Expert-curated AI risk mitigation strategies for enterprise governance. Each mitigation includes implementation guidance, suggested evidence, and links to the specific AI risks it addresses.
Implementing measures to ensure that only authorized individuals can access, modify, or utilize AI systems and their data.
Educating employees about AI systems, their potential impacts, and ethical and regulatory considerations.
Establishing a documented set of policies and procedures outlining expectations for how to use AI tools.
Incorporating techniques for mitigating bias into the model pipeline.
Creating a process that allows individuals to contest decisions made by an AI system.
Using a version control system (e.g. Github) to keep track of all code used during development and deployment.
Gathering data from various sources to ensure AI models are fair, unbiased, and accurate across different scenarios
Maintaining a clear record of the exact data used to train different model versions.
Setting rigorous standards and processes for documenting AI models and dataset.
Combining several base models to produce a more robust final model.
Soliciting participation from impacted stakeholders throughout the AI lifecycle.
Creating a high-quality dataset for evaluating data labeling and model performance
Implementing a mechanism for detecting hallucinations in the output of models.
Creating systems and tools that allows an individual to modify the output of an AI system.
Incorporating human review and approval processes in AI systems.
Creating a plan that outlines roles, responsibilities, escalation paths and external communication protocols for AI incidents.
Building systems from scratch to avoid using vulnerable components from other sources.
Limiting the public release of technical information about the system.
Manually reviewing training data to ensure quality and identify potential biases or errors.
Granting AI Systems access to only the minimum set of external systems and resources needed to function effectively.
Providing users with technical information about the AI system's data, design, performance and capabilities.
Adjusting hyperparameters to control the diversity, creativity, and determinism of the model's outputs.
Implementing a system to track and evaluate the performance of deployed AI systems.
Implementing a system to track different versions of AI models, data, and code.
Using multiple annotators to improve the quality and accuracy of data labeling.
Implementing comprehensive policies and procedures to manage data collection and storage activities.
Defining and enforcing the least acceptable level of accuracy for a model.
Review system performance and conducting new risk assessment on a periodic basis after a system is deployed.
Using prompts that create a clear boundary between the instructions and the user input.
Using prompting techniques, like Chain-of-Thought and Self-Refinement, can reduce the likelihood of LLM hallucinations.
Including an option for a human to intervene and take over the task.
Assessing the impact of the AI System on natural resources and prioritizing renewable energy sources.
Creating restricted development environments that limit access to external resources.
Combining a Large Language Models with an external knowledge bases.
Sanitizing training data prior to use to remove both inappropriate and poisoned content.
Intentionally incorporating and highlighting frictions within the user interface to encourage reflection, critical thinking, and intentional use.
Maintaining secondary models that can be deployed in the event that primary models fail.
Using secure and encrypted transfer methods when moving assets, like data and models.
Hosting externally built models inside of existing architectures or firewalls.
Using structured inputs and outputs over free-form text to improve reliability and safety.
Using synthetic data to augment and expand datasets to more completely cover the types of data seen in the deployed setting.
Providing an instruction to an AI model to guide its responses and behavior according to specific guidelines or objectives.
Creating simple tests that validate whether parts of the system function as expected.
Obtaining consent from the user before using their data to train or fine-tune AI systems.
Taking measures to ensure that external data and model sources are trustworthy.
Systematically examining AI systems and related assets for potential security weaknesses, threats, and signs of malicious activity.
Trustible recommends specific mitigations based on your AI use cases and risk profiles, with evidence tracking and accountability built in.
Explore the Platform
