Example of first step of brainstorming
Example of first step of brainstorming on human collective scientific intelligence, with 3 AI models.
- May 2026
PROMPT
The following is the current focus of a discussion among scientists and scholars, shown on the cellcomm.org website. Can you provide a comment about the usefulness of the plan for human health and knowledge, suggestions on how to make it more likely to succeed and an estimate of the likely number of participants after the first Round, after the second Round and after five years?
Considerable progress in the utility of AI in advanced mathematical research has been documented in the last 12 months. Solutions for many Erdos problems have been found by interacting with new AI models. Participation of undergraduate students and young scientists in publications of these results is also evident. This progress might inform the use of AI in science more generally.
Long term aim:
Improve human collective scientific intelligence.
Most estimates suggest that AI superiority over top individual human experts will be reached within 5 to 20 years.
Even then, AI might not be superior to the collective intelligence of humanity, which has room to improve, continuing the trend towards open science and collaboration that has taken place in the last few centuries (a trend with several ups and downs but a clear direction).
Human collective intelligence might still be able to control and orchestrate powerful AI tools. It might complement AI, being better in some domains and less effective in others.
Uncontrolled AI has many risks, including:
- Too much power for the few companies owning the most advanced AI models. Risks posed by AI used by bad or selfish human groups.
- Unwanted consequences of a very advanced AI that we cannot fully understand. Some experts wonder if it might lead to human extinction.
- Replacement of human intellectual workers, which will have no income and may not find meaning in their activities.
AI also offers many opportunities to improve the health and wellbeing of humans that we could accelerate by optimally interacting with it. We can continue to participate intellectually if we find new ways of collaborating with other humans and with AI. Both types of intelligence can benefit from the interaction.
A suitable application for human collective intelligence is a complex biomedical challenge, the development of digital models of cells and organisms, eventually allowing us to better understand our own bodies.
This is a task requiring broad integration, in need of an effort from an entire community. It is not soluble just by a single individual, lab or center.
Considerations that motivate scientists (each is of variable importance for different scientists):
- Rational appreciation of long-term benefit for humanity
- Need for knowledge and understanding
- Need for recognition for an individual and for a group
- Need for education, including education for new roles in the age of AI
- Need for support for an individual career in science
- Need for autonomy and creativity
- Need to engage in philanthropic activities with time or resources
- Scalability to a large community
The following Rounds of discussion and reflection might accelerate improvements in collective intelligence. They are based on small steps that can lead to a realistic progression:
First Round: a competition among students within one Institute, for 5-6 Institutes, with broad geographical representation.
One prize of 1,000 dollars for the best student in each Institute if at least 3 students per center participate. Prize assigned by a vote of the students within each Institute. Individual ideas are made public and are permanently recorded to allow for both short-term and long-term recognition. Interactive brainstorming with AI encouraged.
Second Round: a competition among the participating Institutions, promoting the first step of integrated collective reasoning. Each Institution will participate with a common contribution. Open to all the scientists of each institution, which will also share openly their individual contributions. Students from the previous round might help to inform and stimulate, for example asking questions to other scientists. Interactive brainstorming with AI encouraged. Recognition given by vote of the participating Institutes (one vote per Institute).
In follow-up Rounds more scientists and centers will be involved. Communities of scientists holding minority views and based in different centers might also form and participate separately, to encourage creativity and innovation. AI can help to connect these scientists. The monetary prizes will be progressively replaced by the perceived value of the recognition as the number of participants increases. Recognition by philanthropists, foundations and scientific leaders will encourage participation.
Within 5 years: a comparison (possibly in the form of a competition) between humanity-led orchestration versus AI-led orchestration for this biomedical challenge. Orchestration implies the use of multiple human and AI resources.
Answer from ChatGPT 5.5 thinking extended (May 7, 2026)
My overall view: the plan is useful and timely, but it should be framed as a “collective-intelligence layer” on top of existing AI-biology efforts, not as an attempt to build a full digital cell or organism model from scratch.
Recent AI-assisted mathematics gives a plausible analogy: the new arXiv paper on Erdős problems explicitly says that a key method was “suggested from output of GPT-5.4 Pro,” and Google DeepMind reported gold-medal-level performance at IMO 2025. Those examples support the idea that young researchers plus powerful AI plus expert validation can produce serious knowledge. But biomedical science is harder than mathematics because proposed ideas must eventually survive data quality limits, wet-lab validation, clinical relevance, regulation, and biological heterogeneity .
