Responsible AI Practices

Overview

Preppr is an AI-powered platform for disaster preparedness that generates exercise scenarios, situation manuals, intelligence reports, live exercise synthesis, and collaborative preparedness assessments. Because our outputs inform real-world emergency planning decisions, we treat AI safety, accuracy, and bias mitigation as core product requirements—not afterthoughts.

This document describes the technical and procedural measures Preppr employs to detect, control, and mitigate bias and inaccuracy across the platform.

Multi-Model Architecture

Avoiding single-source dependency in AI outputs.

Preppr uses models from multiple independent AI providers—including OpenAI (GPT-4o, GPT-4.1), Anthropic (Claude Sonnet 4), and Google (Gemini 2.5 Flash and Pro)—across different platform features. This multi-model strategy is a deliberate architectural decision that provides several bias mitigation benefits:

Diverse training data and alignment approaches. Each model family is trained on different data corpora using different alignment methodologies. Relying on multiple providers reduces the risk that a single provider’s training biases systematically influence Preppr’s outputs.

Cross-validation capability. Where Preppr’s workflows involve multi-step AI processing, outputs from one model can be evaluated or refined by another, reducing the likelihood of compounding errors from a single model’s blind spots.

Provider flexibility. Preppr’s architecture is not locked to any single AI provider. If a model demonstrates persistent quality or bias issues, Preppr can route workflows to alternative models without architectural changes.

Inherited Model Provider Bias Controls

Preppr benefits from the bias mitigation investments of each AI provider.

Each of Preppr’s AI model providers maintains its own bias detection and mitigation programs. By using these models via API, Preppr inherits the safety and fairness controls built into each provider’s platform:

OpenAI (GPT-4o, GPT-4.1, GPT-4.1 Mini). OpenAI implements safety mitigations including reinforcement learning from human feedback (RLHF), red-teaming, and content filtering systems designed to reduce harmful, biased, or discriminatory outputs. OpenAI publishes model cards and system safety documentation for each model release and maintains an active safety research program.

Anthropic (Claude Sonnet 4). Anthropic’s Constitutional AI approach trains models to be helpful, harmless, and honest through a combination of human feedback and AI-assisted evaluation. Anthropic conducts bias evaluations across demographic categories and publishes responsible scaling policies that govern model development and deployment.

Google (Gemini 2.5 Flash, Pro). Google’s Responsible AI practices include fairness testing, adversarial evaluation, and model governance processes. Gemini models are evaluated against Google’s AI Principles, which include commitments to avoiding the creation or reinforcement of unfair bias.

AskNews (Preppr Intelligence). AskNews is purpose-built to reduce bias in open-source news intelligence. It processes hundreds of thousands of articles daily across multiple languages, applying AI-powered bias detection at the article level, surfacing contradictions between sources, and ensuring diversity of perspectives. Preppr Intelligence inherits this bias-reduced data foundation for all intelligence and research features.

These provider-level controls serve as a foundational layer of bias mitigation that Preppr builds upon with its own platform-level safeguards. As providers update and improve their safety measures, Preppr automatically benefits from those improvements.

Groundedness in User-Provided Context

AI outputs are anchored to real organizational data, not open-ended generation.

Preppr’s AI features are designed to generate outputs grounded in user-provided context rather than relying on open-ended generation. Across the platform:

Document-grounded analysis. Ask Preppr and document analysis features extract and synthesize information from user-uploaded plans, policies, and procedures. AI responses are anchored to the content of these documents, reducing the surface area for hallucination or fabrication.

Scenario-specific generation. The Exercise Designer generates exercise materials based on user-defined parameters including hazard type, jurisdiction, participating organizations, and objectives. Outputs are constrained by these inputs rather than generated from unconstrained prompts.

Exercise-grounded synthesis. Preppr Exercise synthesizes findings and report-outs from the actual discussions and responses occurring during a live exercise, grounding its outputs in real participant input rather than generating analysis from general knowledge.

Intelligence grounded in bias-reduced news data. Preppr Intelligence is powered by AskNews, a news intelligence platform specifically designed to reduce bias in open-source information. AskNews processes hundreds of thousands of articles daily across multiple languages, using AI to extract facts, analyze bias, and identify contradictions between sources. By surfacing diverse perspectives, attributing all information to original sources, and applying algorithmic bias detection at the article level, AskNews provides Preppr with a structurally de-biased intelligence foundation. AI analysis is performed against this verified, source-attributed material rather than model memory alone.

Contributor-informed synthesis. Preppr Collaborate gathers collective intelligence from real organizational stakeholders through guided AI interviews. Findings are drawn from actual contributor input—not generated from general knowledge—then aggregated and anonymized before being presented to the campaign manager.

This grounding approach means that Preppr’s AI outputs reflect the user’s actual organizational context, reducing the risk of generic, biased, or irrelevant content.

Human-in-the-Loop Controls

Qualified professionals review and approve all AI-generated deliverables.

Across Preppr’s product suite, AI-generated content passes through human review before it becomes a final deliverable:

Exercise Designer. AI-generated scenario narratives, injects, and situation manual content are presented to the user for review, editing, and explicit approval at each stage of the design workflow. Users can modify, reject, or regenerate any AI output before it is incorporated into the final exercise.

Preppr Exercise. During a live exercise, Preppr synthesizes findings and generates report-outs in real time. A human Facilitator and Controller reviews, edits, and approves all AI-synthesized content before it is presented to exercise participants. No AI-generated synthesis reaches participants without explicit human approval.

Ask Preppr. AI-generated answers and document analysis are presented as responses for the user to evaluate. Users determine whether and how to apply these insights to their work.

