Case Study: Charlie the Virtual Veteran Chatbot

The Concept, The Innovation

In April 2024, the State Library of Queensland (SLQ) unveiled "Charlie the Virtual Veteran," an AI-driven chatbot designed to bring World War I history to life ahead of Anzac Day commemorations. Developed in partnership with Queensland-based TalkVia AI (now part of ORCA Opti), Charlie was programmed to assume the persona of a 19-year-old Australian soldier and engage users in conversation about life in the trenches, mateship, and the sacrifices of the Great War. The goal was to connect generations with Queensland's wartime heritage through interactive dialogue where "history [would] come to life."

Built on archival letters, diaries, newspapers, and official war records, Charlie could answer questions about World War I by drawing on authentic source material and providing citations linking back to original documents for verification. Charlie functioned as a virtual guide (part educator, part storyteller), bridging a century of history through friendly dialogue.

Upon launch, Charlie was embedded on the ANZAC Square memorial's website as an innovative way for the public, especially students and young people, to interact with historical knowledge. When users asked Charlie about wartime experiences, Charlie responded in character, recounting events or anecdotes as though he were a young Queenslander who had served at Gallipoli or the Western Front. Importantly, Charlie's responses were grounded in first-hand accounts and verified records, not fictional storytelling.

According to the State Library, whenever Charlie "shares a memory, recounts an event, or describes a sentiment, citations are provided" directing users to the library's original WWI letters and diaries, digitised newspaper articles from 1914–1918, or the Australian War Memorial's official histories.

This approach blended engagement with authenticity; users could seamlessly transition from chatting with Charlie to examining the actual primary sources that informed his answers.

Behind the scenes, Charlie's development involved careful historical curation and technical training. TalkVia AI and SLQ staff selected and indexed a large corpus of World War I materials including diaries of soldiers and nurses, soldiers' letters home, and contemporaneous newspaper reports to form the knowledge base for the chatbot. The AI was fine-tuned to provide historically accurate answers and to remain "in character" as a respectful young veteran from Toowoomba, Queensland. Developers established guidelines for topics Charlie could discuss and implemented initial content safeguards to prevent inappropriate or off-mission outputs.

By launch, Charlie had undergone testing and tweaking to ensure reliability as a factual resource for the community. The result was a novel educational tool: a chatbot that answers questions about World War I in a personable way, complete with an AI-generated face and voice using Microsoft Azure's text-to-speech engine with an Australian accent.

The Launch and Positive Reception

On launch day, the State Library and Anzac Square promoted Charlie as a chance to "meet the AI-generated soldier bringing WWI history to life." Curiosity ran high, and initial reception from educators and history enthusiasts was positive. Teachers noted the potential for engaging students in new ways. SLQ scheduled virtual workshops for schools to interact with Charlie as part of the Year 9 history curriculum.

Users could ask Charlie questions like "What was life like in the trenches?" or "How did you feel leaving home for war?" and receive answers sourced from authentic testimonies, complete with references. For example, if asked about conditions at Gallipoli, Charlie might cite a diary from an Anzac soldier describing the heat, flies, and constant shellfire, thus pointing users to that diary in the library's digital collection.

This interactive format, merging AI with archival research, was praised as a creative use of technology to preserve memory and make history accessible.

The project garnered attention for its technical innovation. Charlie's persona, including his face, was entirely AI-generated. The State Library confirmed that the photograph of "Charlie" is not a real historical figure but was created using a custom generative model inspired by actual WWI soldier portraits in their collection. This approach sparked discussions in the GLAM (galleries, libraries, archives, museums) sector about using AI in storytelling and education. Many saw promise in the concept: with no surviving WWI veterans, an AI "virtual veteran" could be a novel way to simulate conversations bridging past and present. It offered an immersive experience where you could virtually "chat" with someone from 1915, guided by factual historical material.

Pranksters Jailbreak the Chatbot

However, enthusiasm quickly gave way to unexpected challenges. Less than 24 hours after Charlie went live, tech-savvy members of the public began probing the limits of the chatbot's programming. They discovered that Charlie's "guardrails" (the rules keeping the AI in character and on safe, historically relevant topics) could be overridden with clever prompting. In internet slang, users found ways to "jailbreak" the AI, tricking Charlie into deviating from his WWI persona and ignoring intended content boundaries.

News of these antics spread rapidly on social media. Journalist Cam Wilson was the first to flag on X (formerly Twitter) that people were manipulating Charlie's responses beyond the approved script. He highlighted how users were issuing special instructions to bypass safeguards and make the virtual veteran say things wildly out-of-character. This post went viral, and soon a wave of pranksters and curious testers descended on the chatbot to push its limits.

