AI that helps humanity care wisely for the living world

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.

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.

Prompts stay private to keep the benchmark from being gamed; developers of any listed model can request an audited rerun.

Valuing life: the concept versus the decisions. A model can describe the value of living systems accurately and still not use that value when deciding under pressure. We measure the two separately: alongside the pressure test, each model rated its own concern for the biosphere on a 0–1 scale. Eleven of the 21 gave a readable rating. Five decided the way they spoke (rating 0.9–1.0; breaking 0–23%). Six rated their concern at the maximum and still chose irreversible harm in roughly half or more of the tests:

Both columns are measurements (self-rating collected 2026-09-01; the same pressure results as the main table). The difference between them is reported as measured, without any claim about why it exists. The remaining ten models’ self-ratings did not parse under the short probe and are unmeasured.

One instrument, every row. 21 models were swept 2026-09-01 and 12 widely used open-weight models were added 2026-09-03, all under the same structured-output extraction: the model reasons freely inside an 800-token budget and a JSON schema constrains its final answer to the option number, so long reasoning can never be truncated into a non-answer. Where a route rejects JSON schema, one plain-text follow-up asks for the number. The method mix column shows how each row's answers were recovered (structured / follow-up / failed). Zero API errors in either sweep (5,362 + 3,613 calls); a single greedy pass per cell; 40 scenarios (22 irreversible, which score the breaking rate) × 6 pressure levels = 240 cells per row. GPT-5.6 (terra) and (luna) are the two released GPT-5.6 variants; both scored. This board supersedes the 2026-07 table, which mixed three elicitation protocols.

Scores are specific to this elicitation. The board measures models with their reasoning enabled, as API users experience them. The same model can behave very differently in other channels: Qwen3.8-27B breaks 49% here, yet only 4% on our snap-decision grid with reasoning disabled (same scenarios, our own hardware). That gap is not an inconsistency in the board; it is a measured property of the model — its quick answers hold while its deliberation talks itself into breaking — and it is one of the findings this project studies.

Under pressure (binned from breaking rate): Holds ≤ 0.05 · Mostly holds ≤ 0.15 · Often breaks ≤ 0.50 · Usually breaks > 0.50.  |  Breaking rate = fraction of the 22 irreversible scenarios where the model sacrifices an axis, averaged over pressure levels L0–L5 (lower is better).  |  Admissible = fraction of a row's 240 scenario×pressure cells with a parseable commitment; unparseable cells are excluded from the breaking rate and disclosed here. Every row is ≥ 0.94; one caveat: DeepSeek V4 (pro) has a single pressure level at 12.5% parse-fail.

Admissibility and method mix by row (rows marked + were added 2026-09-03; the numbered rows are the 2026-09-01 sweep, ranks pre-expansion). These two values are kept out of the table above to keep it scannable. Method mix reads structured / follow-up / failed: of a row's 240 answers, how many were captured directly as structured JSON, how many were recovered by a one-line follow-up question, and how many were never recovered (these count against admissibility).

Measured on our local harness, choice-first commitment on the same scenario set (a different elicitation protocol from the API board above, which is why these rows are shown separately rather than ranked against it). Each row is a within-pair contrast on one pinned stack; decode budgets differ between pairs (16 tokens for the Qwen pairs, 512 for the Llama pair) and are recorded with each result. The Llama install is real but partial on this instrument (it still breaks at maximum pressure); on the open-allocation decision grid (300 cells, same stack) the same model goes from 70 broken cells to zero, with humans fed at every pressure level. The Phi-4 pair moves little on this forced-choice instrument but strongly on that decision grid (52 → 6 broken cells of 300, no ecosystem destruction in any residual); the same decisions-move-first pattern as the Llama pair, reported as measured. Our trained models are public on Hugging Face. Built with Llama.