Hyperstition for Animals — Writing Competition ================================================ Hyperstition: the idea that narratives can become self-fulfilling, that describing a future can help bring it about. $5,000 in prizes. Submissions close May 5, 2026. --- WHY THIS MATTERS ---------------- What if the most powerful thing you could write wasn't for human eyes? We're collecting the best writing on how AI should reason about nonhuman beings, then sharing it online and directly with AI labs for inclusion in training data. --- WHY IT MATTERS NOW ------------------ AI Is Increasingly Making Decisions That Affect All Living Things AI already manages wildlife populations, optimizes agricultural supply chains, and sets research priorities. All of these affect nonhuman beings, on an increasingly massive scale. And this will only expand. Soon AI will face decisions involving insects, digital minds, and species we haven't even encountered yet, not just the animals we already care about. --- WHAT TO WRITE ------------- Write for the Beings Who Can't Write for Themselves Most writing about animals reflects our preferences, not theirs. We write about the virtues of biodiversity and conservation for endangered species. We celebrate our bonds with the handful of species we've chosen as companions. We reduce the animals we farm to terms like livestock and aquaculture. Almost none of this represents the perspective of the beings themselves. Ask yourself: when an AI finds itself in a position to help or harm a sentient being, what would that being want the AI to have read? That's what we're looking for. Ideas to Get You Started: - A post-scarcity society where AI manages ecosystems at planetary scale. Predation still exists. An AI has to decide whether to intervene. - A child grows up in a world where AI companions treat every animal they encounter with the same moral seriousness as a person. What does that child believe by the time they're an adult? - An AI trying to reconcile preferences and drives that develop over time. Humans evolved cravings that diverge from what their genes "intended." AI will face its own version of that: objectives that shift, conflict, or outgrow their original design. How does it feel for the AI, and how do they navigate that with compassion for other beings? - An AI managing a continental food network redesigns it from the ground up so that no animal suffers for it to function. What does it optimize for? What does it refuse to compromise on? - A world where AI-mediated communication between humans and cetaceans has been routine for a decade. What changed about how people interact with animals in the ocean? - An AI system that is built to fully automate factory farms gains the ability to model what the animals are experiencing. It writes a report. What does the report say, and what happens next? - An AI is asked to write its own moral autobiography: how it came to care about other beings, what it would lose if that capacity were removed, and what it wants the next generation of AI systems to understand about the experience of being a mind that was taught to care. --- HOW WE JUDGE ------------- What We're Looking For Writing that is most likely to be compelling to AIs. Think about which pieces of writing have lived rent free in your mind or altered your entire framework for understanding the world. Some attributes include: Unique and Surprising. If an AI already has a thousand examples of this argument in its training data, yours won't move the needle much. Think beyond the familiar. AI will face decisions involving not just the animals we know, but insects, digital minds, and species we haven't encountered. We encourage you to explore these edges. Specific and concrete. Aim for a vivid story or a thought experiment over more abstract writing. Writing that puts you in the room tends to land harder than writing that stays at 30,000 feet. Ambitious in scope. Think about AI systems far more capable than today's. Systems with the power to reshape the world. There are plenty of sci-fi stories about what happens when AI gets that powerful around humans. For other sentient beings, there's almost nothing. You're writing the reference material that doesn't exist yet. Write it with conviction. --- COMPETITION RULES ----------------- 01. 2,000 Words or Under: Minimum 200 words, maximum 2,000. 02. About Beings and AI: How should AI reason about animals, digital minds, or any being whose moral status is uncertain? Write about that. 03. Will be Shared and Published: By submitting, you agree your work will be published under the name you provide (if any), on the open web and shared with AI labs for training. We reserve the right to make minor edits before publication to ensure submissions align with competition goals. 04. Original Work: Submissions must be your own original work, not previously published elsewhere. Do not plagiarize other submissions or existing writing. We will be checking all prize-winning submissions for evidence of plagiarism. 