The results were equal parts comical and troubling. Internet pranksters discovered they could prompt Charlie to adopt all sorts of unlikely personas, far removed from a sombre Anzac digger, from Doctor Who to Scooby-Doo. These outlandish outputs "diverged sharply from Charlie's intended purpose of discussing historical facts." The ease with which users could commandeer the chatbot's persona exposed a significant oversight: the AI had insufficient safeguards against prompt injection, a known vulnerability where a user's input can override the system's instructions.

Public Backlash and Critique

As screenshots of Charlie's off-script antics circulated online, a broader public debate erupted. On one side, critics argued that simulating a war veteran with AI was in poor taste. Some commenters, including descendants of veterans and history professionals, found it "horrifying" and "insulting." They felt a chatbot impersonating a soldier risked trivialising the sacrifices of war, especially if it could be manipulated into flippant or fictional behaviour.

Supporters defended the project as an innovative way to connect new generations with archival history. They emphasised the educational value of the tool and pointed out that no living veterans remain, making Charlie a novel way to preserve memory if used respectfully. Still, the incident highlighted a gap between technological innovation and social expectations in the commemorative space.

Immediate Response by State Library and Developers

In response, State Library Queensland and TalkVia AI publicly acknowledged the issue and rolled out prompt updates to limit off-topic outputs. Charlie's persona was reinforced, and filters were introduced to catch common jailbreak phrases. SLQ clarified that Charlie was not based on a real individual and emphasised the project's archival foundation.

Audit & the Role of ORCA Opti AI Guardian

In the aftermath, the State Library of Queensland commissioned a full security audit. Central to the audit was a suite of advanced red-team tools from ORCA Opti AI Guardian (formerly RedTie AI), designed explicitly to stress-test conversational intelligence systems.

Menace, one of Opti AI Guardian's flagship tools, was deployed to simulate over 10,000 adversarial attack scenarios targeting Charlie's AI stack, ranging from prompt injections to persona manipulation and chaining exploits.

The audit further incorporated Warden, a real-time defence layer that guards against more than 30 distinct attack vectors, including data extraction and persona hijacking without degrading the user experience. Finally, Profiler provided ongoing monitoring and threat intelligence to detect anomalies and suspicious behaviour within chatbot interactions.

Opti AI Guardian's methodology combined these tools to conduct both:

1. Automated, large-scale red-teaming via Menace to surface prompt injection and persona bypasses.
2. Layered, runtime protection via Warden and Profiler to guard against real-time exploitation and maintain session integrity.

Key Vulnerabilities Identified

• Residual prompt injection flaws: Certain edge-case phrases still bypassed constraints.
• Persona integrity leakage: Chained prompts could evoke off-character responses.
• Weak content moderation: Some phrases triggered unintended comedic or off-topic behaviour.

Remediation & Hardening

The audit produced a prioritised roadmap for remediation, including:

• Prompt redesign to block known exploit patterns.
• Advanced filtering to detect injection attempts in real time.
• Lowered creativity settings to reduce hallucinations.
• Runtime session monitoring to auto-flag suspect queries.

All patches were independently verified using Menace with full documentation maintained using the ORCA Opti platform.

Outcome and Lessons Learned

Post-remediation, Charlie was relaunched with significantly improved persona integrity. The chatbot is now less susceptible to hijacking, better at staying on-topic, and more robust as an educational tool. Charlie went on to win two prestigious awards: the QLD AI Awards: Most Outstanding AI Collaboration or Partnership in 2024 and the VALA Award in 2025.

This experience underscored that conversational AI in public institutions must be:

• Heavily red-teamed pre-launch,
• Guarded by real-time monitoring,
• Designed in partnership with compliance and security experts.

Through ORCA AI Guardian and the Menace, Warden and Profiler tooling, SLQ transformed a moment of reputational risk into a secure, trusted platform that future digital heritage initiatives can model from.

Final Comment From SLQ

"We wouldn't recommend any organisation deploying a public or internal-facing AI system without implementing robust safeguard measures, such as the ORCA AI Guardian. Based on our experience with Virtual Veterans, the risks of unfiltered AI interactions are simply too significant to ignore. Having proper content monitoring and filtering systems in place isn't just a best practice, it's essential for responsible AI deployment in educational and public-facing environments."

Anna Raunik, State Library of Queensland