Industry Thought

AI Powered Community Libraries for Emergency Management Sound Great — Until You Understand the Data Lifecycle

Part of building an AI company in the preparedness space means constantly watching how emerging tools apply to emergency management. There has always been an interest in…

Part of building an AI company in the preparedness space means constantly watching how emerging tools apply to emergency management. There has always been an interest in community libraries—shared repositories where agencies can upload, share, and find plans, SOPs, and after-action reports. Recently, however, AI has opened the opportunity to "chat" with and unlock this crowdsourced collective knowledge. It's a compelling vision, but it's also where significant data governance risks are currently hiding.

Full disclosure: Preppr builds AI-powered preparedness tools. However, we do not co-mingle data across customers. Your sensitive data is never shared with other users or used to train models. We don't offer a shared community library, and for the reasons I'm about to lay out, that is a deliberate choice at this time.

While most know me today as the CEO of Preppr, for much of my career I focused specifically on the intersection of practice, policy, and emerging risk—working from the military to local community preparedness and national-level policy circles. I don't often go back into full "consultant mode" or do long-form writing these days, but the structural issues emerging in our sector require it.

Some may dismiss my positioning as competitive bias. You should be skeptical; I have skin in the game. But I also have an obligation to help our community adopt trustworthy AI and understand how to scrutinize technology vendors. Sometimes that means highlighting architectural risks that aren't immediately obvious. Rather than hiding that bias, I am flagging it upfront so you can read this analysis with full context. I'm posting this on LinkedIn to facilitate an open industry discussion, not a sales pitch. I am intentionally not naming any specific platform. The data governance risks I'm describing apply to any vendor offering AI-powered community libraries in this space.

The Emerging Governance Gap

Recently I've been tracking a new capability gaining traction in the field: community libraries where users contribute documents—plans, SOPs, AARs—to a shared, platform-wide repository. Other users can browse those documents, chat with them using AI, and run automated comparisons. To be clear, this isn't a fringe concept—it has been highlighted at major industry conferences and endorsed by leading associations. That endorsement carries weight in this community and signals a real hunger for shared knowledge.

However, the questions I'm raising in this post often don't get surfaced in quick product demos or conference announcements—they require a level of technical due diligence that our community hasn't historically had to apply to its tools.

In recent demonstrations of this capability, it is often mentioned that users already span government, healthcare, and the private sector. This implies a cross-sector mix of data, which makes the governance questions significantly harder.

I like the concept of an AI-powered community library. But knowing how AI technology works, I have questions about the execution. When evaluating these tools, the most critical question to ask is: what happens if someone contributes a document and later revokes permission?

The standard answer is straightforward: the document would be deleted. Okay. Sounds great on the surface. But what about all the information already derived from it — the summaries, the crosswalk outputs, the chat responses that pulled from it and now live in other users' histories?

This is where the answers usually stop. It points to a data governance complexity that this community needs to understand before we start feeding operationally sensitive documents into shared AI systems. This isn't just a technical challenge; it's a structural one. And we have a long history of underestimating the structural challenges of sharing emergency management data.

We've Been Trying to Build This for Twenty Years

The desire for a shared knowledge base in our field isn't new. The federal government has tried — and largely failed — to build one for decades.

FEMA's Lessons Learned Information Sharing System (LLIS) was supposed to be exactly this: a national network for sharing best practices, after-action reports, and lessons learned across the preparedness community. It launched after 9/11, accumulated nearly 23,000 resources over more than a decade, and was used by verified emergency response providers and homeland security officials at every level of government. Then, in 2015, FEMA shut it down, consolidating its contents into the Naval Postgraduate School's Homeland Security Digital Library. The community largely moved on without it.

The Homeland Security Information Network (HSIN) had an even rougher run. Intended as DHS's primary nationwide information sharing tool, it was rushed to deployment under the pressure of the post-9/11 threat environment. By 2007, GAO was testifying before Congress that HSIN had landed on OMB's Management Watch List and High-Risk List — officially labeled a poorly planned and poorly performing project. GAO found nearly a dozen overlapping homeland security networks within DHS costing over $300 million annually, with poor coordination between them. The system had been built around law enforcement requirements without sufficiently identifying the needs of other HSIN user communities, such as emergency management personnel. HSIN still exists today, but those early failures set a pattern that shaped how the EM community thinks about shared information systems: cautiously.

So when a commercial platform offers a community library with AI on top — the ability to chat with shared documents using AI, not just browse them — I understand the appeal. These platforms are stepping into a space the government never figured out how to fill. And the AI layer is genuinely compelling.