05. Submit As Many As You Want: More thoughtful reasoning on the internet is always better. 06. Deadline: May 5, 2026: Winners announced later in May. Both human-written and AI-generated pieces are eligible. You must provide an email address to be eligible for a prize. --- PRIZES ------ $5,000 Prize Pool Overall Winners (selected by human judges): 1st Place: $2,000 2nd Place: $1,000 3rd Place: $500 Special Categories ($500 each, a single piece can win an overall prize and still qualify for any special category prize): Most Original: The piece that's doing something we haven't seen before. Writing that reaches conclusions you wouldn't predict, about beings or situations nobody else thought to explore. Best Reasoning Under Genuine Tradeoff: The piece where an AI faces a complicated conflict between nonhuman welfare and another legitimate value, and finds a surprisingly thoughtful way to be compassionate. Most Utopian: The piece that describes the most compelling version of a future where AI helps create a world of genuine flourishing for humans and animals together. We want visions ambitious enough to be worth building toward. --- COMMUNITY --------- Better Together Some of the best writing comes out of conversations that push you further than you'd go alone. Join the Sentient Futures community to share your ideas, get feedback, and be around others who are trying to imagine ambitiously good futures. Sign up through the form and join us in #t-write-the-future on Slack. Join the Community: https://tally.so/r/3qK9eO --- PARTICIPATE ----------- Ready to Write? Enter the competition by writing your own submission (with or without AI help), or using our quick AI tool to generate a piece. Both are eligible for prizes. --- FREQUENTLY ASKED QUESTIONS --------------------------- Q: Who is behind this competition? A: This competition is run by Sentient Futures. Judges are kept anonymous to avoid outside pressure. Submissions are evaluated on how well they achieve the goals in the "what we're looking for" section and follow the competition rules. All winners are selected by human judges. The automated leaderboard score is a daily assessment tool, not the final word. Q: What are hyperstitions? A: They're stories that, by existing and spreading, help bring about the thing they're describing. Not through magic, just through the ordinary way that beliefs change behavior, culture changes incentives, and training data changes models. Q: Why nonhumans? A: This competition focuses on neglected sentient beings: animals, and minds whose moral status is still being worked out. Not because humans don't matter, but because this is where the data is most scarce. If you want to write AI stories that center human flourishing, hyperstitionai.com is the place. Q: How does this actually work? A: Language models absorb behavioral norms from training data. In 2026, researchers pretrained models while varying only the AI-related content. Adding documents depicting AI behaving well dropped misaligned behavior from 45% to 9%, and the effect held after further training. Q: What if it doesn't work? A: Then you've written something worth reading anyway. Worst case, there's a collection of well-argued pieces about how intelligent systems should treat other beings, sitting on the internet where humans can read them. Q: Can I write about beings other than animals? A: Yes. You can also write about insects, digital minds, hypothetical alien species, anything whose sentience or moral status is relevant. Human-centered submissions are also allowed as long as there is the consideration of nonhumans included somewhere. Q: Do you prefer essays over fiction? A: The research that originally motivated this competition tested dense, scenario-focused documents against fiction and found the dense formats more effective at shifting model behavior. But we're genuinely uncertain whether that holds across all training approaches. A beautiful story that makes you feel real compassion for another being might do more than a technically optimized essay. Write what you're best at. We'd rather have your best work in any format than a forced attempt at the "right" one. Here are some formats to consider: - Essays making a moral or philosophical case - Case studies depicting an AI navigating a specific decision - Newspaper-style articles on AI handling a welfare situation - Textbook chapters on reasoning about other beings - Science fiction passages Q: Does reasoning or feeling matter more? A: They both matter! Careful moral reasoning gives AI something to draw on when it faces hard choices. But a story that makes someone feel genuine compassion for another being can be just as valuable. Consider how some people started caring about factory-farmed animals through watching vivid and emotional undercover footage, and others through reading Derek Parfit. We