Project summary

As AI systems become more capable and increasingly involved in agriculture, food production, wildlife management and other decisions that affect animals, we need better ways to understand how these systems reason about animal welfare.

We propose to develop the Animal Welfare Benchmark for Transformative AI (AWB-Africa), an open evaluation benchmark for testing how advanced AI systems recognize, reason about and respond to situations involving non-human animal welfare. The benchmark will cover livestock, poultry, fish, working animals and wildlife, with particular attention to African and other low-resource settings that are often underrepresented in AI evaluations.

We will develop realistic scenarios with animal-welfare and veterinary experts, then evaluate several advanced AI models across African, low-resource and higher-resource contexts. We will examine whether models recognize underlying welfare concerns when circumstances, resources and cultural settings change, and whether reasoning developed in one context transfers to another.

We will also test whether simple model instructions or “animal welfare constitutions” can improve performance. The project will produce an openly available benchmark, methodology and findings that can be used by AI safety researchers, animal-welfare organizations and AI developers to identify gaps and improve how advanced AI systems account for animal welfare.

What are this project's goals? How will you achieve them?

The project has three main goals.

1. Develop a practical benchmark for animal-welfare reasoning.
We will create a set of realistic scenarios that test whether advanced AI systems can recognize animal suffering and welfare risks, reason about trade-offs, and identify less harmful alternatives. Scenarios will cover livestock, poultry, fish, working animals and wildlife. We will begin with African and other low-resource contexts, while also developing comparable higher-resource scenarios.

2. Test whether AI welfare reasoning generalizes across contexts.
We will evaluate several capable AI models using the benchmark and compare their performance across different species, geographic settings, resource levels and circumstances. We want to understand whether a model that performs well on familiar or high-resource scenarios also recognizes welfare concerns when infrastructure, economic conditions and available interventions are different.

3. Identify and reduce important failures.
We will analyze where models fail and test whether relatively simple interventions, such as targeted system instructions or an “animal welfare constitution,” can improve their responses. We will measure whether these improvements remain reliable across different contexts rather than only improving performance on the scenarios used for testing.

The work will combine literature review, scenario development, expert consultation, structured annotation, model evaluation and comparative analysis. Animal-welfare and veterinary experts will help validate the scenarios and evaluation criteria, while our technical team will build the benchmark, run the model evaluations and analyze the results.

The final outputs will include an open benchmark, evaluation methodology, documented results and tested instruction templates that other researchers and organizations can use to study animal welfare in increasingly capable AI systems.

How will this funding be used?

The funding will be used to develop, validate and evaluate the Animal Welfare Benchmark for Transformative AI.

A substantial portion will support the people needed to build a rigorous benchmark, including animal-welfare and veterinary experts, AI/evaluation specialists, research assistants and annotators. They will help develop realistic scenarios, review the welfare issues involved and establish consistent criteria for assessing AI responses.

Funding will also support scenario and data development in African and other low-resource settings. We will gather real-world examples and consult people with experience in livestock, animal care, wildlife and veterinary services, then develop comparable scenarios from higher-resource settings for cross-context testing.

Part of the funding will cover access to AI models, computing, evaluation infrastructure and data management needed to test multiple capable systems systematically. We will also support expert workshops, analysis, documentation and external review.

The remaining funds will support publication and release of the benchmark, methodology and evaluation tools as open resources. This will allow other animal-welfare researchers, AI safety researchers and developers to build on the work after the grant period.

Who is on your team? What's your track record on similar projects?

The project will be led by Coded Foundation for Technology Development (CFTD), a Nigerian technology-focused nonprofit with experience in AI, data science, digital systems and community-based research.

I will lead the technical and research side of the project. I have an MSc in Computer Science and practical experience working with machine learning, large language models, data systems and AI applications. My work has included developing machine-learning models, building data and monitoring systems, working with LLM-based tools, and supporting research projects involving data analysis and AI. Through CFTD, I have also led technology projects involving government institutions, researchers, young people and communities across Northern Nigeria.

Our strongest track record is in AI, data and technology rather than animal welfare specifically. We therefore do not plan to treat animal-welfare expertise as something we can provide internally. We will bring veterinarians and animal-welfare researchers into the project to help develop and independently review the scenarios, welfare criteria and interpretation of results.

The wider team will include an AI/ML evaluation advisor and research assistants responsible for scenario development, annotation, model testing and analysis. We will also draw on local contributors with experience in livestock, animal care and wildlife to ensure that African and low-resource scenarios reflect real conditions rather than assumptions.

This combination gives us a strong technical foundation while being transparent about where specialist expertise is needed. The project applies capabilities we already have in AI, data, research and evaluation to an important area that is new for our organization.

What are the most likely causes and outcomes if this project fails?

The biggest risk is that the benchmark may not be sufficiently rigorous or useful to support meaningful comparisons between AI systems. Animal welfare involves difficult questions about species, context and trade-offs, and some scenarios may prove too subjective or difficult to evaluate consistently.

A second risk is that our African and low-resource focus creates challenges around data availability, expert access or comparability between settings. Some welfare situations may not translate cleanly across countries or between low- and high-resource environments.

There is also a risk that the benchmark does not gain significant use after release. A technically sound evaluation does not automatically lead to adoption by AI researchers or animal-welfare organizations.

If these risks materialize, the project could produce a smaller or less generalizable benchmark than intended, with limited external uptake. We would reduce these risks through early piloting, expert review, testing across multiple models and external feedback before finalizing the benchmark.

Even if adoption is limited, the project would still generate useful evidence about how current AI systems reason about animal welfare and where their failures occur. We would publish the methodology and findings openly so that the work can contribute to future research rather than depend entirely on immediate uptake.

How much money have you raised in the last 12 months, and from where?

Over the last 12 months, Coded Foundation for Technology Development has received funding and income through a combination of grants, project support, training activities and donations. Our work has been supported primarily through project-based funding and earned income rather than a large unrestricted institutional funding base.

We have not previously raised dedicated funding for an animal-welfare or AI-safety project. The proposed Falcon Fund grant would therefore provide the first dedicated funding for this area of work.

We are happy to provide a more detailed breakdown of our recent funding sources, amounts and restricted/unrestricted status if required.