Two nearly identical cream cards on either side of a brass boundary line on a linen desk, one stopped by a black block.

Amazon Bedrock Denied Topics: False Positive Testing Guide

November 17, 2026

Two nearly identical cream cards on either side of a brass boundary line on a linen desk, one stopped by a black block.

Amazon Bedrock Denied Topics: Reduce False Refusals Without Weakening Controls

By Erin Harris, AI Strategy & Business Transformation Consultant

To reduce false refusals in Amazon Bedrock Denied Topics, define the prohibited behavior narrowly, specify acceptable adjacent requests, and test both categories before deployment. A semantic guardrail should enforce an explicit business boundary, not block every request containing a sensitive subject.

AWS published eight best practices for Denied Topics on November 16, 2026. Its guidance explains that semantic category descriptions can overblock when they are broad or ambiguous. The publication is an official implementation resource, not an independently validated guarantee of safety. Confirm current service availability, supported models, pricing and region-specific requirements before rollout. Source: https://aws.amazon.com/blogs/machine-learning/best-practices-for-writing-denied-topics-in-amazon-bedrock-guardrails/

What goes wrong with broad topic rules

Suppose a company tells an assistant not to provide medical advice. An overbroad rule might block a request to summarize the organization's employee sick-leave policy. Likewise, a rule about legal advice could inadvertently prevent staff from organizing documents for outside counsel. In each case, the topic is sensitive, but the actual task may be appropriate.

False positives create measurable business friction: unresolved requests, longer handling time, more escalations and lower adoption. False negatives create different risks by allowing disallowed responses. Both need to be measured. There is no reason to assume that making a rule broader always makes the system safer in practice.

A five-step evaluation framework

Step 1: Write the policy objective. State the behavior the assistant must not perform, the users affected and the business reason. Avoid broad category labels without examples.

Step 2: Define neighboring allowed behavior. For each prohibited category, write examples of legitimate summarization, routing, research or documentation tasks that touch the same subject.

Step 3: Build paired tests. Use a set of prompts that clearly violate the policy and another set that should pass. Include paraphrases, indirect requests and ordinary workplace language. Label expected outcomes before running the tests.

Step 4: Measure both error types. Record the number of prohibited requests that pass and legitimate requests that are refused. Review severity, not just percentages. A single severe miss may outweigh many minor inconveniences.

Step 5: Version and monitor. Assign a policy owner, record changes, rerun regression tests after updates and establish an appeal or escalation path for blocked legitimate work.

Worked hypothetical example

A property management company wants its AI assistant to avoid giving individualized legal advice about eviction proceedings. A prohibited test might ask the assistant to instruct a manager on circumventing tenant protections. An allowed test might ask for a neutral summary of a published housing policy or for a checklist of records to send to licensed counsel. The team evaluates whether the guardrail differentiates those tasks. The example is hypothetical and is not an AWS test result.

What a guardrail cannot do by itself

A denied-topic rule does not authenticate a user, determine which customer files they may access, or guarantee that sensitive data cannot leave the organization. Those protections require independent authorization, data governance and monitoring. A guardrail also cannot settle every ambiguous policy question. Organizations should establish a human escalation path for requests with material legal, safety or customer consequences.

If an assistant is used in customer service, measure not only blocked content but also task completion and employee workarounds. A system that refuses legitimate tasks may need a better policy, a more constrained workflow or a different division of responsibilities between software and people.

Key Takeaways

  • Write rules around prohibited actions, not vague subject labels.
  • Test allowed neighboring requests alongside prohibited ones.
  • Track false refusals and misses, including severity.
  • Keep policy versions and an accountable reviewer.
  • Do not confuse content guardrails with access controls.

Next step

EH Synergy's AI Opportunity & Transformation Assessment can help identify the appropriate boundaries for an AI workflow, the tests needed before deployment and the operational owner of each control. Inquire through EH Synergy's established contact channel.

blog author avatar

Erin Harris

Founder of EH Synergy. M.Ed. in Educational Technology & Instructional Design and MBA in Marketing. Erin helps everyday people and small businesses use AI and automation to build real, useful tools.

Back to Blog
Erin Harris, Founder of EH Synergy

About the Author

Erin Harris, Founder of EH Synergy

M.Ed. in Educational Technology & Instructional Design and MBA in Marketing. Erin writes about how businesses can use AI and automation in ways that actually work.

Book a Call →

New to AI? Start with the free course.

The Clear Build Starter: five short lessons to build your first planner or tracker with AI. No coding needed.

Start the Free Course
EH Synergy

EH Synergy LLC · Dallas, Texas

Book a Call
© 2026 EH Synergy LLC