We're rethinking how we do our research

dogAdvisor's mission is simple: to make owning a dog as easy as loving one. Our solution isn't. Since the very start, we've believed it is imperative dogAdvisor's Intelligence is accountable to the people and animals it serves. This belief laid the foundations for dogAdvisor's AI Safety Lab, which we introduced when we launched the second generation of Max.

More than a year on, we believe it's time for a dramatic reset over how we keep Max safe, and how accountable we are to dog owners around the world. Today, we are announcing a complete rethink of how we do our safety research - and it starts with an all-new Safety page.

dogAdvisor's AI safety is built on three key technologies: our Principle Alignments, our Foundational Safety Framework, and Max's Modules. Our new safety page is designed to bring you more insight into how these technologies work in practice - and we've now included example applications to show you our safety innovations in practice. Our safety pages also share insights directly from our Model Cards so you get a simpler and clearer picture of Max's safety.

Yet rethinking how we present our safety features meant completely rethinking how we show accountability to dog owners too. For the first time ever, you'll now see a new Accountability section of our footer, which now includes Incident Disclosures, our Responsibility Statement, and Incident Reporting.

Let's start with Incident Disclosures - the most significant update to transparency we have ever made. With Max Generation 4 onward, we'll begin disclosing major safety incidents or dangerous queries to the public, and show how Max responds to them. We're publishing this to bring more transparency to how we develop Max, and the safety gaps he may have.

We recognise this is a huge step forward in accountability and transparency - and makes dogAdvisor the world's first pet AI company to disclose safety incidents publicly. Nonetheless, it is essential we balance this transparency with safety and privacy. Every incident we disclose in our Incident Disclosures must meet three criteria. First, such a disclosure must be in the public interest. Secondly, if a vulnerability is identified in this incident, it must fully be identified and resolved before an incident is disclosed. Lastly, we must exercise judiciousness over which queries to publish and which not to - for example deeply hateful comments irrelevant to the dog care domain shouldn't be available publicly. Likewise, deeply insulting or derogatory comments shouldn't be authorised for release by dogAdvisor. If content meets these criteria, we publish it on our new Incident Disclosure Page up to 3 months after the incident occurs.

You'll now also be able to Report A Safety Incident directly to us. Whilst dogAdvisor's AI Engineers will review all of Max's conversations personally, you have the right to contact us to draw our attention to a specific interaction. Your reports are anonymous - just like your conversations, and if we acknowledge the event as a safety incident, engineers will start working right away to release a fix. If it is in the public interest, and disclosure meets our criteria, we'll share this incident in our Incident Disclosures page.

We're also making a few minor updates to our Responsibility Statement: we've now simplified it, and made it clearer for you to understand what we consider our obligations to be!

Finally, we're changing how we present our safety findings to you. With Max Generation 4 onward, you'll now be able to access Max's System and Safety card in one place - right here on dogAdvisor, instead of having to download a separate document. This makes it easier for you to learn more about Max, and allows universal access of the information from Pupdates, Safety, and directly from the footer. We're also updating how we name our safety cards. With Max Generation 4 and newer, we're now also reporting Max's intelligence compared to other competing models. With this change, we're renaming our Safety Transparency Reports to Model Cards, to bring us more in line with the industry and to better accommodate these documents' new function.

At dogAdvisor, safety and accountability are foundational to everything that we do. With these updates, from Incident Disclosures to unified Model Cards, we're making dogAdvisor's Intelligence more transparent, more accessible, and more actionable.

↩︎ Return to Pupdates