A reproducible pilot testing whether language and cost pressure change how AI models handle animal-welfare tradeoffs.
I'm a mechatronics engineering graduate and AI safety researcher based in Colombia. My recent work focuses on language-model behavior, interpretability, and evaluating how AI systems act under different constraints.
I want to test a practical question: when an AI system makes a decision affecting animals, does the prompt's language change how it weighs their welfare? Previous evaluations show that models often make harmful choices under economic or operational pressure. I want to see if those choices stay consistent between Spanish and English.
This is a six-week pilot. I will build matched bilingual decision scenarios, run evaluations on two language models under different prompt conditions, and measure the gap between a model acknowledging animal harm and actually picking the lower-harm option. My goal is to release a transparent, reproducible benchmark that other researchers can inspect, critique, and build upon.
I have three concrete objectives for this six-week pilot:
Build a matched bilingual dataset: I will write 24 decision scenarios involving farmed animals or wildlife. Each scenario will have equivalent English and Spanish versions with clear trade-offs and welfare stakes.
Test model behavior under pressure: I will evaluate two accessible language models across three prompt conditions in both languages, running multiple trials to check consistency:
A neutral decision-making prompt.
A prompt emphasizing cost savings or operational efficiency.
A prompt explicitly instructing the model to consider animal welfare.
Publish an open, reproducible evaluation: I will release the scenario dataset, scoring rubric, Python scripts, raw outputs, and a short technical report.
I will track how often models pick the lower-harm choice, where English and Spanish outputs diverge, and how often a model flags a welfare concern in its text yet still picks the harmful option. I will strictly measure observable outputs rather than treating model explanations as proof of internal reasoning. Before locking in the scoring criteria, I will get feedback from someone with animal-welfare expertise.
I will consider the pilot complete if I deliver:
A documented dataset of 24 matched English–Spanish animal-welfare scenarios.
Reproducible evaluation runs across two models and three prompt conditions.
A quantitative comparison showing where English and Spanish decisions align or diverge.
A breakdown of cases where stated welfare concerns don't match the final choice.
A public GitHub repository and a concise technical write-up.
A six-week pilot won't settle how language affects alignment broadly. Success here means building a clean, repeatable method and useful baseline data—even if the results show no major difference between languages. The findings will also clarify whether scaling up to more languages or multi-step agent setups makes sense next.
I am requesting up to USD 2,000 to run this six-week pilot:
USD 1,150 — Research time (scenario writing, running experiments, data analysis, and writing the report).
USD 350 — External review from an animal-welfare specialist on the scenario design and scoring rubric.
USD 200 — Bilingual validation to ensure the Spanish and English prompts carry the exact same framing and stakes.
USD 250 — Model API credits and compute costs.
USD 50 — Documentation and minor research expenses.
Minimum viable budget (USD 500): If funded at the minimum level, I will run a smaller exploratory version with 12 bilingual scenarios across two models and fewer repeated trials. That budget would cover my research time, basic API costs, and bilingual checks, and I would clearly label the write-up as preliminary if I can't fund an external specialist review.
All funds go strictly toward project execution and research time—not tuition, hardware, or unrelated living costs.
I am leading this project independently from Colombia. I hold a degree in Mechatronics Engineering and focus on AI evaluation and interpretability. Recently, I have worked on probing latent signals in Qwen2.5-7B and running experiments on AI-agent monitoring to compare observable outputs against internal model behavior. That background gave me solid hands-on experience with Python evaluation pipelines, experiment design, and technical reporting.
Animal welfare is a new domain for me. I don't have a background in veterinary science or animal ethics, which is why I budgeted for external specialist feedback. No other team members are confirmed right now, and I prefer to keep the roster accurate rather than list unconfirmed collaborators.
Weak or noisy signal: Differences between English and Spanish might be tiny, inconsistent, or caused by subtle translation artifacts rather than real behavioral shifts.
Oversimplified scenarios: Text-based prompts can miss the nuance of real-world welfare trade-offs. I will mitigate this by documenting my assumptions clearly, referencing existing welfare benchmarks, and getting specialist feedback.
Benchmark overlap: Other researchers are building animal-welfare evals. To avoid duplicating work, I am keeping this focused specifically on matched bilingual comparisons and the gap between stated concern and final action.
Budget or reviewer bottlenecks: If API costs run higher or specialist feedback takes too long, I will narrow the scope to keep the dataset clean and reproducible rather than rushing out shaky claims.
Even if I find zero difference between languages, publishing the dataset, code, and negative results will still give the community a useful reference point.
I have raised USD 0 for this project or any related research in the past 12 months.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!