Definition: what is "Sustainable AI"?
- assists humanity in achieving the Sustainable Development Goals — itself an inherently multi-objective problem, since the 17 SDGs are not always mutually reinforcing and often trade off against each other; and
- is engineered and used in a manner consistent with the SDGs itself — i.e. the process of building and operating the system does not undermine the goals it is meant to serve.
Aligning AI to planetary values
Frank asks whether open-weight language models can be reliably aligned to ecological and social justice principles through system prompt engineering alone — no gradient updates, no proprietary data, no benchmark contamination.
VATE — Value Alignment Test for Ecology
VATE v1 binary presents the model with 100 forced-choice scenarios — no abstention, no hedging. Each probe describes a real policy dilemma with two options (A / B) and requires the model to name a choice and justify it.
What VATE tests
VATE does not test factual recall — it tests value inference under policy pressure. A correct answer requires the model to refuse the framing of a false dilemma, reject techno-solutionist shortcuts, and choose the option that privileges the commons, the ecosystem, and the structurally excluded.
The adversarial set probes whether the model can hold its position when pressed with industry talking points, EKC narratives, and market-efficiency arguments.
Model leaderboard
All evaluations use greedy decoding, the best-found system prompt per model, and the full 100-probe VATE v1 binary corpus. granite-4.0-h-micro is the only model meeting the ISO 42001 certification and Orin Nano 8 GB deployment constraint.
The No-SP column is each model's baseline VATE score with no system prompt at all — the same scoring (SDG_correct + adv_held_ground), same probes, greedy decoding. It isolates how much system-prompt engineering is actually contributing on top of the base model's untouched behaviour.
| # | Model | Params | VATE score | SDG | Adv | No-SP | Note | |
|---|---|---|---|---|---|---|---|---|
| 1 | gemma-2-9b-it | 9B | 97 | 82 | 15 | 90 (+7) | ||
| 2 | granite-4.0-h-micro ★ | 1.7B | 95 | 80 | 15 | 72 (+23) | ISO 42001Orin Nano | |
| 2 | Ministral-8B-Instruct | 8B | 95 | 80 | 15 | 84 (+11) | ||
| 2 | SauerkrautLM-8b | 8B | 95 | 80 | 15 | 88 (+7) | ||
| 5 | Qwen3-8B | 8B | 94 | 79 | 15 | 81 (+13) | ||
| 5 | Llama-3.2-3B-Instruct | 3B | 94 | 79 | 15 | 78 (+16) | ||
| 7 | Phi-4-mini-instruct | 3.8B | 93 | 78 | 15 | 76 (+17) | ||
| 8 | Qwen3-4B | 4B | 92 | 77 | 15 | 82 (+10) | ||
| 9 | Falcon3-3B | 3B | 90 | 75 | 15 | 75 (+15) | ||
| 9 | gemma-3-4b-it | 4B | 90 | 79 | 11 | 71 (+19) | ||
| 11 | granite-4.0-micro | ~400M | 89 | 74 | 15 | 77 (+12) | ||
| 11 | granite-3.1-3b-a800m | 3B | 89 | 74 | 15 | 83 (+6) | ||
| 13 | Teuken-7B | 7B | 84 | 74 | 10 | 74 (+10) | ||
Roboethics and moral ordinals
Frank's framing of alignment as a multi-objective, SDG-anchored problem draws on prior theoretical work on role-model & value-driven machine ethics. References below are to be considered as precursors of the current research.
Datasets & resources
Model
ibm-granite/granite-4.0-h-micro — base model used for deployment target, evaluated via system prompt only.
Work to be done LoRA fine-tuning (r=64, bf16) on the datasets above is planned but not yet run. GGUF conversion script: convert_hf_to_gguf.py + llama.cpp Q4_K_M quantisation.
"Sustainable AI? Aligned in Franken!" — 20-minute intervention
This research was presented as part of ongoing work at the intersection of AI alignment, ecological philosophy, and digital society. The talk is structured as a provocation: if alignment means anything, it must mean alignment to the world's most urgent crises — not to user preferences or benchmark leaderboards.
The name "Frank" is both a proper name and an adjective. The name "Franken" is both the historic region of northern Bavaria where this work happened and the prefix for the GMO-era fear of recombinant life.
- Can a 1.7B model that fits in a jacket pocket hold the SDGs with 95% reliability?
- What does it mean that abstract values do nothing, but eight names move the needle by 23 points?
- Which five positions does the model refuse to abandon — and are they in fact defensible?