The Most Capable AI in the World Went Dark in 72 Hours. The Questions It Leaves Behind Are Yours.

On June 9, one of the largest AI companies in the country released what it described as its most capable publicly available model.

By June 12, that same model was offline.

According to the company’s own statements and public reporting, Anthropic released Claude Fable 5 as a model designed to operate for extended periods across software development, research, analysis, and complex problem-solving. Three days later, the company said it had received a U.S. government export-control directive citing national security concerns. The directive required Anthropic to suspend access to Fable 5 and a related model, Mythos 5, for users across the platform while the issue was reviewed.

The reported concern was that someone had found a way around the model’s safety limits and pushed it toward information the model was designed to refuse. Anthropic has said it believes the directive was based on a misunderstanding and is working to restore access.

The technology world will debate the details. Whether the government response was appropriate. Whether the model was truly unsafe. Whether the company moved too quickly. Whether the decision was about national security, politics, competition, or all of the above.

For public safety leaders, the more useful question is closer to home.

What happens when a powerful tool becomes available quickly, starts being used before the rules are fully understood, and then changes, fails, or disappears faster than the organization can respond?

That is the part of this story that should get the attention of police chiefs, command staff, municipal managers, fire and EMS leaders, accreditation managers, and anyone responsible for agency operations.

Because in public safety, AI is not some distant future issue waiting to arrive through a formal rollout. It is already here. In some agencies, it is already being used in the gray.

The greatest AI risk for public safety may not be the tool everyone is talking about. It may be the one already being used quietly, informally, and without a policy.

The Gray Area Is Already Here

Most agencies would not describe themselves as “using AI.” That is understandable. No one may have purchased an AI platform. No one may have asked the governing body to approve an AI contract. No one may have issued a special order announcing artificial intelligence as an agency initiative.

But that does not mean AI is absent.

An officer may be using a consumer chatbot to clean up the wording of a report narrative before submitting it. A detective may be using a tool to summarize a lengthy document, organize interview notes, or prepare a rough outline for a search warrant. A supervisor may be using AI to draft a performance evaluation, disciplinary summary, press statement, training memo, or email to a resident. An accreditation manager may be using it to reword proofs, draft policy language, or summarize a standard. A command officer may be using it to help prepare talking points, grant narratives, staffing justifications, or public-facing communications.

None of those uses necessarily begin with bad intent. In many cases, the employee is trying to save time, improve clarity, or keep up with a workload that has outgrown the staffing available to handle it. That is exactly why the issue is difficult. AI is useful enough that people will find it. It is accessible enough that they can use it without asking permission. And it is familiar enough that many employees may not fully recognize when they have crossed from basic editing support into something that affects official agency work.

That is the gray area.

The concern is not that every use of AI is automatically improper. The concern is that some uses may involve case facts, criminal justice information, personnel matters, internal deliberations, protected records, evidence summaries, victim information, investigative strategies, or official reports. If those details are entered into a public tool, stored by a third party, used to train a model, or later questioned in court, the agency may be forced to explain a practice it never approved, never trained on, and may not have known existed.

That is not a technology problem. That is an awareness problem.

Chiefs May Not Know What Their Own Agency Is Already Doing

One of the most uncomfortable parts of this issue is that a chief may honestly believe the agency does not use AI and still be wrong.

That is not a criticism of chiefs. It is a reflection of how these tools enter the workplace. They do not always come through procurement, policy review, municipal approval, or command staff discussion. They often arrive through individual use. An officer downloads an app. A supervisor tries a chatbot. A records employee uses an online summarizer. A vendor adds a feature. A platform changes its capabilities after an update. By the time leadership becomes aware of it, the practice may already feel routine to the people using it.

That creates a gap between official agency policy and actual agency practice.

In public safety, that gap matters. If AI touches a report, a public statement, a disciplinary document, a personnel evaluation, an investigative summary, a training record, or an accreditation file, someone will eventually need to stand behind the final product. If the chief, solicitor, prosecutor, municipal manager, or accrediting body asks how the work was created, “we did not know” is not a strong position.

The first step is not panic. It is curiosity. Chiefs and command staff need enough visibility to know whether AI is being used, where it is being used, what information is being entered, and whether the use affects official records or decision-making.

