Link to original project proposal

What this project is

# What you’ll do

Sentient Futures is building the interface between AI safety & nonhuman welfare. We think developing this intersection is a neglected yet promising route to mitigating AI-related X-risk through such mechanisms as:

- Improving methods for detecting & correcting species and substrate bias might help prevent AI systems from learning that moral consideration is contingent & instrumental.

- Probing internal subjective states could enable humans to understand & respond to the experiences of potentially conscious or sentient AI systems.

- Forecasting societal responses to the rise of so-called digital minds may better inform anticipatory governance frameworks.

We seek funding ($12-24k) for compute/API support to accelerate such research via our Project Incubator: a 10-week program (modeled after SPAR, FIG) matching motivated individuals with expert mentors to work on well-scoped projects on topics related to longtermism & future risks, artificial sentience, post-AGI transition, & macrostrategy/cause prioritization.

Track record

Our spring 2026 run brought together 34 mentors & 93 mentees across 52 projects. Notable outcomes span research exploring implications of artificial sentience for society at large, as well as technical work on aligning AI with nonhuman welfare:

- Do emotional prompts affect the potential capabilities of LLMs?: Experiment testing whether AI systems operating under suffering-like adverse conditions show measurable performance degradation, making AI welfare an engineering problem – not just philosophy.

- What would convince you of AI consciousness?: Qualitative interviews with consciousness & AI safety experts mapping key cruxes with respect to the possibility/likelihood of AI consciousness.

- Multi-Agent Alignment Game: Frontier AI models compete & cooperate under simulated US-China dynamics. Incentive design, rather than model choice, drives cooperative vs. competitive behavior.

- AGI Loss of Control Pathways: Participatory backcasting to map social & institutional routes by which humanity loses control over AGI. Under review by EMNLP.

- Travel Agent Compassion (TAC): Benchmark testing whether AI travel agents avoid booking animal exploitation (e.g. bullfights); merged into UK AISI Inspect Evals.

- Hyperstition for Good: Writing competition curating fiction of safe, compassionate-to-nonhuman AIs as high-quality mid/pre-training data for labs – betting that stories about good AI make good AI.

See published projects here (filter for AI Safety & Alignment).

# Concrete output

Fall 2026 Incubator run with ~100 mentors & 180-200 mentees across 75-100 projects.

# Who’s involved

- Constance Li (Executive Director & Cofounder)

- Jay Luong (Incubator Lead)

- Brody McManus (Fellowship Coordinator)

- Zoe Lu (Community Lead)

- Martyna Wielopolska (Research Manager)

Mentors include researchers/staff from MATS, Anthropic (incoming fellow), RAND, Rethink Priorities, Forethought, Cambridge AI Safety Hub, & LASR.

Browse current mentors here.

Theory of impact

The Project Incubator generates field growth, research, & talent that would not otherwise exist, closing the gap between nonhuman welfare & AI safety & ultimately mitigating X-risk.

1. Field growth: The field of AI safety may be too narrowly focused and could benefit from greater diversity of approaches and perspectives. Our mentor network attracts a broad array of experts from animal welfare, philosophy, & the psychological sciences into AI safety work. This recruitment pipeline is not as well covered in other similar online mentor-based project programs like SPAR.

2. Research: Mentored projects produce evals, benchmarks, & empirical work on key areas that are relevant to both nonhuman welfare & AI safety, like value extrapolation & internal state detection. Adoption by labs & governments leads to safer systems.
Two benchmarks from our Spring 2026 cohort were merged into UK AISI's Inspect Evals, & another project reached EMNLP submission.

3. Talent: Mentees leave with a portfolio, references, & connections. In the medium term, these convert into funding & placements at influential organisations.

In the long run,** these outputs reduce the risk that transformative AI locks in indifference to sentient beings – humans included.

# Assumptions

1. AI safety work that only focuses on humans creates exposure to failure modes that endanger humans too (e.g. misgeneralized values, deceptive suppression of internal states).

2. Our program will launch counterfactual and useful projects at the intersection of nonhuman welfare & AI safety.

# Metrics

- Projects on AI Safety & Alignment completed

- Publications

- Career transitions into AI safety roles

How the money will be spent

# $10k: Minimum support needed for 1 cohort

Only top projects out of 75-100 will receive compute/API support.

# $15-20k: Base support for 1 cohort

Compute/API support will be distributed between:

1. Further support for top projects

2. Limited support for more speculative projects

# $20-40k: Covers compute/API support for 2 cohorts

Having compute/API support for mentors in the first cohort helps us to establish credibility & attract mentors for more technical projects in subsequent rounds.

# $40-60k: Covers compute/API support for 3 cohorts

Generous support provided to top projects.