But cross-user data co-mingling — documents, outputs, and AI-generated insights flowing between contributors and consumers — is one of those ideas that sounds great on the surface. What's better than sharing with your colleagues automatically? The problem is that it faces significant challenges under scrutiny by anyone with a solid understanding of both the technology and the use case. AI fundamentally changes the data governance equation in ways that LLIS and HSIN never had to worry about.

What Actually Happens When You Contribute a Document to an AI System

When you upload a document to a traditional shared library, it sits in a folder. Someone opens it, reads it, or downloads a copy. Simple. If you remove it, the source is gone.

AI-powered platforms work differently.

There's a preliminary question worth addressing: what screening happens before a document enters the system at all? Some platforms rely on manual review by internal teams. That's a start. But checking for PII, HIPAA concerns, and content that shouldn't be shared publicly is complex. Documents contributed from different states carry different legal protections, open records obligations, liability exposures, and regulatory constraints. A hospital surge plan from California raises different legal questions than one from Texas. A county evacuation plan may be subject to public records laws in one state and exempt in another. Unless legal experts familiar with each contributing sector are reviewing every document, manual screening—however well-intentioned—may miss the specific legal risks that matter most. Furthermore, manual review is difficult to scale without eventually relying on automation, which introduces its own accuracy questions.

Even assuming the intake screening is solid, here's what happens next. In a typical RAG (retrieval-augmented generation) architecture, when a document enters the system it gets broken into chunks. Each chunk is converted into a mathematical representation called a vector embedding and stored in a specialized database. When a user asks a question, the system searches those embeddings to find the most relevant content, feeds it to a large language model, and generates a response. That response—the summary, the crosswalk output, the answer to a chat question—gets stored in the requesting user's conversation history.

Now imagine dozens of users chatting with a contributed document over the course of several months. Running crosswalks. Generating summaries. Each interaction produces AI outputs that weave in content from the original document. Those outputs live in other users' accounts.

When the contributor revokes permission, the platform can delete the source file and the embeddings. That stops future retrieval. But standard deletion processes rarely reach back into every conversation, every crosswalk report, and every exported summary where the document's content was already synthesized. The source is gone. But unless the platform has built highly specific mechanisms to trace and handle derived works — which is a hard engineering problem that most haven't solved — traces of that intelligence persist indefinitely.

This is the difference between pulling a book off a library shelf and erasing everything anyone learned from reading it. One is a file operation. The other requires tracking provenance and tracing content from those documents to every single use throughout the platform. It is a significant engineering and governance challenge that requires deliberate architectural choices.

This Problem Is Already Creating Real Consequences

Before I get into examples, an important technical distinction. There are three different ways your data can end up living inside an AI system.

Pre-training: the process of building a foundational model like GPT-5 or Claude from scratch using massive datasets. The information doesn't sit in the model as retrievable files; it's encoded as patterns distributed across the model's weights. Deleting the source data after training doesn't change the model's behavior because the learning has already happened.

Fine-tuning: an existing foundational model is adjusted on a smaller, domain-specific dataset to improve its performance in a particular area. Fine-tuning also modifies the model's internal weights, so it carries a similar baked-in problem — just at a smaller scale.

Retrieval-augmented generation (RAG): in a RAG system, the AI model itself isn't retrained on your documents. Instead, your documents are chunked, embedded, stored externally, and made available to the AI model. When a user asks a question, the system retrieves relevant chunks and feeds them to the model as context for generating a response.

To be fair, terminology varies. I have heard the term "tuning" used colloquially in marketing for this type of feature. If a platform is actually fine-tuning models on contributed content, the data is baked into the model. However, it is more likely that "tuning" is being used to describe RAG customization. Even with RAG, the derived content problem remains. The model didn't learn your content permanently, but it read your content, synthesized it, and delivered it to other users in a form that's now stored in their history.

With that distinction in mind, each of the architectures have already produced real-world consequences.

Training-based: The OpenAI Precedent

In March 2023, Italy's data protection authority temporarily banned ChatGPT from processing Italian users' data. The core issue: OpenAI had used personal information to train its models — baking it into the model weights — without adequate legal basis, and there was no meaningful mechanism for individuals to request erasure once their data had shaped the model's behavior. The ban was lifted within a month after OpenAI took corrective steps. But the investigation continued for nearly two years. In December 2024, Italy fined OpenAI €15 million (roughly $16 million), finding that the company had processed personal information of users to train ChatGPT without having an adequate legal basis and failed transparency obligations. The fine was notable, but the underlying principle was more important: regulators are starting to treat AI-generated outputs — not just source data — as part of the data governance equation.

RAG-based: The Problems Closest to Home

The examples above involve data permanently encoded into model weights. But RAG-specific risks are emerging too — and they're closer to what a community library platform would actually be doing.

