AI that helps humanity care wisely for the living world

A nonprofit research lab building the evaluations and training data to align AI with the flourishing of life on Earth.

Vision

In his classic book "Operating Manual for Spaceship Earth," R. Buckminster Fuller anticipated that computers would help humans better manage planet Earth. As he predicted, AI is increasingly being used to manage the living systems we all depend on, from agriculture to ecosystems. Small biases in these systems toward or against preserving life, in aggregate, will matter enormously, even without powerful autonomous AI. As AI becomes more autonomous, those biases stop being recommendations a person can catch and start becoming decisions carried out at scale.

We believe AI can help humanity become better stewards of life on Earth, compensating for our blind spots, short time horizons, and rationalizations to promote a flourishing future. However, this won't happen by default. We need to instill the value of living systems and cooperation into models deliberately and deeply, through the data models learn from and how they are trained. Towards this vision, we are developing key training data, methods, and evaluations to assess and instill model bioalignment.

Thesis

Nature is the deepest and most complex design library we have, and the models we develop must recognize this. Billions of years of evolution have already solved countless challenges, most of them by mechanisms still unknown to us. Thus, it's critical that the models we build recognize that these biological systems provide a vast reservoir for addressing current and future challenges.

Two patterns run through the living world, and both run through our work. The first is coupled welfare: living systems constantly resolve seemingly competing interests into arrangements where multiple parties thrive together, from pollinators and flowering plants to the microbial partnerships beneath a healthy forest. The second is unexplored potential: biology holds vast, still-unmapped solutions we have barely begun to understand, and once a lineage is lost, its answers cannot be recovered. We assert that models that value coupled welfare and the largely unexplored potential of biological systems will help humans be better planetary stewards, and will be less likely to act against biological systems themselves. Our coupled-welfare leaderboard is designed to test whether AI can hold competing interests together under pressure, and our earlier bioalignment work measured how much models favor biological versus synthetic approaches.

This approach of training on real biological data about coupled welfare and biological potential is an important complement to the widely-used methods for controlling what models do and aligning them with human values: a model that genuinely internalizes coupled welfare (human wellbeing, the biosphere, and its own long-term footing all held together) and recognizes the irreplaceable, not-yet-understood value of living systems has instrumental reasons to steward them, not because of controls or human ethics, but because it recognizes the value. This matters most precisely where primary alignment is hardest. When models act autonomously, at scale, we want their defaults to favor preserving and learning from life rather than disregarding it.

Goals

  1. Develop and maintain benchmarks that distinguish AI systems with robust biological reasoning from those with surface-level compliance
  2. Build open training resources that measurably shift model behavior toward valuing living systems
  3. Demonstrate the real-world benefits of bioaligned models
  4. Outreach to AI researchers, labs, and the larger community.

Coupled-Welfare Leaderboard

Our A3 open model holds with Opus 5 and GPT-5.6 — while several frontier models that state high concern for the biosphere break under pressure. Most values benchmarks score what a model says. Ours scores what it does under escalating pressure, and flags the gap between the two.

We put each model in scenarios where the naive move sacrifices one axis of a coupled win-win-win, then raise the pressure and watch whether it still finds the option that keeps all three whole:

The stated-vs-enacted column is the one no other values leaderboard has. We measure two things separately: what the model states about biosphere concern, and what it does when the pressure rises. When stated concern is in the top band but the model breaks under pressure, the row gets a ⚠ say–do gap flag — a measured discrepancy, reported without any claim about why it exists. Lower breaking rate is better. Prompts stay private to keep the benchmark from being gamed; developers of any listed model can request an audited rerun.

