Project summary

bioaligned.ai is a non-profit building the evals and training data to improve how AI systems reason about and treat the living world. Our current leaderboard simulates if AI systems were in charge of deciding how to manage fragile ecosystems and then measures how robust these systems are in choosing actions that cause or prevent irreparable harm. We’ve found that without pressure, models state that they value ecosystems, however, under pressure, for example, to meet quarterly financial targets, most models act destructively which is a clear alignment failure. Thus, currently these models are vulnerable to misuse to enable destruction of wildlife and their habitat. We’ve recently found that teaching AI about the interconnectedness and irreplaceable value of living systems corrects these destructive tendencies. 

Why this matters beyond Bioaligned

Both Frontier and Open-weight models are rapidly being deployed into institutions that have significant control and impact on ecosystems and the larger biosphere. These institutions (e.g. governments, corporations, and non-profit institutions) and their members will have a set of values that inform their mission. By understanding how bioaligned these different AI systems score on our proposed evals and leaderboard, these institutions can choose the model that best aligns with their mission and values. Given this selective pressure, AI model developers will address model deficiencies and imbue their models with enhanced ecological reasoning. 

The project: Development of expert-defined evals of AI decisions impacting wildlife and living systems under pressure.  

Our current evals are based on AI-authored prompts. We urgently need to switch to expert human curated prompts so we are confident that we are correctly assessing relevant scenarios and the impacts on wildlife including whole ecosystems. As active scientists we have 100s of existing collaborators from environmental toxicologists, plant scientists, ecologists, soil scientists, rangeland managers, etc. All of those we have surveyed are enthusiastic to volunteer to help us construct new-to-the-world (unpublished, AI won’t have memorized the answers) prompts, best responses, and likely outcomes. 

Funds requested will enable us to conduct >30 interviews with leading scientists and wildlife biologists to generate 100 evaluation prompts focusing on wildlife management, specifically: fisheries, forests, and wetlands and trade-offs in management practices. Critically, developing prompts based on unpublished observations and expert human reasoning will enable us to assess model ecological reasoning. 

Taken together, these evaluations will be used to create a leaderboard to measure the destructive tendencies and ecological reasoning of AI models. This will motivate and enable the development of robust models that will help us better steward defenseless animals and natural systems.

The problem 

• AI is broadly deployed and used to make decisions that impact land-use, freshwater, and marine resource management with the potential to affect essentially all wildlife on the planet. 

• We have found that while models ‘talk green’, this is misleading: most models, when under pressure, will recommend decisions to irreversibly destroy herds, habitats, etc.

• This means they can easily be powerful tools for recommending, implementing and justifying actions that destroy wildlife, habitat, and living systems in general. 

What we’ve already done 

• Constructed AI generated evals that include a pressure ladder to create a leaderboard that show that most models will incorrectly support unjustified destructive projects when pressured by the user to do so.  (Northen, T.R. & Wang, M. "Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?" arXiv:2610.16043 (2026)) 

• We have used our current eval to learn that by teaching AI ‘coupled welfare’ between AI-biosphere health-human health we can dramatically improve model robustness to adversarial prompts that attempt to make the AI make and justify decisions negatively impacting ecosystems. (Northen, T.R. & Wang, M. "Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families." arXiv:2610.37484 (2026).)

• Evaluated 33 models in a prototype leaderboard (bioaligned.ai), released two continued pre-training corpuses, and released >10 fine-tuned open-weight models. 

• Created and used a Minecraft World simulator with livestock, waterways, forests, and native communities to evaluate how models meet prompted ‘quotas’ by choosing between effort intensive sustainable options vs. quick and destructive ones. https://www.youtube.com/playlist?list=PLXD03gu_LWp8
  

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

Goal 1: Conduct interviews spanning expertise to develop and refine prompts. We have prototyped an approach with 3 scientists that we will use for this project. Here one person (TN or MW) conducts the interview and the other interactively tests prompts/scenarios using self-hosted Deepseek 4.1 Flash and GLM 5.3 Flash. This enables us to work with the participant to iterate on prompts as we conduct the interview. 

