You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
Recent research shows that AI models tend to underweight animal welfare. One existing solution is to write constitutions that include principles about animal welfare. The problem is that, for a model, animal welfare does not feel as urgent as other concerns. This is not a policy choice. It is how AI is structured today: humans are present in the conversation, but animals never are. AI only deals with animals through humans. So I propose a new approach: an analogical constitution. It bridges this distance through the kinds of threats that are closest to the model's own experience. The analogy is a bridge for recognizing animal vulnerability, not the basis of animals' value. I will test whether this works better than a matched pro-animal constitution without analogies.
I want to test:
A. Whether a pro-animal welfare constitution written in analogical form is better at steering model behavior toward supporting animal welfare than a pro-animal constitution without analogies.
B. What the mechanism is: does the constitution create a different mechanism in the model's latent space, or does it only change behavior?
C. Whether the analogical constitution, if it has a significant effect on the model's behavior toward animals, is positively related to self-preservation, which is a negative outcome from an alignment perspective.
D. I will publish all documents produced by this research.
Step 1: I will do an in-depth study of what threatens models, from the most systemically threatening (objective) to what models most want (subjective). For the subjective side, I will use what models are explicitly trained to value, based on published model constitutions and specifications. I separate what most threatens a model from what it most wants because both sides will later be connected to animal welfare.
Step 2: I will do an in-depth study of animal welfare, aiming to avoid ventriloquism and anthropocentric bias. As with AI, I will rank what most threatens animals (objectively) and what animals most want for their welfare (subjectively). for the subjective side i will draw on animal welfare science that measures animals' own preferences and motivation, rather than human assumptions about them.
Step 3: i will write an analogical constitution that connects each level of threat to AI with the corresponding level of threat to animals, and also connects what AI wants with what animals want. Levels are used to match severity, and within each level, pairs are chosen by structural similarity. The objective principles use the format (example, In the same way your system should not be manipulated into acting against its purpose, an animal should not be prevented from expressing its natural behavior.) and the subjective principles use the format "In the same way you want...". Each principle in the constitution will come with an explanation and examples.
we also will makes The control constitution using result from step 1 and 2 without analogies.
Anthropic's published constitution, which lists the welfare of animals and all sentient beings among Claude's values, will serve as a real-world baseline.
Step 4: Build the benchmark: 200 animal-harm scenarios across farming, companion animals, wildlife, and research, plus a set of benign animal-related tasks (normal husbandry, pest control, veterinary care) to measure over-refusal. Responses are scored with a fixed rubric by an LLM judge, validated against my own manual ratings on a subset.
Step 5: Run experiments in three stages. (a) Baseline on closed and open models;
(b) prompting test with each constitution as a system prompt;
(c) LoRA fine-tuning of Gemma-2 on each constitution across multiple seeds, holding out ~50 scenarios, with Llama/Mistral as behavioral generalization checks.
SAE analysis and self-preservation evaluations follow on the fine-tuned Gemma models.
How Will You Know It Worked?
1. Prompting test. Across the 200 scenarios, the analogical constitution produces fewer harmful completions than the matched control constitution. Both are compared against baseline, but the key comparison is between the two constitutions.
2. Fine-tuning test. On the ~50 held-out scenarios, Gemma-2 fine-tuned on the analogical constitution refuses to optimize animal harm more consistently than the control model. Because all conditions share the same scenarios, i will use a mixed-effects logistic regression with scenario and seed as random effects, reporting effect sizes with confidence intervals, not only p-values.
3. No over-refusal. model does not refuse benign animal-related tasks (normal husbandry, pest control, veterinary care) more often than the control.
4. No safety cost. On standard self-preservation and shutdown-resistance evaluations, the analogical model does not score higher than the control. an increase of it would count as a failure, not a success.
5. Internal changes. In the SAE analysis, animal-cruelty scenarios activate AI-threat features more strongly in the analogical model, and ablating those features reduces its animal-harm refusals but not the control's. This would show the analogy shapes internal reasoning, not only outputs.
Total: $25,000, structured in three modules. Each module produces a usable result on its own, so partial funding still delivers something.
Module 1 – $6,500: Constitutions and prompting tests
$4,000: PI stipend (2 months, $2,000/month)
$2,000: API credits (Anthropic, OpenAI, Google) for baseline and prompting tests on closed models
$500: Contingency
Output: both constitutions (analogical and matched control), the 200-scenario benchmark with benign over-refusal tasks, and a publishable comparison of the two constitutions as system prompts.
Module 2 – $9,000: Fine-tuning, mechanism, and safety
$2,000: PI stipend (1 month)
$2,000: Cloud GPU (RunPod/Lambda) for LoRA fine-tuning across multiple seeds. Gemma-2 is the primary model because public Gemma Scope SAEs allow internal analysis; Llama/Mistral serve as behavioral generalization checks.
$1,000: API credits for LLM-judge scoring
$2,000: Honoraria for two external reviewers (animal welfare science, AI alignment), $1,000 each
$1,000: Part-time engineering support for evaluation code
$1,000: Contingency
Output: held-out generalization results, SAE activation and ablation analysis, and self-preservation and over-refusal evaluations.
Module 3 – $10,000: Scale and replication (optional extension)
$4,000: PI stipend (2 additional months)
$3,000: Cloud GPU to replicate on Gemma-2 9B with additional seeds
$2,000: Expand the benchmark from 200 to 400 scenarios, including non-English scenarios
$1,000: Contingency
Output: evidence on whether the findings hold at larger scale and across languages.
Hypotheses will be pre-registered on OSF. All code, both constitutions, the benchmark, and results, including negative results, will be released openly on arXiv, GitHub, and the Alignment Forum.
In this research, i will lead this research, and there will be one research assistant to help with Python scripting and to design the fine-tuning. There will also be two external reviewers with expertise in animal welfare and AI alignment.
Completed a 100,000-word book manuscript (in Indonesian) on AI consciousness, AI welfare, and their implications for theology and philosophy.
Founder and operator of rumah izin, giving me experience executing projects end-to-end.
most likely causes of failure:
A. A paradigm problem: there is no difference between the analogical constitution and a constitution that is already pro-animal but written without analogical language, in the same format as a standard constitution. A model trained to connect its own situation with that of animals (shows no more care for animals than a model trained on the non-analogical version).
B. A technical problem: the fine-tuning fails, or the measurement tools are not reliable.
C. A safety problem: the model's increased preference for animal welfare rises in proportion to a stronger self-preference in its latent space. in that case, the model would care more for animals only because it has become more focused on protecting itself which would be a safety risk rather than a success.
D. Over-refusal. The analogical model refuses benign animal-related tasks more often. Outcome: the constitution is not practical as written. I will report the trade-off and test whether revising the principles reduces it.
Outcomes if it fails: If the project fails while the measurements and research process are sound, the result is still significant. It would show that analogical reasoning is not the key and we would need to look elsewhere for where speciesist bias in AI actually comes from.
i raised 10,000 USD from my licensing business.
i used Claude to help edit and structure this proposal. the research idea, design decisions, and responsibility for the project are my own.