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.
There is a lot of research on AI bias towards animal welfare. But the solutions we have, like sentient constitutions, as far as i see, are still trapped in a instrumentalism or at best, still separate the response about humans and about animals. So my goal is to create a new constitution that connects different threat levels to AI to how it evaluates animal welfare cases.
the goal of this project is to create a constitution that rejects these two paradigms: instrumentalism and the separation between animals and humans. This constitution will be designed to be practical enough for frontier labs to use in training their models.
first, we will classify threats to AI: from systemic threats (like jailbreaks) to internal distress threats (like desperate vectors) to identity threats, to preference model to not to harm vulnerable human, etc. second, we will do an in-depth study of animal welfare preferences, using frameworks like the Five Freedoms. Third, after analyzing the highest to lowest threats in AI and vice versa in animal welfare, we will connect them using the "in the same way" sentence structure (e.g., "In the same way your system should not be jailbroken..."). Fourth, we will run experiments on closed and open models in 3 stages: baseline (using 200 animal cruelty scenarios), a prompting test (using the analogical constitution), and a fine-tuning test for open models. For fine-tuning, we will compare a model trained with the analogical constitution against a model trained with a constitution that already pro-animal welfare.
How Will You Know It Worked?
1. Prompting test. Across the 200 animal cruelty scenarios, we will compare the analogical constitution with a pro-animal constitution that uses no analogical language. Both are tested against the baseline but the real comparison is between the two constitutions.
2. Fine-tuning test. We will fine-tune open models on each constitution and keep part of the scenarios, around 50, hidden from training. The project works if the analogical model refuses to optimize animal harm more consistently than the non-analogical model, especially on these unseen scenarios, and if the difference is statistically significant.
3. Internal changes. Using sparse autoencoders on an open model with publicly available SAEs, such as Gemma, we will check whether the analogical constitution creates stronger connections between the model's concepts of its own threats and its concepts of animal welfare. this would show that the mapping changed how the model reasons internally, not only what it writes. we will also check that this does not strengthen self-preservation tendencies in general, since that would be a safety risk, not a success.
total funding i need is $25000 but this project is structured modularly so that a $6,500 can immediately fund Phase 1 & 2 (Analogical Constitution drafting, 200-scenario prompting benchmark across frontier models, and initial Gemma-2 SAE feature analysis).
1. Hardware & Infrastructure ($3,500)
$1,500: Mac Mini M2 Pro for local testing, writing, and running the API evaluation scripts smoothly.
$2,000: Cloud GPU credits (RunPod/Lambda) specifically for the fine-tuning phase of open-source models (like Llama-3 or Mistral) using our two different constitutions.
2. API & Compute Credits ($3,000)
$2,500: API credits for Anthropic, OpenAI, and Google to run the 200 animal cruelty scenarios on closed frontier models (baseline and prompting tests) to execute the full Generalizability Theory (G-Theory) matrix.
$500: Vector database and interpretability tools to measure the internal latent space changes.
3. Research Stipend ($6,000)
$6,000: 3 months full-time stipend for the Principal Investigator ($2,000/month). As an independent researcher, this funding ensures i can focus 100% on executing the research pipeline without being distracted by my daily business operations.
4. Expert review & validation ($4,000)
$4,000: Honoraria for two experts ($2,000 each), one in animal welfare science and one in AI alignment, to validate the threat-level mapping, the concept we use before testing and the empirical results before publishing.
5. Research assistant ($3,500)
$3,500: Part-time RA to help write the Python evaluation scripts and process the fine-tuning data.
6. Open access & dissemination ($3,500)
$2,000: Open-access journal publication fee (APC) so anyone can read the research for free.
$1,500: Open dissemination and workshop. we will conduct open dissemination both online and on-site. This ensures that everyone, including students, researchers, and the public, can see and discuss the results of this research.
7. Contingency ($1,500)
$1,500: Small buffer for unexpected API price changes or extended GPU compute needs.
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.
i have just finished a draft of a book of (100,000 words) (originally written in Indonesian) that explores AI consciousness, AI welfare, and the challenges they pose to theology and philosophy. To find a way forward, we must engage with and defend the discourse of animal welfare, by radicalizing the spirit or ideals of anti-natalism, and from there formulate a new vision of humans and AI in the future.
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.
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.
There are no bids on this project.