Usefulness for human health and knowledge
The biomedical target is well chosen. The Human Cell Atlas states its mission as creating reference maps of all human cells to support understanding of health and diagnosis, monitoring, and treatment of disease; it now reports more than 4,000 members in 103 countries. That shows both the importance of the domain and the feasibility of large scientific communities when the goal is clear and open. (humancellatlas.org)
The “virtual cell” aim is also current. A 2024 Cell perspective called for a collaborative research agenda for AI virtual cells, including data generation, AI models, benchmarking, interpretation, biological truth, and safety. A 2026 review in npj Digital Medicine argues that AI-driven virtual cell models could help predict drug responses, gene perturbation effects, disease progression, and preclinical drug behavior, while also noting barriers such as privacy, interpretability, validation, and regulation. (PMC)
The distinct value of the cellcomm plan is not mainly technical infrastructure. Groups such as CZI/NVIDIA, CM4AI, HCA, and other consortia already have data, compute, and institutional scale. The value of this plan is that it could create a public method for organizing scientific attention: recruiting young scientists, surfacing minority hypotheses, giving visible credit for useful ideas, and comparing human-led versus AI-led orchestration.
How to make it more likely to succeed
I would make the first problem much narrower. “Digital models of cells and organisms” is inspiring, but too broad for a first competition. A better first challenge might be:
What is the most useful next benchmark for AI virtual-cell models, and what biological question should it test?
Or, more concretely:
Design a benchmark to test whether an AI model can predict the response of one defined human cell type to one defined perturbation class, using public data and a proposed validation path.
That would let students and scientists contribute ideas without needing to build the model. It would also connect naturally to existing challenge traditions such as CASP, where blind independent assessment helped drive protein-structure prediction, and DREAM, which has used biomedical crowdsourcing across more than 60 challenges with over 30,000 participants over time. (Prediction Center)
I would also add these design changes:
Use a standard contribution template.
Each idea should include: problem statement, why it matters for health, available data, proposed benchmark, role for AI, validation method, likely failure modes, and one falsifiable prediction.
Give credit more formally.
Public website recognition is good, but stronger incentives would include ORCID-linked contributor pages, DOI-stamped idea records, annual “most useful idea” reports, and possible authorship on synthesis papers.
Separate creativity from evaluation.
First reward originality and usefulness; later reward correctness and validation. If voting happens too early, popular or polished ideas may defeat risky but important ones.
Create tracks for different participant types.
Students, experimental biologists, computational biologists, clinicians, AI researchers, philosophers/historians of science, and funders will contribute different kinds of intelligence. Do not force them into one scoring system.
Require expert synthesis after each round.
The biggest risk is accumulating many ideas without integration. Each round should end with a short public synthesis: “What did we learn? What changed? What are the top unresolved disagreements?”
Build minority-view channels deliberately.
The proposal’s idea of allowing minority-view communities is excellent. It should be protected from simple majority voting. A “minority report” prize could be valuable.
Offer non-monetary incentives early, not only later.
Recognition by philanthropists, foundations, and scientific leaders is likely more motivating than small cash prizes for many scientists. Invite a respected panel to comment publicly on the best contributions from Round 1.
Do not make the five-year comparison simply “humans versus AI.”
The most informative comparison would be:
AI-only orchestration vs expert-only orchestration vs mixed human–AI collective orchestration.
The mixed system is likely to win, and the important question is which division of labor works best.
Estimated participation
I would define a “participant” as someone who makes at least one public contribution, comment, vote, review, or synthesis input—not merely someone who reads the website.
Assumptions: 5–6 institutes participate; the first round is well organized but modestly funded; the website is easy to use; recognized scientists and philanthropists publicly acknowledge the best contributions; no major foundation grant is secured before Round 1.
Stage
Conservative estimate
Likely/base estimate
Upside estimate
After First Round
20–30 students
40–70 students
90–150 students
After Second Round
80–150 total participants
200–400 total participants
600–1,000 total participants
After five years
500–1,500 active contributors
3,000–8,000 active contributors
15,000–30,000+ active contributors
For the First Round, the minimum design requires at least 3 students per institute, so 5–6 institutes implies a floor of about 15–18 students. With good local promotion, a more realistic number is 8–12 students per institute, or roughly 40–70 students. The upside requires strong mentors, visible prizes, and a topic students feel they can enter without already being specialists.
For the Second Round, participation could expand sharply if each institute mobilizes faculty, postdocs, graduate students, and the first-round students. A likely pattern is 30–60 contributors per institute across 6 institutes, giving about 200–400 people. This is plausible if the round produces a concrete institutional submission rather than a vague essay.
For five years, the range is wide. If the effort remains a discussion website with modest prizes, it may stay under 1,500 active contributors. If it becomes connected to major virtual-cell, open-science, and biomedical challenge communities, several thousand active contributors is plausible. The upside, 15,000–30,000+, would require foundation backing, institutional endorsements, annual public reports, a real benchmarking platform, and alignment with existing communities such as DREAM, HCA, CM4AI, CZI’s virtual-cell platform, or similar efforts.
