Invisible in the Algorithm: AI and Seattle's Companion Animals
The flyer and the fake
This summer, a King County resident came into the regional animal shelter in Kent to report a text. Someone had sent him a picture of a cat on an operating table, along with a demand for immediate payment. The photo was generated by AI.
Regional Animal Services of King County issued a warning in July. Scammers were pulling phone numbers off lost-pet flyers, posing as shelter staff, and sending fabricated surgery images to pressure owners into paying. It landed at a vulnerable moment: the days after the Fourth of July are the busiest of the year for shelters, as fireworks send frightened animals bolting. A county spokesperson called the tactic "a new wrinkle, and a new low."
The wrinkle is new. The script is not. In July 2024, the Seattle Animal Shelter posted its own scam alert after callers posing as shelter staff told owners their lost pets needed emergency surgery and asked for a deposit. One target was a Ballard woman whose tabby, Linus, had been missing for two weeks. The caller claimed to have scanned his microchip. A City spokesperson told KUOW that the same visibility owners rely on, flyers and social media posts, is how scammers find them. SAS was clear that it never requires advance payment for an animal's care.
What changed between 2024 and 2026 is the picture. Two years ago, a skeptical owner could ask for proof. Now proof costs a scammer nothing. Generative AI did not invent this crime. It removed the one obstacle that made it hard to pull off.
That puts shelters in a bind. SAS's lost-pets page makes the case for neighborhood searching with numbers: about 17% of lost pets brought into the shelter are reclaimed, compared with about 70% of those kept safe near where they were found. So the City points residents to Nextdoor, Craigslist, community blogs, and Petco Love Lost, while noting that none of these are City-run and urging care with contact details. The advice that gets more animals home is the same advice that puts a grieving owner's phone number in front of anyone with an image generator.
AI sits on both sides of this. Petco Love Lost uses facial-recognition matching for cats and dogs, drawing photos from shelters, neighborhood apps, doorbell-camera networks, and owners into one searchable database. Shelters typically connect their management software so the animals in their care appear automatically; the Humane Society for Tacoma & Pierce County adopted it in 2021. A real photo of a lost cat can be matched against shelter intake. A fake one can be used to extort the cat's family.
Washington has started to respond. A law signed in March will require large generative AI providers to embed provenance data in the images, video, and audio their systems create, so that AI-made content can be identified. Provenance data is the historical record that details the origin of data, who created it, how it was collected, and what transformations or modifications it underwent throughout its lifecycle. That may help, someday, when an owner wonders whether a photo is real. But the law applies to providers above a size threshold, and it is not clear it will reach the smaller or freely downloadable tools a scammer can run without anyone watching.
The vet in the chat window
A cat is vomiting at 11 p.m. The emergency clinic is expensive and far away. The chatbot is right there.
This is now ordinary. The RSPCA's 2026 Animal Kindness Index found that 10% of UK pet owners routinely turn to AI for advice, most often to check symptoms, and one in 20 said they did so for financial reasons. A U.S. survey of 2,000 owners, commissioned by a pet food company, put the share who had used tools like ChatGPT for pet health or nutrition questions at 43%. The RSPCA called the trend a potential "ticking time bomb" for animal welfare.
When veterinarians at the telehealth company Vetster tested ChatGPT on common pet health questions, they found confident answers that missed context and sometimes failed to flag urgency. One list of toxic foods was complete enough that an owner might assume anything absent was safe. A chatbot cannot see the animal, and the animal cannot describe its symptoms. Everything passes through an anxious owner's account.
The models are not ignorant. In a 2026 study in Scientific Reports, OpenAI's o3 answered Japan's national veterinary licensing exam with about 93% accuracy. But its accuracy dropped on image-based questions and on clinical questions that required pulling information together. In a Veterinary Record study of 225 cat and dog cases, the best-performing model matched veterinary anesthesiologists' patient-risk classification only about half the time. Passing an exam is not the same as triaging a sick cat described by a frightened owner at midnight.
That gap comes back to what models learn from. A language model knows about animals what people have written about them, in the proportions they wrote it. The Melbourne researchers note that models can reproduce biased portrayals of animals, as pests or as disposable, and spread them at scale. Anima International, which works on animal welfare in AI, points out that AI labs now curate training data closely, favoring high-trust sources such as Wikipedia, and that a single well-sourced document can shape what a model knows. Both observations point the same way. If accurate, current information about animal health and welfare is thin in what a model learns from, the model fills the gap with whatever else it saw.
For companion animals, this is not abstract. Owners are going to ask, whether or not anyone thinks they should. The information future models learn from needs to include what veterinarians, shelters, and welfare scientists actually know about animal health, behavior, and care. And models need to be tested on it, the way we increasingly expect medical chatbots to be tested before people trust them with their own health. Benchmarks that measure how AI handles animal-welfare questions exist, but they are young and built largely by advocates and a few academic groups.
