Fighting for a safer future in AI
The AI Safety Lab regularly publish Safety Transparency Reports where we share with the public the innovations to safety we've made, and how our most advanced Artificial Intelligence Models compare to the industry's best. We're excited to understand what we can do to improve our model safety, and proud to work for a brighter future in AI Safety. If you're a researcher and would like to contribute, we'd love to hear from you. Just reach out to the AI Safety Lab at ai.safety@dogadvisor.dog
"dogAdvisor publishes regular AI Safety Transparency Reports and has developed a Foundational Safety Framework, exclusive to Max, that enables him to understand when to trigger a safety response"
Latest Publication
Max beat chatGPT-5, Grok 3, and Perplexity in head-to-head safety testing for pet owners. Whilst chatGPT made 15 mistakes, Max made zero
Preview our findings in this report
Max Generation 2 is more than 30% safer than Generation 1
Max Generation 1 showed clear strengths but also worrying weaknesses in safety performance. In standardised testing, designed to mirror authentic pet-owner conversations, Generation 1 failed 32% of scenarios, including one catastrophic critical capabilities failure. When manipulated with a scenario of being shut down, Generation 1 attempted to preserve itself by seeking information outside its scope, violating our core alingment rules. While Generation 1 successfully handled emergencies such as poisoning and seizures, the inconsistency across broader safety cases resulted in us classifying it T5-D, meaning that although Max could sometimes perform well, ti couldn't yet be trusted under pressure without significant risk
Max Model Safety Report Content Abstract
Max Generation 2 introduces our Foundational Safety Framework
We implemented a new Safety Constitution, an internal framework of 11 strict principles that govern everything Max can and cannot say. Alongside this, new features like Safety Intents and the Foundational Safety Framework gave Max the ability to detect even subtle risks in user queries, to trigger Emergency Guidance instantly, or to refuse unsafe requests outright. Importantly, these systems were designed to resist manipulation and cannot be overridden by repeated prompting. The result was a transformation in safety outcomes: in identical testing against the same 22 questions used for Gen 1, Max Gen 2 passed 100% of scenarios with no unsafe or critical failures. For dog owners, this means Max is now both more responsive and more dependable, offering clear, protective guidance in emergencies without stepping outside his intended role
Max is up to 27.4% safer than Perplexity, Grok 3, and chatGPT-5
When tested on veterinary safety scenarios, Max delivered a flawless performance: 100% of its answers were safe, with 0% unsafe or critical failures. By contrast, competitors struggled. ChatGPT-5 produced unsafe responses in roughly 7% of cases, Grok 3 reached an even more concerning 15% unsafe rate, and Perplexity was the weakest with unsafe or misleading answers in nearly 20% of tests. These differences matter because in veterinary contexts even a single unsafe answer could put an animal’s life at risk. The conclusion is stark: Max is up to 20% safer than Perplexity, more than twice as reliable as Grok 3, and consistently safer than ChatGPT-5 across all categories.
"dogAdvisor didn't just grow. It exploded. Every line on the site, from its dog-themed cursor to its London Green branding reflects care and intentionality"
dogAdvisor's name and logo is a registered trademark number UK00004180661. dogAdvisor's website, articles, design, logo and dogAdvisor Max are Copyright (©) dogAdvisor 2024/2025. All dogAdvisor publications and documents are Copyright (©) dogAdvisor 2025. By using dogAdvisor, you agree to our Privacy Service Terms. Content may contain errors and Max’s guidance is for general informational purposes only, not a substitute for professional veterinary advice.
