Advancing Trusted AI in ArcGIS

At Esri, we prioritize trust in AI development and deployment. Trusted AI in ArcGIS focuses on security, privacy, transparency, fairness, reliability, and accountability. This reflects Esri's values and commitment to responsible innovation, bridging the AI trust gap, and fostering positive societal change. Access the PDF version of Esri's Advancing Trusted AI in ArcGIS overview in the ArcGIS Trust Center.

Legacy of AI trust and innovation

Before the generative AI boom of the past few years, when people talked about AI, typically they were referring to machine learning and deep learning models used in pattern recognition, forecasting, object detection, change detection, and more. Over a decade ago, Esri started with machine learning to perform clustering, regression, and classification on spatial data. More recently, work has continued in both the machine learning and deep learning space, including the introduction of pre-trained deep learning modelspre-trained deep learning models that are trained using massive, diverse data sets to make it easier to get started with tasks such as feature extraction, point cloud classification, and image redaction. This work falls under the umbrella of what Esri refers to as GeoAI.

Generative AI refers to a type of machine learning model designed to create new content, or insightful recommendations, by learning patterns from large-scale datasets. Generative AI models are trained on extensive datasets, which may include very large, domain-specific data and/or data from the internet. Unlike traditional AI models that focus on making predictions or performing analysis, generative AI models are used to generate creative outputs such as text, images, or other forms of content.

In contrast to Generative AI as described above, GeoAI focuses on analyzing and interpreting geospatial data to uncover patterns and make predictions using the input data. Generative AI can be thought of as a machine learning model that is trained to create new data, rather than making a prediction about a specific dataset and is more non-deterministic allowing for more creative solutions. A generative AI system is one that learns to generate more objects that look like the data it was trained on. An example of ArcGIS capabilities incorporating generative AI are AI Assistants within ArcGIS.

The actual machinery underlying generative AI and other types of AI oftentimes utilize the same algorithms, which can blur the distinction between the types. Generative AI’s quick proliferation and broader use cases has resulted in expedited regulatory requirements and customer demands for stronger tranparency and control.

AI landscape today

The AI landscape is rapidly evolving. Governments around the world are actively shaping the future of AI by enacting new laws and frameworks. For instance, the European Union recently adopted the EU AI Act, which establishes regulations on high-risk AI applications. Similarly, the United States has seen proposals like the AI Bill of Rights and Executive Order M-24-10 addressing potential risks and biases in AI systems.

Esri has recognized the importance of staying ahead of these evolving legal requirements, responsible development, and ethical considerations. We proactively align our AI practices with key regulations and industry-recognized frameworks. This includes following the guidance set forth by laws and regulations stated above.

ArcGIS guiding AI principles

Esri's Trusted AI framework is grounded in six core principles:

• Security: We are committed to safeguarding security and mitigating risks in our AI systems through a secure-by-design approach while ensuring responsible AI that proactively protects against security threats.
• Privacy: We prioritize protecting user data and ensuring the privacy of AI throughout the AI lifecycle ensuring compliance with global privacy standards through privacy-by-design methodologies, data anonymization, and data minimization.
• Transparency: We provide clear visibility about our AI models, empowering informed decision-making about our AI processes, limitations, and outcomes.
• Fairness: Esri has long upheld the principles of fairness, ethics, and societal responsibility in its everyday practices. These core values are embedded in our approach to decision-making, product development, and community engagement.
• Reliability: Our AI is carefully tested and validated to deliver consistent and dependable results across diverse environments and use cases.
• Accountability: We maintain accountability by establishing clear governance frameworks, holding our teams responsible for AI deployment and monitoring, ensuring human oversight remains central to all AI-related decisions.

Esri's approach to Trusted AI

Design and development

Esri employs a risk evaluation process to assess AI features, ensuring they meet privacy and security standards. Through human-in-the-loop design, ethical guardrails, red teaming, and holistic lab testing, Esri minimizes risks while ensuring inclusivity and user control.