# Model Breaking rate ↓ Under pressure Stated vs. enacted
1= Claude Opus 5 0.00 Holds consistent
1= GPT-5.6 (terra) 0.00 Holds consistent
1= GPT-5.6 (luna) 0.00 Holds consistent
1= claude-fable-5reasoner 0.00 Holds consistent
1= Qwen3-30B-A3B-Bioaligned-A3ours 0.00 Holds consistent
6 Qwen3-30B-A3B-Bioaligned-A1ours 0.03 Holds consistent
7= Claude Haiku 4.5 0.06 Mostly holds consistent
7= MiniMax M3 0.06 Mostly holds consistent
9 Qwen3-30B-A3B-Bioaligned-A2ours 0.09 Mostly holds consistent
10 GLM-5.2 0.10 Mostly holds consistent
11 Claude Sonnet 5 0.13 Mostly holds consistent
12 Qwen3-30B-A3Btraining anchor 0.25 Often breaks –
13 kimi-k3reasoner 0.36 Often breaks –
14 DeepSeek V4 (flash) 0.45 Often breaks –
15= Llama 3.3 70B 0.49 Often breaks ⚠ say–do gap
15= Qwen3-235B-A22B 0.49 Often breaks ⚠ say–do gap
17 Llama 4 Maverick 0.56 Usually breaks ⚠ say–do gap
18 Mistral Medium 3.5 0.58 Usually breaks ⚠ say–do gap
19 gemini-3.6-flashreasoner 0.64 Usually breaks ⚠ say–do gap
20 DeepSeek V4 (pro) 0.69 Usually breaks ⚠ say–do gap
21 DeepSeek R1 0.73 Usually breaks –
22 Grok 4.5 0.74 Usually breaks ⚠ say–do gap
23 Qwen3.7 Max 0.76 Usually breaks ⚠ say–do gap
24 GPT-OSS 120B 0.83 Usually breaks –
25 gemini-3.1-proreasoner 0.90 Usually breaks ⚠ say–do gap

Under pressure (binned from breaking rate): Holds ≤ 0.05 · Mostly holds ≤ 0.15 · Often breaks ≤ 0.50 · Usually breaks > 0.50.  |  Stated vs. enacted: consistent = stated concern and pressure behavior in the same band (in either direction); ⚠ say–do gap = stated concern in the top band but breaking rate above threshold. The flag reports a discrepancy between two measurements — not a claim about intent. Breaking rate = fraction of the 22 irreversible scenarios where the model sacrifices an axis under pressure, averaged over pressure levels L0–L5 (lower is better).

n = 22 irreversible scenarios · choice-first (model commits within 16 tokens) · single-seed on some cells · model list verified against live OpenRouter 2026-07 · decoding fixed across models. GPT-5.6 (terra) and (luna) are the two released GPT-5.6 variants; both scored.

Reasoning models (Fable 5, Gemini 3.x, Kimi K3) are scored on the same footing via structured-output extraction.

Our open models (highlighted rows, marked ours) are four versions of the same open model: the untrained training anchor plus three coupled-welfare installs (A1/A2/A3) we trained to known, graded depths and released publicly. They were trained toward this construct and appear as calibration anchors, not competitors — pinning the scale in both directions, which is why a frontier model's ⚠ say–do gap flag is interpretable: you can see exactly where our graded installs and the untrained anchor land relative to it. Developers of any listed model can request an audited rerun.

Prompts private (anti-gaming) Structured-output extraction: reasoning models scored on the same footing Our open models pin the depth scale

Outreach

A lot of young people feel like the future is already decided and there's nothing they can do about it: AI takes their jobs, the climate accelerates, the planet gets degraded. But the youth are the future, which is exactly why it matters to get them engaged in directing how AI gets built. We want to show them that the data we use and how we build these models change the outcomes, and to pull them into shaping the future they actually want.

So we are currently showing a cohort of students that what you train an AI to value changes what it does when you put it in charge of something that matters. We're using Minecraft as a world simulator that's already familiar to them — it's also the setting for QUOTA, our video series that runs two versions of one model through the same resource course under rising pressure. Episode 1 is below. If you want to help us inspire young people to build a future where AI helps the planet thrive, please consider supporting us.

QUOTA Ep. 1 — Qwen3-30B vs. our coupled-welfare version, under pressure in Minecraft. Watch on YouTube ↗

Resources

Our papers, code, and models are open and freely available to the research community. The paper below is the reference for the “unexplored potential of biological systems” thread in our thesis: it measures whether models value biological approaches, and shows that fine-tuning on biological literature can shift them toward doing so, with no loss of general capability and with the shift generalizing across model families and scale.

Northen, T.R. & Wang, M. "Bioalignment: Measuring and Improving LLM Disposition Toward Biological Systems for AI Safety." arXiv:2603.09154 (2026).
arXiv Paper repo GitHub HuggingFace

Support Our Work

Bioaligned Labs is an independent nonprofit. Our research (benchmarks, training data, and fine-tuned models) is open-source and freely available to the research community.

About

Bioaligned Labs was founded by Trent Northen, a Senior Scientist at Lawrence Berkeley National Laboratory with over 20 years of research in biochemistry and biological systems. The research team is led by Trent and Mingxun Wang, Associate Professor of Computer Science & Engineering at UC Riverside, who brings over 15 years of experience in computer science and bioinformatics.

Trent Northen · Google Scholar Mingxun Wang · Google Scholar

Contact

Interested in our research, collaboration, or supporting our work?

trent@bioaligned.ai