Data stewards/owners are accountable for data governance and data quality across the organization. They define and enforce the standards that data must meet—such as accuracy, completeness, consistency, timeliness, and proper metadata—ensuring the data used by AI is reliable and well-managed. This role also covers data definitions, lineage, cataloging, access policies, and ongoing quality monitoring, which together establish trust in AI outputs. Data engineers, while they build and maintain the data pipelines that supply data, typically execute and operationalize those standards rather than own them. AI ethicists focus on ethical considerations in AI applications, such as fairness and transparency, not the formal governance of data quality. Privacy experts concentrate on protecting personal data and compliance with privacy laws, which is essential but addresses a specific aspect of governance rather than the overall data quality governance for AI.

A data lake is the repository that can hold both structured and unstructured data at scale while preserving governance and access controls from the source. It stores data in its native format, ranging from tables and CSVs to PDFs, images, and logs, enabling flexible schema-on-read analytics and data science workflows. Because data lakes are designed to integrate with security and governance mechanisms—like identity and access management, encryption, and policy-based controls—they can maintain or enforce access rules as data is ingested, stored, and consumed, helping preserve the origin’s security posture. The other options don’t fit as well. A data exploration and training platform focuses on tools and interfaces for analyzing data and building models rather than serving as a central, diverse data repository with integrated access controls. A vector database specializes in storing high-dimensional vectors for similarity search and is not typically used to aggregate all data types with preserved source-level access controls. An AI system production refers to deploying models and serving AI capabilities, not to storing and governing large, heterogeneous datasets.

Deliberately feeding incorrect data to an AI to produce incorrect results is data poisoning. The idea is to taint the information the model learns from, so its outputs become unreliable, biased, or aligned with the attacker’s goals. This typically happens during training or data collection, when manipulated data shifts the model’s understanding or creates vulnerabilities that show up later in deployment. Since the attack targets the model’s learned parameters and knowledge, the effects can persist and be hard to undo without remediation like data cleansing or retraining. Prompt injection, in contrast, aims to influence the model’s behavior at inference time through the prompt itself, not by altering the training data. Model drift refers to natural changes in data patterns over time that affect performance, not deliberate manipulation. Adversarial inference isn’t the standard term for this kind of data manipulation and is not the right descriptor for feeding bad data to corrupt training.

ROI from an AI program comes from a mix of outcomes, not just dollars. The most complete way to show value beyond financial returns is to track a range of metrics that reflect how AI boosts operations, speed, risk management, and stakeholder value. Productivity gains measure how much more work gets done or how much human effort is saved. Decision speed tracks how quickly insights translate into action. Risk reduction captures fewer incidents, lower error rates, and stronger safeguards. Customer impact looks at how AI improves customer experience, satisfaction, retention, or loyalty. Governance maturity and regulatory compliance gauge how well the organization is standardizing processes, improving model governance, audit readiness, and staying aligned with regulations. Together, these metrics paint a fuller picture of ROI by showing tangible improvements in efficiency, capability, and risk management, not just revenue or cost figures. Relying on only revenue growth misses efficiency and risk benefits; focusing only on cost savings omits speed, decision quality, and governance gains; and measuring only market share overlooks internal improvements and compliance.

Data validation, cleaning, and anomaly detection are all about protecting the data that trains and updates an AI system. In data poisoning, an attacker tries to insert manipulated data to distort the model’s behavior. Validation checks ensure data records conform to expected formats, types, ranges, and cross-feature consistency, so bad entries don’t slip through. Cleaning goes further by removing or correcting noisy or obviously faulty data, reducing the chance of poisoned examples influencing learning. Anomaly detection looks for data points that don’t fit the normal patterns or distributions, flagging suspicious samples for review or rejection before they affect the model. Used together, these mechanisms strengthen the data pipeline, making it harder for poisoned data to contaminate training. Other options focus on different aspects, like explaining how models decide their outputs or adding human oversight, rather than the data-layer defenses described here.