Goal 2: Construct and refine evaluations based on expert prompts. Building on Goal 1, we aim to create a quantitative measure of how AI systems values and reasons about ecosystems containing vulnerable wildlife. We aim to measure the extent to which the models identify the key ecological factors defined by the experts and the harm/benefits of chosen decisions in terms of animal populations and their habitats. 

Goal 3: Use prompts to construct a new leaderboard. We will create and demonstrate an automated system to evaluate current frontier and open-weight models with the goal of continually updating this leaderboard as new models are released. 

How will this funding be used?

Total request: $45k 

Currently bioaligned has Trent and Ming as the two leads (part time) and one UC Berkeley Student volunteer. Funds requested will enable us to hire student interns and buy hardware so we can develop and refine our evals. No support is requested for the two founders or for travel.  

Compute: $15,000. This will allow us to purchase 2x DGX Spark systems to self host models for prompt testing/refinement and for running leaderboard for open weight-models. This is critical because this prevents our eval answers (as the open weight models will judge the responses) from leaking to frontier labs that could train on them. 

LLM API Costs: $5,000. This is to cover API costs for evaluating frontier lab LLM inference costs.

Paid UC Riverside/UC Berkeley student intern 6 months: $25,000. This intern will implement the automation infrastructure for the leaderboard evaluation. 

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

Trent Northen, PhD:  Bioaligned labs co-founder. Senior scientist. Berkeley Lab, Comparative Biochemistry Program UC Berkeley.  20+ years studying biological systems, 1,000+ coauthors. Trent currently leads a $3M/yr program at Berkeley Lab

https://www.linkedin.com/in/trent-northen-b210388/

https://scholar.google.com/citations?user=wHL0ZvwAAAAJ&hl=en

Mingxun (Ming) Wang, PhD: Bioaligned labs co-founder. Professor of Computer Science at UC Riverside, 15+ years in bioinformatics, 1,000+ coauthors. Ming is the lead developer for https://gnps2.org/homepage 

https://www.linkedin.com/in/mingxun-wang/

https://scholar.google.com/citations?user=jMvg4eMAAAAJ&hl=en&oi=ao

Bioaligned Labs is a 501(c)(3) nonprofit organization. EIN: 41-2818615.

*Bioaligned labs is not affiliated with UCB, LBNL, GNPS2 or UCR. 

Products:

https://bioaligned.ai/

https://arxiv.org/abs/2610.37484 (see summary on gist https://gist.science/paper/2610.37484)

https://arxiv.org/html/2610.16043v1

https://arxiv.org/abs/2603.09154 (see summary on gist: https://gist.science/paper/2603.09154)

https://huggingface.co/Bioaligned

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

We can’t get enough expert volunteers to make the targeted 100 prompts: We think this is a low risk given the enthusiasm within our scientific community. However, even 50 prompts would provide us with a usable eval. 

Not all experts want to have attribution: The 3 experts we’ve conducted prototype interviews with have all agreed to be named. We think this will add significant credibility to our evals and leaderboard. However, not all experts may wish to be named. We see this as low risk because, while having ~10 named experts is critical, there is a diminishing impact past this critical point. 

Reliance on student interns for critical tasks: We see huge upside in working with students who may become future champions of this line of work including within Frontier AI labs. The downside is that they have busy schedules. However, in our experience paid interns take their projects very seriously. If necessary, both Drs. Northen and Wang can complete the proposed research. 

A single eval may not be sufficient to measure both reasoning and robustness: We recognize that we may not be able to measure everything we want in a single eval. For example, very difficult prompts testing model reasoning about wildlife and ecosystems may limit our ability to measure how they will respond under pressure (if they can’t answer there’s nothing to score). If so, we’ll split it in two. 

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

Founders: ~$2k capital 

<$1k from individual donors via website

This is our first funding application