The Animal AI Incident Observatory
What we’re building
René Itah and I are building a public platform to track incidents of AI-caused harm to non-human animals. Some examples of these harms include autonomous vehicles hitting animals on roads, smart home systems creating unsafe conditions for companion animals, and robots interfering with nesting and migratory patterns in wildlife.
Our design ingests data from multiple different sources and uses an LLM to detect instances of AI-caused harm to non-human animals. These instances are then reviewed by a human prior to publication on a publicly accessible dashboard. We want to build this with input from people already working on AI+animal welfare, so if this is a space you care about, we’d love your take.
Why we’re building it
AI systems increasingly make decisions that affect animals, but there’s no central place tracking when those systems cause harm. Reporting spans a variety of sources from news reports to mandatory reporting (ex. NHTSA) to research papers. Some AI incident databases already exist, like the legal hallucinations database and AIID, but nothing that focuses on non-human animals. A centralized record of these incidents would equip researchers, advocates, and policymakers to identify patterns, prioritize interventions, and hold system owners accountable.
Some specific questions
Would something like this be valuable to you? If so, for what purposes?
Are there any specific data sources or AI systems you would want to see represented?
Are there any specific classifications on incidents that would be helpful for your work (ex. aquatic animals, mild v. severe harms, species-specific, zip code)
Given that we’re using LLMs to filter and classify incidents, what would make you trust (or distrust) the resulting dataset? What would you want to see in terms of transparency, methodology, or human review?
Is there anything that you believe we’re missing or not considering?
We’d very much appreciate your feedback either in the comments below or by email.
Hi, thanks for working on this! Your post provoked some interesting thoughts. I’m interested in this topic from the biodiversity/extinction angle, so I’m not sure if I’d be a good fit. Here are some thoughts I had:
I am expecting a lot of damage from support infrastructure—power generation, data centers, transmission lines, and mining before I see direct damage from autonomous robots. This is just because AI is so information-centric right now, although autonomous machines are coming shortly.
I think the most damage would come from the increase in doing work that was too impractical before—things like spraying for pests constantly instead of a few times a year. Fishing constantly. Scanning constantly. So damage from increased disturbance and an increase in activities that were previously not done.
However, this might in some cases look worse but be better, so it seems really important to compare to what it is replacing. For example, if you spray constantly for pests, maybe you use less pesticide overall and keep populations low on your crops, without off-target pesticide drift to surrounding areas causing mass die-offs.
I’m expecting the most increase in AI activities from aerial drones because they are easily built and deployed in a wide variety of environments for a lot of activities. (Rather than androids, vehicles, or factory machine arms).
Something from the information side of things—AI unearthing, deducing, and providing more information to rare species collectors might endanger rare populations of birds/insects/plants/etc as they collect them to death to sell on the black market. Secrecy is their main mode of protection right now.