations can leverage generative AI capabilities without their private data being used to train the LLM.
Trusted generative AI
Salesforce’s Einstein generative AI solutions are designed, developed, and delivered based on five principles for trusted generative AI.
-
Accuracy: We prioritize accuracy, precision, and recall in our models, and we back our model outputs up with explanations and sources whenever possible. We recommend that a human check model output before sharing with end users.
-
Safety: We work to detect and mitigate bias, toxicity, and harmful outputs from models used in our products through industry-leading detection and mitigation techniques.
-
Transparency: We ensure that our models and features respect data provenance and are grounded in your data whenever possible.
-
Empowerment: We believe our products should augment people’s capabilities and make them more efficient and purposeful in their work.
-
Sustainability: We strive towards building right-sized models that prioritize accuracy and reduce our carbon footprint.
To learn more about trusted AI, see Salesforce Research: Trusted AI(Link opens in a new window)
The Einstein Trust Layer in action
Tableau AI is powered by Einstein AI and inherits The Einstein Trust Layer and security controls.
With Tableau Pulse we ground the insight summaries that we generate using templated natural language insights and values calculated using deterministic statistical models. Tableau Pulse is also based on a metric layer that provides a bounded, safe space for insights to be detected.
Tableau Pulse uses generative AI to enhance and synthesize the language of the insights generated by Tableau. The result is summarized insights in easy to understand language that the user can quickly engage with.
To enable Einstein Copilot for Tableau to return a viz, a calculation, or an asset description, we first need to ground Einstein Copilot in your data.
When you launch Einstein Copilot, we query the data source that you're connected to and create a summary of field names, descriptions, data types, and the first 1000 unique field values (non-numerical) per field. This summary is sent to the Large Language Model (LLM) to create vector embeddings so that Einstein Copilot can understand the context of your data. The summary creation happens within Tableau and the summary context data is forgotten by the LLM as soon as the vector embeddings are created.
When you type a question or request into the conversation pane in Einstein Copilot, a combined prompt consisting of the user's input, metadata describing the current state of the viz (web authoring), and historical context from the conversation pane flow through the Einstein Trust Layer to the LLM. Sensitive information is detected and masked with placeholder text to maintain context. For more information about PII masking see Einstein Trust Layer Region Language Support(Link opens in a new window) in the Salesforce help.
The response flows back through the Einstein Trust Layer to check for toxicity and unmask any masked data. Due to our zero data retention policies in place with our third-party LLM providers, any data sent to the LLM isn’t retained and is deleted after a response is sent back.
The result is a viz, a calculation, or an asset description ready for you to review.
Techniques like this ensure our products adopt generative AI in a trusted manner. At the same time, your customer data isn’t used to train any global model.
Want to learn more about the Einstein Trust Layer? See Einstein Trust layer: Designed for Trust(Link opens in a new window) in the Salesforce help, or take the Einstein Trust Layer(Link opens in a new window) module on Salesforce Trailhead.
Reviewing generative AI outputs
Tableau AI is a tool that can help you quickly discover insights, make smarter business decisions, and be more productive. This technology isn’t a replacement for human judgment. You’re ultimately responsible for any LLM-generated outputs you incorporate into your data analysis and share with your users.
Whether it’s generating calculations to use in a Tableau Prep flow, summarizing insights for metrics you follow, creating visualizations for you from your data, or drafting descriptions for your data assets, it’s important to always verify that the LLM output is accurate and appropriate.
Focus on the accuracy and safety of the content before you incorporate it into your flows, visualizations, and analysis.
-
Accuracy: Generative AI can sometimes “hallucinate”—fabricate output that isn’t grounded in fact or existing sources. Before you incorporate any suggestions, check to make sure that key details are correct. For example, is the proposed syntax for a calculation supported by Tableau?
-
Bias and Toxicity: Because AI is created by humans and trained on data created by humans, it can also contain bias against historically marginalized groups. Rarely, some outputs can contain harmful language. Check your outputs to make sure they’re appropriate for your users.
If the output doesn’t meet your standards or business needs, you don’t have to use it. Some features allow you to edit the response directly before applying it to your data, and you can also try starting over to generate another output. To help us improve the output, let us know what was wrong by using the thumbs up and thumbs down buttons where available and provide feedback.