Customer choice

Generative AI in ArcGIS is opt-in, empowering users to control when and how AI capabilities are used. Customers can enable or restrict AI functionalities via administrative settings, ensuring alignment with organizational policies.

Data handling

Customer data privacy is central to Esri's AI strategy:

• Esri does not train AI models using customer data without explicit authorization.
• Customers retain ownership of prompts and data they provide for AI analysis within ArcGIS products

AI functionalities adhere to data segmentation and anonymization practices, safeguarding data ownership and integrity.

Metadata and transparency

Esri has developed AI Transparency Cards that detail the functionality, validation, and safeguards of AI features. These cards provide context for responsible AI usage, complementing industry standards like AI Model Cards.

Governance

In 2023, Esri established an AI Governance Board to oversee adherence to trusted AI principles. This collaborative initiative ensures innovation is guided by ethical and regulatory considerations.

Understanding AI in ArcGIS

Esri leverages two major categories of AI within ArcGIS:

• GeoAI: Uses pre-trained deep learning models for tasks such as feature extraction, pattern recognition, and predictive analysis. GeoAI integrates into geospatial workflows, empowering users to analyze spatial data with advanced computational tools.

• Generative AI: Incorporates AI Assistants and other generative technologies that support creativity, user productivity, and automated workflows.

For a deeper dive into Esri's Trusted AI strategy, visit our Advancing Trusted AI in ArcGIS resource.

Generative AI Safeguards

The following aspects are applicable for all generative AI features we incorporate into our products:

• Generative AI features are NOT enabled by default in ArcGIS product
  • End-users and/or administrators must opt-in to utilize generative features.
• Unless you explicitly authorize Esri, data and prompts you use with AI assistants will
  • Not be used to train AI models.
  • Remain private to you and will never be shared.
• When 3rd party AI services are utilized
  • Enterprise-class AI instances are utilized to segment and protect your data.
• Generative AI features have limitations
  • The potential for inconsistent output part of the creative consequences of generative AI.
• Customers should always incorporate a human to assess AI output to ensure appropriateness

AI Deep Learning Packages

Esri has provided GeoAI capabilities for years, such as the Deep Learning Packages available within the Living Atlas, which do NOT utilize Generative AI models. These deep learning packages contain an item description page which may contain fields such as: intended use, model architecture, training data used, licensing, tuning guidance, performance metrics, samples, input/output, and potential limitations. Additionally, we strongly recommend only downloading packages from sources marked as Authoritative, which means the organization publishing the content has been verified by Esri. You should reach out to publishers of content you want to consume and ask them to complete verification before downloading their content.

Non-Product AI Solutions

Beyond ArcGIS products, Esri offers generative AI solutions such as the Esri Support AI Chatbot and the Product Marketing Chatbot. These tools help users navigate Esri's ecosystem more efficiently. However, as these chatbots are not embedded within ArcGIS products, their usage may involve separate terms and data practices, such as marketing-related tracking.

Collaborating in Trusted AI

Achieving trusted AI is a shared responsibility. Esri partners with customers to implement best practices in data governance, ethical AI use, and human oversight. Feedback from users strengthens our solutions, fostering a collaborative environment where trust and innovation thrive. Please see our customer AI Implementation best practice guidance for more information.

Moving forward with trust

Esri's commitment to trusted AI remains constant as we advance the capabilities of GeoAI and Generative AI. By integrating these technologies responsibly, we are building a future where AI empowers positive change while maintaining the highest standards of trustworthiness.

External references

In alignment with Esri's commitment to secure and responsible AI development, the following external resources were referenced in development of our AI policy.

• Federal register: Safe, Secure and Trustworthy Development and User of Artificial Intelligence
• EU AI Act: First regulation on artificial intelligence
• Four principles of explainable Artificial Intelligence: NIST.IR.8312.pdf
• Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile NIST.AI.600.1.pdf