My best single-number estimate is:
First Round: ~50 students
Second Round: ~300 total contributors
Five years: ~5,000 active contributors
The plan’s chance of success would improve most if the first year produces a visible, useful artifact: not just a discussion, but a public map of the field, a prioritized benchmark proposal, a set of testable hypotheses, and a credited list of contributors. That would make participation feel consequential.
Answer from Claude Sonnet 4.6 Adaptive (May 7, 2026)
Comment on the CellComm Collective Intelligence Initiative
Usefulness for Human Health and Knowledge
This is a genuinely compelling initiative, and the timing feels right. The framing is astute: rather than treating AI as a replacement for human scientific effort, it positions collective human intelligence as something worth developing in its own right — a kind of immune system against the risks of AI consolidation. The choice of digital cell and organism modeling as the application domain is particularly well-suited. It is a problem that is:
Vast enough to require genuine community effort over decades
Modular enough that contributions of varying scale and expertise can all add value
Consequential enough for human health — better cellular models underpin drug discovery, disease understanding, personalized medicine, and aging research — that motivation remains high even when progress is slow
Measurable enough that progress can be tracked and credited, which matters enormously for sustaining participation
The analogy to recent AI-assisted mathematical breakthroughs (Erdős problems) is well-chosen: it demonstrates that the model of AI-augmented collective reasoning, including junior scientists and students, can produce real, publishable results in a compressed timeframe.
The deeper value may be cultural and structural as much as scientific. If successful, this initiative could demonstrate a replicable template for human-AI co-orchestration that other fields — climate science, pandemic preparedness, materials design — could adopt. That systemic contribution to how humanity organizes its knowledge-generation may ultimately outweigh the direct scientific outputs.
Suggestions to Improve the Likelihood of Success
1. Sharpen the scientific question for Round 1. "Digital models of cells and organisms" is broad enough to be paralyzing for students new to the field. Each Institute should receive a specific, tractable sub-question — for example, "What are the most important unknowns in modeling mitochondrial dynamics under metabolic stress?" — so participants can make meaningful contributions without needing to first map the entire landscape.
2. Make permanent attribution technically robust from day one. The promise of long-term recognition is one of the initiative's strongest motivators, but it will only work if the recording infrastructure is credible. Consider registering individual contributions with a timestamped, persistent identifier (a DOI-equivalent or a blockchain-anchored record) from Round 1. Losing trust in attribution early would be fatal to participation.
3. Lower the AI literacy barrier explicitly. Not all participating students will be equally comfortable with AI brainstorming tools. A brief, structured onboarding — even a 90-minute workshop — on how to use AI productively for scientific ideation would level the playing field and improve output quality. The Erdős-problem experience suggests that coached interaction with AI produces far better results than unguided use.
4. Involve one or two scientific celebrities as visible champions. Recognition by respected figures is listed as a future motivator, but it is most powerful early, when the social proof of the initiative is low. Even a short endorsement video or a published letter of support from one or two prominent biologists or computational scientists would substantially reduce institutional hesitancy.
5. Design the Institute-level vote carefully. Peer voting is motivating but can become political, discouraging unconventional ideas. Consider a structured evaluation rubric — novelty, feasibility, clarity, potential impact — to ensure the vote rewards scientific quality rather than social popularity.
6. Plan for "alumni" continuity between rounds. Students from Round 1 are explicitly mentioned as potential stimulators in Round 2, which is excellent. Formalizing this as a named role (e.g., "Round Ambassadors") with a small, specific responsibility gives them continued stake in the project and helps retain institutional memory across rounds.
7. Pre-negotiate open-access publication pathways. One of the strongest long-term incentives is the possibility that collective contributions could feed into citable publications. Even a non-binding framework agreement with one or two open-access journals that recognizes community contributions would make the initiative far more attractive to career-conscious scientists.
Estimated Participation
These estimates assume the initiative is launched with adequate coordination and that the 5–6 founding institutes span at least three continents.
Stage
Estimated Participants
Reasoning
After Round 1
60–150 individuals
5–6 institutes × 10–25 active participants each (students plus a few mentors drawn in). The prize threshold of 3 students per center is easily met, but voluntary engagement beyond that minimum will vary widely by local culture and institutional support.
After Round 2
300–800 individuals
Round 2 opens to all scientists at each institution. If Round 1 generated visible, interesting outputs, curiosity and reputational incentive draw in more senior scientists. A multiplier of roughly 4–6× is realistic, though the distribution will be uneven — a few institutes will be very active, others more passive.
