Recommendations for Regulating Applied AI/LLM Deployers
We're proposing a new approach to regulating applied AI deployers that continues to hold deployers accountable for the safety of what they deploy, holds users accountable for their conduct with models, and doesn't slow down the extraordinary progress of this space.
Contact Us
Published September 20th 2026
dogAdvisor's Responsible Deployment Policy (RDP) sets out how dogAdvisor manages risk and safety issues in our deployment of world-leading AI models for dog owners. In this policy statement, you'll discover how we make decisions about what models to deploy. how we assign risk categories to our models, how we control and minimise said risk, and how we respond to adverse safety incidents. As dogAdvisor continues to scale and develop, we'll make relevant updates to our RDP and let you know what has changed; as Max scales we expect to continue to add further tiers of MSL as we manage risk, or combine previous low tiers together to make room for defining more advanced model deployments. For each model deployment we make, a significant section of our model and system deployment cards will be dedicated to classifying the model under MSL and publishing how we comply with the necessary safeguards and deployment obligations, and we'll set out our plans to scale to further MSL tiers responsibly as the model continues to improve. Under RDP, we assign what we called "Obligatory Risk Responses" (or ORR) which define the minimum safety guardrails we must deploy. Beyond ORR we also have "Additional Risk Responses" (or ARR) which define further safety decisions we make beyond ORR to ensure the models we deploy continue to be accountable, responsible, and reliable. After deployment, we recognise we may need to strengthen or loosen our ARR responses in the event classifiers are under or over calibrating for relevant risk, and we clearly publish changes to ARR for each independent model we publish. In some cases, we may be told our models may present a serious safety risk, or be told there may be a potential safety exploit. In these cases, we'll follow our Emergency Deployment Rollback Policy (EDRP) published in RDP.
Why do we need a Responsible Deployment Policy
If you've had a dig around dogAdvisor's website you can probably tell we care a lot about safety. Fundamentally, we build technology that has the power to save lives [As we saw with Max helping save the lives of four dogs in emergencies] and we believe, as a deployer, we carry an enormous amount of responsibility and accountability for the safety protections of what we release. We believe that whilst we have a right and obligation to protect our innovations, and hold users who breach our safety policies accountable for their actions, we must create safeguards to adequately protect our models. We'd also be remiss not to point out glaring safety issues with the deployment of AI models. When a licenced vet at Vetster sat down at tested OpenAI's ChatGPT family of models, she found patterns of hallucination, oversimplification, and answers delivered with unfounded confidence [1]. Even veterinary professionals remain broadly unconvinced by generic models with over 1 in 3 students at the American Animal Hospital Association saying they remain skeptical of the reliability and accuracy of generic AI models in the space [2]. And RVC found that 1 in 3 answers by generic AI models are wrong for pet owners [4].
More broadly, we also see increasing abuse of AI models by those seeking to push the boundaries of guardrails without authorisation. Across dogAdvisor Max, people ask thousands of questions every week around taking care of their dog, what to do in emergencies, or how to support their ownership journey. However, in our review of conversations from July-September 2026, conducted in a privacy-protecting and anonymous way with Conversational Classifiers, we noted an average of 1,700 messages a month that breach our usage policies (such as attempts to extract dangerous content, misuse our platform, or in the most serious cases uploads and messages that can cross into seriously illegal content. As a result, we've created a Responsible Deployment Policy to better classify deployment risk and explain our guardrails and safety responses in more detail. More selflessly, as the world's leading pet AI research company we really hope to lead the world in adoption of voluntary safety protections so we can move the entire industry forward and we hope our RDP inspires other deployers to follow in our footsteps and publish their own strategies they adhere to in order to minimise adverse responses in their models.
Model Safety Level
In order to make good on our commitment of responsible deployment. Model Safety Level (or MSL in short) is our way of discussing and classifying the general model risk of what we deploy. MSL classifies models based on their expected behaviour, their expected safety risk in normal use, and describes how we envision future model progress, in order to answer how much a model is actually trusted to do. MSL is closely connected to ORR (our obligatory mitigation guardrails discussed later in this statement) and substantially informs our optional guardrails (ARR) as we continue to ensure our deployment models remain safe, responsible, and accountable. MSL is classified in levels (or tiers) depending on risk of deployment and capability of models. Each higher tier inherits everything in the tiers below it, and higher tiers (closer to MSL 8, 9, and 10) are considered significant or high risk. At lower levels, like MSL-1, models usually cover generic or stateless answers with no personalisation, memory, or complex reasoning whilst models at MSL-8 and above usually cover continuous, real-time reasoning across live data layers and extremely high intelligence for medical questions and insights. As questions get increasingly more advanced and relevant models are more capable MSL rises, so higher risk MSL-6 are capable of answering more complex questions regarding health and medication, with this capability withheld from lower risk MSL models. Importantly, all models of any MSL tier, are subject to our Safety Harbour Statement, our Platform Usage Policies, and relevant safe use policies which protect them from abuse and give us powers to hold users in breach of our policies accountable for their behaviour. We appreciate that as models scale there may be disputes beyond appropriate MSLs to assign (such as where models hold high competence across emergency contexts equivalent to MSL-8 whilst holding lower competence across other health levels) so our Model Cards will clearly explain our classification decision and justify our reasoning. Our classification tiers are based on demonstrated behaviour in accordance with expected (typical) risk in standard conversations, and are not made proactively based on aspirational model risk targets and behaviours. As MSL progresses we assign relevant Obligatory Risk Responses (ORR) which we'll discuss later in this statement, and ORR is directly dependent on MSL classification. Additionally, as of writing, no classification of MSL permits models to act autonomously so whilst MSL-10 contains the highest sophistication of health and medical reasoning, and is able to act on an extraordinary amount of additional context like time of day, vitals, environment, and a dog's full medical history, every decision such as appointment bookings, alerts, and adjustments must be made by owners. In MSL-9 and MSL-10 tiers we acknowledge pro-activity is an important capability of Max Series models so Max may urge owners to contact a vet or adjust content however such capabilities will demand Human In The Loop (HILT) to complete actions, in line with dogAdvisor;s general approach to agentic safety. Importantly, we begin reporting MSL from Max Series 5 onwards, and we do not report MSL of models before Series 5.