Before an agency can decide what to allow, it has to understand what is already happening.

Before an agency can decide what to allow, it has to understand what is already happening.

The Vendor Issue May Be Even Harder

Informal employee use is only one side of the problem. The other side may be harder for leaders to see.

Many agencies are already using software that includes AI-enabled or AI-assisted features. The product may not be marketed primarily as artificial intelligence. It may be a records management system, body-worn camera platform, redaction tool, evidence management system, scheduling product, training platform, transcription service, analytics dashboard, wellness application, recruiting tool, policy system, or grant-writing support tool.

The agency may have purchased the software years ago for a completely different purpose. Then, over time, the vendor adds new functionality. Automated transcription. Facial detection. Object recognition. Sentiment analysis. Report drafting. Smart search. Predictive analytics. Auto-generated summaries. AI-assisted quality control. Suggested classifications. Automated redaction. Workflow recommendations.

From the vendor’s perspective, these features may be improvements. From the agency’s perspective, they may introduce questions that were never asked during the original purchase.

What data does the system process? Where is that data stored? Is agency information used to improve the vendor’s model? Can the agency disable the feature? Is there an audit trail? Can the agency explain how an output was generated? Does the tool merely organize information, or does it influence a decision? Does it touch evidence, reports, personnel records, training files, or public communications? Has anyone reviewed the contract terms, data retention practices, or disclosure implications?

These are not anti-technology questions. They are basic operational questions.

A chief does not need to become a software engineer. But chiefs and municipal leaders do need to understand enough about the systems they implement to know when those systems are doing more than storing information. If software begins analyzing, generating, recommending, prioritizing, summarizing, or drafting, the agency should know that. It should also know whether the feature is optional, whether staff are using it, and whether the output is being treated as official work.

The risk is not just that AI may be wrong. The risk is that AI may be invisible.

Invisible AI Is Hard To Defend

Public safety agencies are used to defending their decisions. Reports are challenged. Policies are reviewed. Training records are examined. Evidence handling is questioned. Personnel actions are appealed. Public statements are scrutinized. Accreditation assessors ask for proof. Defense attorneys ask how something was created, reviewed, retained, or disclosed.

AI adds another layer to those existing responsibilities.

If an officer uses AI to draft part of a report, does the agency require disclosure? If a supervisor uses AI to prepare an evaluation, is that acceptable? If an AI tool summarizes body-camera footage, who verifies the summary? If a redaction platform uses automated detection, who confirms the final redaction? If a system flags records for review, does a human understand why? If a vendor-generated summary is wrong, who catches it before it becomes part of the file?

The answer cannot simply be, “the software did it.”

Public safety work still requires human accountability. AI may assist with speed, organization, drafting, review, search, analysis, or formatting. But it should not blur who owns the final report, decision, policy, communication, disclosure, or operational action.

That is especially true when AI use is not obvious. If no one knows a tool was used, no one can review it properly. No one can disclose it. No one can train on it. No one can limit it. No one can defend it with confidence later.

That may be the most important lesson for agencies right now. The issue is not whether AI is good or bad. The issue is whether AI is visible enough to be managed.

The risk is not just that AI may be wrong. The risk is that AI may be invisible.

The Fable 5 Lesson Is About Control

The Fable 5 story is dramatic because of the speed. A tool was released, adopted by users, and then suspended within days. But the deeper lesson is not about that specific model. Most public safety agencies were not building critical workflows on Fable 5.

The lesson is about control.

Agencies are increasingly surrounded by tools they do not fully control, built by companies they do not oversee, updated on timelines they do not set, and governed by legal, political, commercial, and technical decisions made far outside the agency. That does not mean agencies should avoid modern technology. Public safety cannot operate effectively by refusing every new tool. But it does mean leaders should be careful about allowing unofficial dependencies to form around systems they do not understand.

If a transcription tool changes its terms, what happens to the records it processed? If an AI-assisted report-writing feature is later restricted by law, what happens to the reports already created? If a vendor removes a feature, can the agency still complete its workflow? If a court, prosecutor, or accrediting body asks about the use of AI, can the agency answer clearly?