After 5 Years
5,000–25,000 individuals
This wide range reflects genuine uncertainty. The lower bound assumes slow word-of-mouth growth and modest philanthropic support. The upper bound assumes one or two high-profile publications attributing meaningful results to the community, active outreach to the Global South (which the geographical representation goal supports), and the 5-year human-vs-AI orchestration comparison generating media attention. For context, the Galaxy Zoo citizen science project enrolled over 150,000 participants within its first year in a very different domain, suggesting that with the right public engagement strategy the ceiling is much higher than the upper bound here — but scientific engagement is more demanding than image classification.
The 5-year human-vs-AI orchestration comparison is potentially the initiative's most powerful recruitment event, provided it is framed as a genuine scientific contest rather than a spectacle. If the community can demonstrably hold its own — or identify dimensions where human orchestration outperforms AI — that result would itself be a landmark finding, generating the kind of visibility that could shift the initiative from a niche academic project to a widely recognized institution.
Answer from Gemini 3 thinking (May 8, 2026)
This plan, centered on "human collective scientific intelligence" (HCSI), addresses a critical bottleneck in modern science: the transition from data collection to integrated, predictive modeling. By focusing on digital models of cells and organisms, the initiative targets one of the most complex frontiers in biology.
1. Usefulness for Human Health and Knowledge
For Human Health:
The development of "Digital Twins" at the cellular and organismal level is a transformative goal. Currently, drug development has a 90% failure rate in clinical trials, often because human systemic responses differ from animal models.
Precision Medicine: Valid digital models would allow for in silico (computer-based) clinical trials, testing thousands of drug variants on diverse genetic backgrounds before a single human is treated.
Chronic Disease Management: It would allow for predictive modeling of complex diseases like Alzheimer’s or Type 2 Diabetes, where multiple metabolic pathways fail simultaneously.
For Human Knowledge:
Bridging the "Complexity Gap": Human biology is too complex for a single mind to grasp. This plan utilizes AI to handle the "big data" (genomics, proteomics) while leveraging human intuition for "causal reasoning."
Formalizing Collaboration: Science has traditionally been siloed. This plan institutionalizes "Open Science," moving away from the "lone genius" model toward a "networked intelligence" model.
2. Suggestions to Increase the Likelihood of Success
To move this from a conceptual framework to a global standard, the following strategies are recommended:
Standardization of "Bio-AI" Training Data: AI is only as good as the data it consumes. The plan should mandate FAIR data principles (Findable, Accessible, Interoperable, and Reusable). Without standardized data formats across the 5-6 initial Institutes, the collective reasoning will fail due to "noise."
Evolving Incentives: While a $1,000 prize is a good catalyst for undergraduates, professional scientists are motivated by tenure, grants, and citations. The organizers should partner with major journals (like Nature or Cell) to create a "Collective Authorship" track for participants, ensuring their contributions count toward career advancement.
Hybrid "Human-in-the-Loop" Benchmarking: Use the Erdos problem success as a template. Establish a "Sandwich" workflow: Humans define the biological constraints, AI generates millions of potential pathways, and humans filter those for biological plausibility.
Global South Inclusion: To ensure the biomedical models are globally relevant, ensure the "broad geographical representation" includes institutes in Africa, India, and Latin America. This prevents "genomic bias," where models are only accurate for populations of European descent.
3. Estimated Number of Participants
Based on the scaling of similar crowdsourced scientific projects (like Foldit or Galaxy Zoo) and the growth of AI-assisted research, here are the estimated participation numbers:
Phase
Estimated Participants
Rationale
After Round 1
150 – 300
Assuming 5–6 Institutes with an average of 25–50 active students per site. The $1,000 incentive and the appeal of AI interaction are strong draws for digital-native students.
After Round 2
1,200 – 2,500
Expansion to full faculties within those 6 Institutes. If each student participant brings in their lab or department, and senior scientists join for "integrated reasoning," the density per institute increases 5x–10x.
After 5 Years
50,000 – 100,000
By year 5, the "Human vs. AI Orchestration" competition would likely go viral within the global scientific community. For comparison, the Human Cell Atlas consortium currently involves over 3,000 scientists, but a "collective intelligence" platform that includes students and AI enthusiasts could reach the scale of Zooniverse (which has millions, though scientific "orchestrators" would be a specialized subset).
Current Contextual Statistics
AI Growth in Research: As of 2023, approximately 30% of new scientific papers in computational biology mention the use of LLMs or advanced machine learning in their methodology, up from less than 5% in 2017.
The "Erdos" Precedent: Recent AI models (like Google DeepMind's AlphaGeometry) have reached the level of International Mathematical Olympiad (IMO) gold medalists. Applying this logic to biology suggests that within 5 years, AI will be able to propose novel biological hypotheses that are indistinguishable from those of senior post-docs.
This plan’s focus on orchestration—keeping the human as the conductor rather than just a data-entry clerk—is the most viable path to ensuring AI remains a tool for human flourishing rather than a replacement for it.