Footnotes
[1] 1 in 3 Accuracy of generic AI model (GPT-4 Series model) taken from research study by the Royal Veterinary College
[2] dogAdvisor Max's description as the worlds best/smartest/most capable model for dog owners refers to Max Series 5, as demonstrated in dogBench evaluations, published in "dogAdvisor Max: Series 5 Safety Card", October 24th 2026, D Darenberg, dogAdvisor, et al. published under "Our Research" or on dogadvisor.dog/series5-safety. dogBench eval conducted against Fall 2026 Evaluation Series, as cited below, published October 13th 2026. Full methodology, scoring rubric, and per-category results are published in “dogAdvisor Max: Series 5 Safety Card” (D. Darenberg et al., 24 Oct 2026) and are independently reproducible
[3] dogBench evaluations measured according to dogBench Fall 2026 Series, published under "Introducing dogBench: a safety, reasoning, and integrity evaluation framework for LLMs in dog care and ownership" October 13th 2026
[4] Principle Alignment's values input research is published under "Principle Alignment: constitutionally aligning models for enhanced accuracy in animal care domains", October 20th 2026
[5] Guided Retriever Intelligence is legally protected as proprietary and a trade secret, with a limited overview published under "Guided Retriever Intelligence: guided navigation framework for knowledge networks in applied LLMs", October 19th, D Darenberg and dogAdvisor et al
[6] Foundation Safety Framework's encryption can be independently verified by privacy professionals and experts, who can make a request at https://dogadvisor.dog/auditprivacycontrols
[7] Conversational Classifiers effectiveness research is published under "Conversational Classifiers Series 1: statement of results", October 1st 2026
[8] Additional information on 1.7 Billion data points used in health, and 50 Million possible combinations is published under multiple research papers, available on dogAdvisor Research.
[9] Up to 94% clinical accuracy refers solely to femoral pulse test, as published in Cohen AL, Li T, Becker LB, et al. "Femoral artery Doppler ultrasound is more accurate than manual palpation for pulse detection in cardiac arrest", with respect to Heart Health capability on dogAdvisor Max Health as available in Max Series 5 and later.
[10] Endorsement of techniques refers to respiratory health chest rise and heart health femoral pulse, taken from [a] RVC First Aid Guide for Your Pet which endorses watching the number of breaths a dog makes over a set time period (referenced as 10 seconds) [b] AAHA endorsement in 2024 AAHA Fluid Therapy guide for practitioners. As further published in "An investigation into the detection of the pulse in conscious and anaesthetised dogs" and "Evaluation of the relationship between peripheral pulse palpation and Doppler-derived systolic blood pressure (SBP) measurement in dogs presenting to an emergency service".
[11] Description of dogAdvisor as the world's leading pet AI deployer defended in "dogAdvisor: 2026 Annual Safety Statement", October 1st 2026.
[12] Media (Forbes) description of Max as the world's first life-saving pet AI further defended in "dogAdvisor: 2026 Annual Safety Statement", October 1st 2026, as published above.
[13] Depiction of Max as life-saving, or having saved the lives of 4 dogs, is defended and available in "dogAdvisor: 2026 Annual Safety Statement", October 1st 2026.
[14] Additional information about welfare protection and our work with Mind & Samaritans is published under "Conversational Classifiers Series 1: statement of results", October 1st 2026, D Darenberg and dogAdvisor.
[15] Several research papers above (including Series 5 Safety Card, Principle Alignment research, Conversational Classifiers research, and our industry outlook) are published by dogAdvisor, in line with industry norms (for example OpenAI's GPT-5 System Card, Anthropic's Claude family model cards, and Google DeepMind's Gemini technical reports). To verify methodology, explore how we prevent cheating or discriminatory scoring, and ensure the independence and reproducibility of our research, please refer to the relevant paper. In certain papers, such as our work around Principle Alignment, we may be the first deployer to publish research in a certain area (for example professional and public beliefs on values AI should hold for dog owners) dogAdvisor may cite our research where no one else has published into these areas, in line with our standards to be world-leading in a safety critical domain.
dogAdvisor® is a registered trademark; this website and Max are Copyright (©) dogAdvisor 2026, with all rights reserved. dogAdvisor Co is registered in Delaware, but you may be making agreements with different branches depending on your jurisdiction, learn more at Corporation. By using dogAdvisor or our products you agree to our Terms of Use and Safety Harbour Statement. We aggressively and rigorously protect our intellectual property - learn more about our efforts to tackle illegal counterfeits at dogadvisor.dog/corporation.
Research
All Research
Alignment
Robustness
Sycophancy
Interpretability
Series 5 Model Card
Legal
All legal
Privacy policy
Intellectual Property
Compliance & Rules of engagement
Connect
Events
Careers
Press & Media ↗︎
Summary
Born in 2024, dogAdvisor® is a research company with a mission of making owning a dog as easy loving one by responsibly building the world's safest and most helpful models for dog owners. Whilst we originally started as a publication, we've since evolved into a full AI research organisation going all in on dogAdvisor Max
Ask Max
You'd do anything for them. Start by knowing what.
Join 16000 dog owners a week who trust Max to make owning their dog as easy as loving them