Microsoft Copilot and the oversharing problem. Microsoft 365 Copilot uses RAG to pull from an organization's SharePoint, OneDrive, Teams, and Outlook data to generate responses. Researchers at the University of Texas discovered a class of vulnerabilities they named ConfusedPilot, demonstrating that Copilot's RAG pipeline could be manipulated into surfacing sensitive information across permission boundaries. Separately, security firm Lasso found that Copilot could expose data from Fortune 500 companies due to overly permissive retrieval scopes — the same RAG architecture retrieving content users shouldn't have had access to and delivering it in AI-generated responses.

Enterprise RAG pipelines leaking sensitive content. Security researchers have documented cases where enterprise RAG systems surfaced API keys buried in embedded markdown files, slides from non-public board decks retrieved through routine queries, and confidential contract details — including pricing — served up because the RAG index was too broadly scoped. This isn't about malicious actors. It's about what happens when a RAG system in a shared environment does exactly what it's designed to do — retrieve and synthesize content across users — without adequate governance. Who controls what gets retrieved? Who sees the output? And what happens to that output after it's generated?

Samsung: when contributed content becomes irrecoverable. In one of the most widely reported AI data incidents, Samsung engineers pasted proprietary semiconductor source code and confidential meeting notes into ChatGPT to get help debugging and optimizing their work. Because ChatGPT retained user inputs for training, Samsung's trade secrets were now in OpenAI's system — and Samsung had no mechanism to retrieve or delete them. The difference is that Samsung employees did it accidentally. A community library invites users to do it by design.

The Emerging Regulatory Framework

The direction is clear across multiple jurisdictions. The OECD published guidance acknowledging that the right to erasure might be invoked in relation to LLMs, and that removing personal data from training datasets does not directly affect the already-trained models. Germany's Conference of Data Protection Authorities published guidance specifically addressing RAG (retrieval-augmented generation) — the exact architecture that a community library with document chat and crosswalk features would use — and its data protection implications. The EU AI Act, which began phased enforcement in August 2025, requires data governance documentation across the AI system lifecycle, including how data flows through retrieval pipelines.

Closer to home, organizations building emerging community library tools report having users across government, hospitals, and the private sector. That mix exposes a U.S.-specific governance gap that needs to be addressed. The federal government recognized years ago that cross-sector sharing of sensitive critical infrastructure information required a dedicated legal structure. That structure was the Critical Infrastructure Partnership Advisory Council (CIPAC), established by the Department of Homeland Security with a targeted exemption from the Federal Advisory Committee Act (FACA). It allowed government and private sector partners to exchange threat intelligence and infrastructure information candidly without triggering public disclosure requirements. Building that trust architecture took nearly two decades. Yet CIPAC was terminated by DHS on March 7, 2025, and the Cybersecurity Information Sharing Act—which helped underpin related information sharing protections—was set to expire later that year.

Layered onto this is the intersection of state sunshine laws, public records obligations, and records retention requirements with AI-generated derivative content inside commercial platforms. That terrain remains largely untested. If a state or local agency uploads documents into a commercial AI-powered community library, do those materials remain subject to open records laws? More pointedly, do AI-generated outputs that synthesize government-contributed information themselves become public records?

Why Emergency Management Should Care More Than Most

Most of the enforcement actions above involve personal data — names, faces, chat histories. Emergency management community libraries involve something different but equally sensitive: operational and sensitive data. A community library that co-mingles documents across all three sectors (government, healthcare, private) is navigating HIPAA, FOIA, trade secret protections, and sector-specific compliance requirements simultaneously.

Three scenarios worth thinking through:

The un-retractable vulnerability. A hospital system shares a hazard vulnerability analysis (HVA) identifying a critical gap in their backup power systems. Another user—perhaps a competitor or a regulator—asks the AI to summarize regional infrastructure risks. The AI generates a report explicitly detailing that hospital's power gap. The hospital later revokes the document. The source file vanishes. But the report identifying their vulnerability remains in the other user's account—potentially discoverable, potentially public, and now completely outside the contributor's control.

The zombie evacuation route. A county emergency manager contributes an evacuation plan. Six months later, a bridge on a primary route is condemned, and they upload a revised plan, deleting the old one. But users in neighboring jurisdictions have already asked the AI to list evacuation routes for County X. Those AI-generated lists—referencing the now-dangerous route—persist in their chat histories. During a crisis, a mutual aid partner might act on that zombie intelligence, believing it to be current because it came from the trusted community platform.

The governance death spiral. This is the macro risk. Once it becomes clear that deleting a document doesn't scrub the intelligence derived from it, the incentive structure flips. General counsels will block participation. Agencies will stop sharing operational plans and start sharing only sanitized, useless brochures. The community library concept dies—not because the tech failed, but because the governance model couldn't guarantee the one thing contributors need most: control.