Preppr Intelligence. AI-synthesized intelligence reports, built on AskNews’s bias-reduced news data, are delivered to users for professional review and interpretation. Preppr does not take autonomous action based on intelligence findings.

Preppr Collaborate. A subscribing user (the campaign manager) defines the scope of work with Preppr, then invites contributors to be interviewed. Preppr conducts those interviews autonomously. It then generates discussion guides for each contributor, aggregates the input, removes PII, and draws findings that are presented back to the campaign manager for review. The campaign manager decides how to use those findings—including, if they choose, as input to exercise design in a separate product. See Autonomous Workflow Safeguards below for the additional controls that apply to the contributor interview and synthesis process.

This design ensures that emergency management professionals—not AI models—are the final decision-makers for all deliverables produced through Preppr.

Autonomous Workflow Safeguards

Where AI operates with greater autonomy, additional controls ensure integrity and auditability.

Preppr Collaborate includes workflows where AI operates without real-time human oversight. Contributors are interviewed directly by Preppr’s AI without the campaign manager present, and the synthesis process—aggregating contributor input, removing PII, and drawing findings—runs autonomously. For these workflows, Preppr implements additional safeguards:

Comprehensive logging. All AI-generated interview prompts, contributor responses, discussion guides, and synthesis outputs are logged throughout the Collaborate workflow.

Neutral third-party audit trail. Preppr partners with a neutral third party to maintain encrypted, tamper-evident logs of autonomous AI interactions. These logs provide an independent record that can be used for auditing and to demonstrate that AI-generated content has not been altered after the fact.

PII anonymization. During synthesis, contributor data is processed through Microsoft Presidio to detect and redact personally identifiable information (persons, organizations, locations) before AI analysis. This ensures findings are drawn from the substance of contributor input, not identifying details.

Campaign manager review. The autonomous interview and synthesis process produces findings that are presented to the campaign manager for review. The campaign manager determines how to interpret and apply these findings. Autonomous AI processing produces analysis for human evaluation, not final deliverables.

Output Quality Control

Systematic measures to monitor and maintain AI output quality.

Prompt Engineering for Bias Mitigation

Deliberate prompting techniques reduce bias at the point of generation.

Beyond architectural and provider-level controls, Preppr employs specific prompt engineering techniques designed to minimize bias in AI-generated outputs:

Specificity and nuance. Preppr’s prompts are detailed, neutral, and objective rather than broad or open-ended. Specific prompts reduce the likelihood that models fall back on biased default responses or stereotypical assumptions.

Requesting diversity of perspectives. Where appropriate, prompts explicitly request multiple viewpoints, perspectives, or scenarios. This is particularly important in exercise design and intelligence analysis, where a single-perspective output could lead to incomplete preparedness planning.

Challenging assumptions. Preppr’s multi-step processing chains include prompts that challenge initial outputs, requesting alternative viewpoints or counter-arguments to surface blind spots in AI-generated analysis.

Defining demographic context. When dealing with scenarios that involve sensitive topics—such as community impact assessments, public health scenarios, or resource allocation—prompts define specific demographic context to avoid generalized or biased assumptions about affected populations.

System-level behavioral constraints. Preppr utilizes the system prompt and instructional layer of each AI model to define behavioral constraints regarding fairness, objectivity, and professional standards. These system-level instructions establish baseline expectations for every AI interaction across the platform.

Domain-specific constraints. All prompts are designed specifically for emergency management use cases. Outputs are constrained to relevant, professional content, minimizing the risk of inappropriate, biased, or off-topic generation.

Structured output formats. Where possible, Preppr’s AI workflows produce structured outputs (e.g., exercise injects with defined fields, situation manual sections with standardized formats, exercise report-outs with consistent structure) rather than freeform text. Structured formats reduce variability and make quality issues easier to identify.

Multi-step processing chains. Complex outputs such as exercise designs and live exercise synthesis are produced through multi-step AI processing chains where each step builds on and can correct the outputs of previous steps, rather than relying on a single generation pass.

Model updates and monitoring. Preppr monitors AI provider model updates and evaluates their impact on output quality. Model versions and providers are documented transparently on the Technical Specifications page. Preppr’s SOC 2 Type II continuous monitoring controls extend to AI system performance.

Organizational Controls

Security and compliance infrastructure that supports responsible AI use.

SOC 2 Type II compliance. Preppr’s SOC 2 Type II examination (June–September 2025) resulted in an unqualified opinion with no exceptions on Security trust services criteria. This includes controls over system monitoring, access management, change management, and risk assessment that apply to AI systems.

California State AI Safety Review authorization. Preppr holds a California State AI Safety Review authorization, enabling all California state agencies to procure Preppr without additional AI safety review. This authorization reflects an independent assessment of Preppr’s AI safety posture by a state regulatory authority.

US-based infrastructure and personnel. All data is stored on US-based AWS infrastructure. Access to user data is restricted to background-checked, US-based personnel on a strict need-to-know basis.

No training on customer data. None of Preppr’s AI providers use customer data for model training by default. This ensures that organizational data submitted to Preppr does not influence future AI model behavior for other users. Data retention by AI providers is limited to 30–55 days for abuse monitoring purposes only.

Continuous Improvement

Responsible AI is not a static achievement. Preppr is committed to evolving its practices as AI capabilities, risks, and regulatory expectations develop. This includes monitoring emerging AI governance frameworks, evaluating new bias detection tools and techniques, and incorporating feedback from the emergency management professionals who use the platform daily.

For questions about Preppr’s responsible AI practices, contact connect@preppr.ai.