In 2025, the Stanford AI Index recorded 362 documented LLM safety incidents, a 55% increase from the year before. These are not obscure edge cases. A court sanctioned a lawyer for submitting filings containing fabricated citations generated by an AI tool. In another case, Google’s AI-generated search summaries confidently recommended adding glue to pizza and eating rocks for digestive health.
If you are building LLM-powered AI features, LLM safety needs to be a core design requirement that runs through every decision you make about how your product works, what it accepts, what it produces, and what it tells users.
Communicate AI Limitations Clearly
Product output shapes decisions. Users who do not know when the output is AI-generated, what the LLM can do, or when performance is degrading will not question it before acting on it. They will make decisions based on incorrect data.
To reduce this risk, alert users when they ask for something the product cannot do, or when a conversation has run long enough that quality is slipping, before they act on the next output.
Constrain and Validate LLM Inputs
Without input constraints, users can push the LLM beyond what it was designed to handle. That makes the output unpredictable and increases the risk of harmful content. Decide what your AI feature is for and enforce that through LLM input design. Character limits, input type validation, and scope constraints reduce the chance for misuse while keeping the feature focused on what it is actually designed to do. Most use cases do not need a free-form input field. Constraints built around actual use cases serve them better while making misuse harder.
Moderate and Filter LLM Outputs
LLMs can generate harmful content. Moderation is what stands between that content and your users. Define what acceptable output looks like for your product, then build the filtering and moderation tools that enforce it on every response before it reaches users.
Make Issue Reporting Accessible
Moderation will not catch everything. When harmful or incorrect output reaches users, they need a clear, easy way to flag it. Issue reporting is a safety mechanism, not just a way to collect feedback. Identifying problems early means you can fix them before they reach more users and erode trust.
Design Against AI Bias
LLMs inherit bias from their training data, and it shows up in outputs. The International AI Safety Report 2025 documents AI systems producing discriminatory outcomes across race, gender, age, disability, and other dimensions of identity. This is a safety issue, not only an ethics one. Biased LLM outputs cause real harm to real users.
Before you ship, define what fair output looks like for the people who will use your product and make it a condition of shipping. Without that standard, you have no way to evaluate whether the AI product is ready. After launch, monitor for unequal outcomes. LLM bias that does not surface in testing can appear in production.
Handle User Data Responsibly
LLMs can surface sensitive information in outputs, whether from training data or from earlier in the same conversation. A response can be on-topic and still include information that should not have appeared. Know what data your LLM has access to and limit what it can surface in outputs.
Users also need to know what happens to the data they submit: whether it is logged, how long it is kept, who can access it, and whether it is used to improve the model. A user who does not know the terms of the interaction cannot make an informed decision about what to share, and will not extend trust to a product that cannot answer that question.
The Design Is the Safety
Safety built into a product from the start is what makes it reliable. That is what users come back to.