Those questions become harder when AI use grows quietly.

This is where public safety leaders may need to shift their thinking. The issue is not only whether to approve a new AI product. It is whether the agency has enough internal awareness to recognize when AI has already become part of its operations.

Policy Matters, But Inventory Comes First

A policy is important. Training is important. Disclosure rules are important. Vendor review is important. But if an agency starts by writing a policy without first understanding how AI is already being used, the policy may miss the reality on the ground.

The more useful starting point may be an inventory.

What consumer AI tools are employees using, formally or informally? Are officers using AI for report narratives, grammar, summaries, warrant language, emails, or administrative writing? Are supervisors using AI for evaluations, discipline, scheduling, training, or internal memos? Are command staff using AI for public statements, council presentations, grants, policy drafts, or strategic planning? Are accreditation personnel using AI to prepare proofs, summaries, or policy crosswalks?

Then the agency needs to look at its systems.

Which vendors have AI-enabled features? Which features are turned on? Which ones are optional? What agency data do they touch? Are outputs retained? Are prompts or inputs stored? Can the agency audit use? Can the feature be disabled? Has the agency reviewed contract language or data practices? Is anyone responsible for approving new AI functionality before staff use it?

Only after that kind of review can a policy become practical. Otherwise, the agency risks creating a document that sounds good but does not match actual behavior.

The goal is not to create a policy that stops all innovation. The goal is to create enough clarity that employees know where the lines are, supervisors know what to review, vendors know what the agency expects, and leadership can explain the agency’s approach when asked.

A policy that does not match actual practice is not governance. It is paperwork.

The Standard Is Already Moving

This is not just a local management issue. The broader public safety environment is beginning to move.

CALEA has already begun addressing artificial intelligence in the accreditation conversation, including responsible use, quality control, and how AI systems are applied in policing. Some states are also beginning to move toward more specific requirements, particularly around AI-generated reports, disclosure, documentation, and retention.

That matters because public safety agencies should not assume the absence of a local mandate means the absence of risk. In many jurisdictions, the law may still be catching up. Accreditation standards may still be developing. Prosecutor expectations may not yet be fully defined. Vendor practices may be changing faster than agency policy. But the operational reality is already here.

Agencies still have an opportunity to think through the issue before they are forced to react to legislation, litigation, accreditation changes, prosecutor expectations, labor concerns, or a public controversy. That is a better posture than waiting for a rule, a lawsuit, an assessor, or a crisis to define the agency’s approach for them.

The agencies that wait for a specific mandate may eventually get one. But by then, AI use may already be embedded in officer habits, vendor systems, report-writing practices, evidence workflows, personnel processes, accreditation files, and administrative decision-making.

That is not the ideal time to begin asking basic questions.

A Conversation Chiefs Should Be Having Now

The Fable 5 story should not be treated as proof that AI is too dangerous for public safety. It should also not be dismissed as a technology-sector drama with no connection to local agencies. Its value is that it highlights a pattern public safety leaders should recognize: tools become available, people start using them, vendors build them into existing systems, and policy often catches up later.

That pattern is already familiar in public safety. Technology routinely moves faster than policy. The difference with AI is how quickly it can touch language, judgment, records, evidence, personnel matters, analysis, and public communication.

So the questions for chiefs and municipal leaders are not theoretical.

Are personnel already using AI tools in unofficial ways? Are supervisors aware of it? Are agency systems using AI-enabled features? Are those features approved, optional, or simply turned on by default? Does the agency know what data is being entered into public tools? Does it know whether vendor systems store or reuse agency information? Are AI-assisted outputs being reviewed by a human before they become part of an official record? Would the agency know how to answer if a prosecutor, defense attorney, assessor, solicitor, governing body, or resident asked whether AI was involved?

Those questions are not meant to create alarm. They are meant to start the conversation where it should start: with reality.

AI is not coming to public safety. It is already here. Sometimes it is official. Sometimes it is embedded. Sometimes it is informal. Sometimes it is helpful. Sometimes it is risky. And sometimes it is operating in the gray because no one has taken the time to identify it, define it, and manage it.

That is the work ahead.