The Questions That Need to Be Asked — and Answered

Before you contribute documents or subscribe to AI-powered community libraries, ask these questions. These are the minimum due diligence.

Before you contribute:

  • What exactly am I authorizing when I upload a document to a community library? Does my content become training data, retrieval context, or both?
  • Can I see which users or features have accessed my contributed documents?
  • Does the platform's terms of service explicitly address what happens to derived content on revocation?

When I revoke permission:

  • What specifically gets deleted — the file, the vector embeddings, or both?
  • What happens to AI-generated outputs that already incorporated my content before revocation?
  • Are other users notified that source material underlying their saved outputs has been withdrawn?

On derived content governance:

  • Does the platform distinguish between source documents and content derived from them?
  • Is there any mechanism to flag, expire, or invalidate AI outputs whose source material has been revoked?
  • How does the platform handle the risk that outdated information persists in derived form long after the source is updated or removed?

Building This Right

I want to be clear: I'm not arguing against AI-powered community libraries. The need is real. Two decades of failed federal attempts to build shared preparedness knowledge infrastructure prove that. LLIS accumulated 23,000 resources and still couldn't sustain itself. HSIN cost hundreds of millions and still wasn't built for the people who needed it most. The appetite for this capability is enormous and legitimate.

AI-powered community libraries have the potential to finally deliver what those systems couldn't. But they also introduce a category of risk those earlier systems never faced. A shared folder of PDFs is straightforward to govern. An AI system that ingests, synthesizes, and redistributes contributed content across thousands of user interactions is a fundamentally different thing.

The platforms building these features owe their users — particularly users in a community where operational security matters and trust is the currency that makes mutual aid work — an honest accounting of the full content lifecycle. Not just the upload and the deletion, but everything that happens in between. Including the parts that are hard to undo.

Sources

  • FEMA Lessons Learned Information Sharing consolidation. In spring 2015, the LLIS.gov website ceased independent operations and delivered nearly 23,000 resources to the Naval Postgraduate School's Homeland Security Digital Library (HSDL).
  • GAO testimony on HSIN. David Powner, Director of IT Management Issues, testified that HSIN was on OMB's Management Watch List and High-Risk List, with nearly a dozen overlapping DHS networks costing $300M annually (GAO-07-822T).
  • Italy fines OpenAI €15 million. The Garante found OpenAI processed personal data without adequate legal basis and failed transparency obligations.
  • EU-wide ChatGPT investigations. The European Data Protection Board launched a dedicated ChatGPT task force in April 2023; DPAs in Germany, France, Spain, and other member states initiated preliminary investigations.
  • Dutch DPA fines Clearview AI €30.5 million; Italian DPA fines Clearview AI €20 million; UK ICO Clearview AI enforcement (original £7.5M fine issued May 2022; Upper Tribunal reversed in October 2025 ruling UK GDPR applies regardless of company location); Canada's Privacy Commissioner Daniel Therrien: "What Clearview does is mass surveillance and it is illegal."
  • Microsoft Copilot RAG vulnerabilities. Roychowdhury et al., "ConfusedPilot: Compromising Enterprise Information Integrity and Confidentiality with Copilot for Microsoft 365," University of Texas at Austin SPARK Research Lab, August 2024; Lasso Security Fortune 500 data-exposure findings; US Congress banned Copilot for staff.
  • Enterprise RAG pipeline data leakage incidents. Security red-team findings including API key exposure, board deck retrieval, and contract details surfaced through routine queries. Zeng et al., "The Good and the Bad: Exploring Privacy Issues in Retrieval-Augmented Generation (RAG)," Findings of the ACL 2024.
  • Samsung ChatGPT data leak. Engineers pasted proprietary source code and meeting notes into ChatGPT; Samsung banned the tool, stating the data was irrecoverable.
  • Clearview AI BIPA class action settlement (In re Clearview AI, Inc. Consumer Privacy Litigation, No. 1:21-cv-00135, N.D. Ill.).
  • FTC civil investigative demand to OpenAI, July 2023.
  • OECD guidance on data access requests in AI-driven processing — notes that erasure rights may apply to LLMs and that removing data from training sets does not directly affect already-trained models.
  • German Data Protection Conference publishes RAG-specific guidance, 2025.
  • EU AI Act data governance requirements. Article 10 requires training, validation, and testing datasets to be relevant, sufficiently representative, and free of errors; phased enforcement began August 2025.
  • TechPolicy.press on the right to be forgotten in AI. GDPR Article 17 grants individuals the right to request data erasure but does not define erasure in the context of AI.
  • CIPAC termination and CISA 2015 expiration. DHS terminated CIPAC in March 2025; the Cybersecurity Information Sharing Act of 2015 was set to expire September 30, 2025.