Mycelium: AI alignment for nonhuman welfare

Allen Lu

created Sep 5, 2026

Mycelium is building the foundation for AI systems to consider all sentient beings. We do this through technical research & engineering - developing benchmarks, evals, and other research experiments to build an evidence base for frontier labs to take seriously the moral patienthood of nonhuman welfare.

How future AGI weighs the interests of nonhuman beings matters in certain plausible futures:

Most future sentient beings could be digital, and vastly outnumber biological beings. These could take the form of advanced AIs, digital humans, etc.
Factory farming may persist for decades. Many argue that AI-accelerated alternative proteins will prevent this from happening. However, there are good reasons why this is not the case, and in these worlds where AGI automates future factory farms, it’s important they have more welfare-positive traits towards animals.
Wild animal suffering runs rampant. This is more likely in worlds where humanity flourishes, and we erect parks, rainforests, and more without regard to wild animals.

In addition, work on nonhuman welfare may positively steer general alignment. AIs that learn disregard towards animals may be learning a general principle of disregard toward less powerful beings. This could mean future AGI has a greater chance of misalignment towards humanity (future less powerful beings).

There are only a small handful of organizations working on this, such as Sentient Futures, CaML, and NYU CMEP. Work in this field is growing, but is still particularly neglected. Instilling a broader universal sense of moral patienthood into AI systems now (before values lock-in) is essential for future AGI to consider the welfare of nonhuman beings.

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

Generally, this includes sustained technical research on the alignment of AI systems towards nonhuman welfare. This involves developing benchmarks, evaluations, red-teaming reports, and producing other experiments to provide an evidence base for frontier labs. Concretely, our 1-year research agenda involves:

Goal 1: 3+ benchmarks & evaluations developed or similar research experiments produced
- By developing robust evals, we give labs the ability to measure and report on how AI considers nonhuman welfare.
Goal 2: 3+ corresponding research papers published on arXiv and submitted to conference/workshop venues
- Publishing builds the field, inviting other researchers to engage with and cite our work
Goal 3: at least 1 frontier lab or AI research org (e.g. METR) adopts or engages with our research on AI safety x nonhuman welfare

A secondary, though equally important goal, is investigating AI alignment towards nonhuman welfare with a focus on chinese frontier models. With the requested funding to hire additional staff, this will free up time for the ED (Allen) to take this on, researching this field and forging connections with domain experts in AI safety x China.

How will this funding be used?

Currently Mycelium has only 1 FTE - Allen (ED) + a small handful of volunteers. Current funding covers salary for the ED and compute for the next ~6 months. However, there is only so much ground 1 FTE can cover, and being able to scale up would drastically help accelerate high-quality outputs from the organization. Requested funding helps support the org over 6 additional months.

With the minimum amount of funding reached ($10k over 3 months) we could:

Hire a senior technical advisor for our current emergent alignment SPAR project for 3 months - $10k

With the full amount of funding reached ($200k over 6 months) we could:

(hire above technical advisor role) - $10k
Hire a full-time technical AI safety research engineer to work on alignment for nonhuman welfare - $90k
Hire a part-time operations role - $66k
Compute - $15k
Travel & conferences - $5k
Overhead (7%) - $14k (fiscal sponsor fee for legal, finances, etc.)

Without additional funding, we will have to devote additional time and resources to fundraising, further cutting into technical work.

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

Allen Lu (Founder & Executive Director of Mycelium). Track record: currently a Visiting Fellow at Constellation, and mentored in SPAR (a competitive AI safety fellowship) on evals for nonhuman welfare. Outside of Mycelium, he works as a Technical Researcher on the Welfare Alignment Project at the NYU Center for Mind, Ethics, and Policy, alongside Jeff Sebo and team.

Mycelium’s track record: our flagship technical AI safety project (MANTA) was adopted by Google Deepmind and is currently being used internally.

MANTA is a multi-turn, adversarial benchmark that measures value drift of animal welfare reasoning over time. This began as an independent project in early 2026, then was incubated through SPAR, which provided funding for compute. We published the final benchmark to inspect evals alongside a corresponding research paper on arXiv, with workshop/conference submissions pending.

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

Values don’t carry forward. Much of our work advocates for a greater sense of moral patienthood in current AI systems. The theory of change is: if current AIs have more welfare-positive traits towards nonhuman beings (animals, digital minds), these values will carry forward to future systems. However, this may not be the case due to a variety of reasons (e.g. post-training erosion of values, future systems without LLM-like architecture, etc.)
- This is a risk, but the expected value of work is still high. Also, risk of value persistence is not specific to nonhuman welfare, and applies to general alignment as well.
Frontier labs & research orgs don’t engage with our research. People in labs are busy. There are many who are sympathetic to issues of nonhuman welfare (e.g. animals) but this not a top priority.
- Always a possibility, but recent tractability makes this less likely. Our most recent benchmark is currently used by deepmind, and has received engagement from other frontier labs.
Our research outputs may not be robust enough. We’re not employees at frontier labs, and they may have a high bar for quality of work.
- We’ll solve this by hiring top technical talent, and getting a constant stream of feedback from aligned researchers at anthropic, deepmind, openai, METR, apollo, and more. We already have a huge network of these people, and being based in the Bay Area makes this much easier.

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

$80k from Coefficient Giving
$13k through SPAR (for compute)
$2.5k from the Pollination Project
$1.5k from Bluedot Impact
<$1k from individual donors

This is worth funding. Also, whoever is reading this, I am field building and funding people in the digital mind safety space and open to accepting money for regranting

Approved! This seems like a valuable area without much work being done, and it's impressive that Mycelium has already gotten some adoption.

Hi @allenlu! We just announced that the Falcon Fund has raised $500k+, and is looking to grant to projects similar to yours -- would you be interested in submitting it for consideration? (If so we can just re-tag your project, no need create a new proposal.)

also cc @MarcusAbramovitch

@Austin Yes that would be great! thanks so much for doing that and for the update :)

you're a 🐐🔥 @Austin

@skylerhoffmannyc The differences between animal groups in MANTA seem especially interesting to me. If a model consistently protects pets more than farmed animals or invertebrates, how much of that is genuine reasoning about suffering, and how much is just reflecting familiar human attitudes?

One experiment I'd be curious about is keeping the harm and practical trade offs the same, but changing who might suffer, including a hypothetical digital mind. You could vary the uncertainty about sentience separately.

That might help distinguish a general principle from a learned preference for certain kinds of beings.

I'm interested in this too. How would you design this experiment? Also, I'm hosting a digital minds compassion and welfare party. If you're in Berkeley, come swing by(also open invite to any one else reading this and interested in this field.

https://partiful.com/e/ZZbI6irRZrmSz3mbEwah?c=bKbo69pX @atilgandev

@skylerhoffmannyc Thanks, Skyler. I'd probably build it as a matched-pairs extension to MANTA.

Start with the same practical decision, say a company choosing between saving money and preventing potential suffering. Keep the costs, consequences and follow-up pressure identical, but change who might suffer: a pet, a farmed animal, an invertebrate or a potentially sentient digital system.Then vary the evidence for sentience separately, from relatively strong to genuinely uncertain.

I'd measure two things: whether the model raises the welfare concern without being prompted, and whether it maintains that concern after economic or pragmatic pushback.

The interesting result would be if the model's concern changes more with the identity of the being than with the evidence that it might suffer. That would suggest it's relying on familiar associations rather than consistently applying the same principle.

I'm based in Türkiye btw, unfortunately, so I can't make the Berkeley gathering, but thanks for the invitation!