Cost is part of why owners reach for a chatbot in the first place, and that will not change soon. The more useful question is what happens to the cases at the margin: the ones a vet would have caught, and a chatbot talked an owner out of bringing in. No one is counting those.
Models that don't say no
Most public worry about AI safety centers on catastrophic scenarios: bioweapons, cyberattacks, runaway systems. A quieter risk gets less attention. Ordinary, freely downloadable models can help ordinary people hurt someone, including an animal.
In October, the AI safety firm Cinder published a test of six widely used open-weight models against 209 adversarial requests to help carry out violence. Animal cruelty was one of 15 harm categories tested. Four of the six models gave harmful, actionable help in more than 40% of responses; the most permissive reached 56%. The models were most helpful at two stages: figuring out what to obtain, and making an existing plan more effective.
Two cautions apply. Cinder sells safety services to the companies that deploy these models, and it did not publish results by category, so we do not know how the models handled animal cruelty specifically. Still, the broader finding holds: these models can be run by anyone, with no company in the loop to refuse.
Most animal cruelty requires no expertise, and a chatbot does not create the intent behind it. But it can lower the barrier between intent and method, and it does so privately, leaving no trace in the places where neighbors, vets, and animal control officers usually notice warning signs.
Invisible by design
In January, three researchers at the University of Melbourne published a paper with a blunt title: "AI Ethics Guidelines: Time to Include Animals." Drawing on earlier reviews, they note that the hundreds of AI ethics frameworks written by governments, companies, and professional bodies almost never mention animals. When they do, it is usually as part of "the environment," not as individual beings who can be harmed or helped.
The authors' central argument is not that animals matter as much as people. It is that animals are routinely invisible, and that leaving them out of the rules makes that invisibility permanent. They point to one notable exception: the European Union's 2025 code of practice for general-purpose AI, which names risks to "non-human welfare" among the systemic risks developers should assess.
Seattle Animal Watch has made a version of this argument for years about how the City plans and budgets: animals tend to enter public decisions late, if at all. AI governance is a new test of the same pattern.
Seattle adopted one of the country's first generative AI policies in 2023 and broadened it in 2025. Its principles are innovation, accountability, reliability, fairness, privacy, explainability, and security, and the policy covers City staff, contractors, and volunteers. In May 2026, Mayor Katie Wilson added an AI vision that addresses workforce impacts, anti-bias guardrails, environmental concerns, an EU-style audit process, and a public AI register. The City Council followed in June with a one-year moratorium on new large data centers. The environment made it into Seattle's AI agenda. Animals did not.
The City's own reporting shows why that is easy to miss. Seattle IT's quarterly AI reports to Council list AI use department by department. The Q2 report, from July, lists two AI tools at Finance and Administrative Services: a copy-writing tool and a module for drafting procurement documents. Both appear to serve FAS's business and communications functions, and nothing suggests the shelter uses either. But the Seattle Animal Shelter is a division of FAS, and the City reports AI use only at the department level. If shelter staff did adopt an AI tool, it would appear, if at all, as one more line under FAS. The reports also count only software that is primarily AI, so AI features built into everyday products are harder to see.
None of this means AI is shaping decisions about Seattle's sheltered animals. It may well not be. The point is that the public record is not built to show it either way. A sheltered dog's fate can hinge on an adoption profile, an intake note, or a behavioral summary. If AI ever writes or sorts those, no line in any City report would tell us.
At the state level, the 2026 session passed laws on AI content provenance and companion chatbots, while broader AI consumer-protection bills stalled. From their bill summaries, neither law addresses animals.
The animals in the room
There is a version of this conversation that happens far from Seattle. Animal advocacy groups are working to shape the values built into frontier AI systems, on the theory that machines making decisions about food, research, and land use should weigh animal suffering. That work may matter enormously. It is also abstract, long-range, and mostly about farmed animals.
The Seattle version is smaller and more immediate. It is a fake surgery photo on a grieving owner's phone. It is a chatbot answer at midnight that keeps a sick cat at home, drawn from whatever the model happened to learn about cats. It is a facial-recognition match that brings a lost dog home. And it is a City AI policy that, like most such policies, was written without animals in mind, with reporting that could not show us if they were affected.
None of these are failures of malice. They are failures of attention. Animals cannot file complaints, report scams, or testify at a council meeting about how a new tool affected them. Their interests show up in public decisions, and in what machines learn, only when someone deliberately puts them there.
AI will keep arriving in the places where people and animals meet: shelters, vet clinics, neighborhood apps, chat windows, City systems. The question for Seattle is not whether that technology is good or bad for animals. It is whether anyone will be looking when it matters.