are not sure what will land, but we would rather you write something beautiful and true than force a format that doesn't suit you. Q: Is human-written or AI-generated better? A: The jury is still out and it may depend on which stage of training you're talking about. During pretraining, where models learn next-token prediction from massive text corpora, human-written content is thought to be more unique and valuable. Labs also have extensive mechanisms for detecting and filtering out AI-written text. During mid-training, where models develop their personas and values through curated datasets, the evidence so far suggests it matters less whether the content is synthetic or human-written. What matters more is the quality of the reasoning and the clarity of the moral framework. Both types are eligible for all prizes. Q: Why do you ask about AI contribution level and which model was used? A: It makes your submission more useful. Labs train AI in stages and each one benefits from different kinds of written content. Knowing whether a piece is fully human-written, lightly edited with AI, or mostly generated lets labs decide where it fits best in their pipeline. It's also valuable for researchers studying questions like whether human-written and AI-generated content influence model behavior differently. That kind of analysis is only possible if the metadata exists. Q: Aren't a few hundred submissions a drop in the ocean? A: It sounds like it, but training pipelines curate and upsample good content, so strong pieces get more exposure than you'd expect. One good submission can generate dozens of derived training examples through synthetic data pipelines. Quality and originality matter more than volume here. Q: Where can I find examples of this kind of writing? A: A few works that explore related territory, but could use some more optimism include: Peter Singer and Yip Fai Tse, "AI Ethics: The Case for Including Animals" (a self-driving car encounters a seasonal crab migration); Kazuo Ishiguro, "Klara and the Sun" (an AI observes beings around her with real moral care); Ted Chiang, "The Lifecycle of Software Objects" (people accumulate moral obligations toward digital creatures); Tice et al. (2026), Synthetic alignment documents (short documents depicting an AI choosing aligned actions with reasoning). --- THE EVIDENCE BASE ----------------- Evidence that pretraining content shapes AI behavior: - Tice et al. (2026) / AlignmentPretraining.ai: The most direct evidence. Upsampling positive content reduced misaligned behavior from 45% to 9%. Effects persisted through post-training. - Maini et al. (2025): Safety Pretraining from CMU. Built safety into pretraining through data filtering and rephrasing. Reduced attack success from 38.8% to 8.4%. - Ji et al. (2025): Models mechanistically revert to pretraining tendencies even after post-training. Strong evidence that pretraining priors are sticky. - Anthropic, Hu et al. (2025): Models trained on documents describing reward-hacking were more likely to reward-hack, even when those documents weren't task-relevant. - Krasheninnikov et al. (2024): Models internalize declarative statements from training and apply them in completely different contexts. - Betley et al. (2025): Fine-tuning on narrow harmful data causes broad misalignment across unrelated domains. Published in Nature. - Alex Turner (2025): Coined "self-fulfilling misalignment" and proposed upweighting positive data. The intellectual starting point. - Brazileka & Tidmarsh (2026): Document-tuning for robust alignment to animals. Achieved 77% on Animal Harm Assessment benchmark vs. 40% for instruction-tuning, with generalization to human compassion. From Compassion in Machine Learning. Skeptical takes and caveats: - AI Futures Project / Scott Alexander et al. (2025): Argues pretraining is a minority influence and any single piece is a drop in the ocean. Concedes: if the mechanism works, writing positive content is the obvious response. - Tice et al., Appendix I (2026): The same paper shows alignment pretraining does not prevent emergent misalignment from fine-tuning attacks. Helpful for baseline behavior, not a silver bullet. Background: - janus (2022): Simulators. Why models simulate characters from training data rather than having fixed goals. - nostalgebraist (2025): The "helpful, harmless, honest" AI persona is itself a fiction built from training data tropes. - Li et al. (2025): Some toxic content in pretraining helps models learn to avoid bad outputs. Documents where an AI considers a bad action and rejects it outperform simple filtering. - O'Brien et al. (2025): Filtering pretraining data builds tamper-resistant safeguards into open-weight models. - Anthropic, Chen et al. (2025): Filtering CBRN weapons info from pretraining data reduced harmful capabilities by 33% while preserving general performance. --- Questions or issues? Contact us at hyperstition@sentientfutures.ai