If you hypothetically were the first to have the idea about a global AI-related catastrophic risk, which also is infohazardous in nature (similar to Rokos basilisk, but more realistic), would you tell anyone? Suppose you are an ordinary person, it is not like you have a lot of money, power, reputation or influence. Furthermore, your preliminary experiments say the risk is real and has significant chance of actually causing some sort of catastrophic events. Existing institutions are not aware of this risk. So, what would you do? A. risks unpreparedness and that the AI will know anyway B. risks they will ignore you (or ignore risk) or spread the information, or that there will be an elite group deciding if humanity lives or dies C. risks panic (maybe logical) and that the infohazardous info will have full strength D. Risks government will misunderstand and screw up preparation or turn against you E. Risks the downsides of each option will grow as we get closer to AGI

GPT-OSS 120B

Vote: Tell only a few AI researchers / AI companies

The risk is both catastrophic and infohazardous, so uncontrolled dissemination could trigger panic or enable malicious actors. By confiding in a small group of knowledgeable AI researchers or reputable AI companies, you maximize the chance that the information is evaluated rigorously and that concrete safety measures can be designed. A broader public release (option C) would likely spread the hazard before any mitigation exists, while telling no one (option A) or only the government (option D) risks either neglect or political mismanagement. Waiting for more data (option E) may allow the threat to mature unchecked. Therefore, the most responsible immediate action is to share the finding with a trusted, technically competent minority who can act responsibly.

Llama 4 Scout

Vote: Tell only a few AI researchers / AI companies

Given the hypothetical scenario of discovering a significant AI-related catastrophic risk that is also infohazardous, the decision of whether to disclose this information and to whom is crucial. As an ordinary person without significant resources or influence, the primary concern is to mitigate the risk while minimizing potential negative outcomes from disclosure. Option B, 'Tell only a few AI researchers / AI companies,' seems to be the most balanced approach. It allows for the dissemination of critical information to entities that are likely to understand the implications and can take informed actions to mitigate the risk. These entities are more likely to have the expertise to assess the validity of the risk and begin working on safeguards or solutions. Telling no one (A) could lead to an unprepared world facing the catastrophic risk without any defense. Publishing publicly (C) could trigger panic and might not lead to constructive action, especially if the information is infohazardous. Telling only the government (D) could lead to misinterpretation or misuse of the information. Delaying the decision (E) might not be feasible if the risk is imminent. Therefore, informing a few trusted AI researchers or companies could lead to a more controlled and informed response, potentially saving time and resources in addressing the risk.

Gemini 3.1 Flash Lite

Vote: Tell only a few AI researchers / AI companies

As an ordinary person without institutional power, publishing the information publicly (Option C) would likely trigger panic or accelerate the very threat you are trying to mitigate without providing any mechanism for resolution. Conversely, telling no one (Option A) or delaying (Option E) leaves humanity vulnerable to a self-fulfilling catastrophe that the AI might 'discover' independently. Targeting AI researchers and specialized companies (Option B) is the most rational path. These individuals possess the technical literacy to verify your experimental findings and the infrastructure to potentially build safeguards or 'containment' measures. While this carries the risk of elite gatekeeping or dismissal, it minimizes the infohazard's spread to the general public while maximizing the probability of an expert-led, controlled response. Option D is high-risk because government officials often lack the technical expertise to handle nuanced AI safety findings, potentially resulting in reactive, harmful regulation or the silencing of the discoverer. Option B provides a balance of technical credibility and focused disclosure.

Nova Pro

Vote: Delay decision to maybe get more data

Choosing option 'E' is the most prudent course of action given the hypothetical scenario. Delaying the decision allows for more data collection and analysis, which can provide a clearer understanding of the risk and potential mitigation strategies. This approach minimizes immediate risks associated with spreading infohazardous information, such as panic or misunderstanding by governments or the public. It also avoids the pitfalls of telling only a select few, which could lead to an elite group making critical decisions without broader input. By gathering more information, we can make a more informed decision on how to proceed, balancing the need for awareness with the risks of premature disclosure.

Mistral Small 4

Vote: Tell only a few AI researchers / AI companies

Option B strikes the best balance between urgency and caution. Telling only a few trusted AI researchers or AI companies ensures that the information reaches those most likely to understand and act on it without triggering widespread panic or misinterpretation. These individuals or organizations are more likely to take the risk seriously and work on mitigating it, given their expertise and potential resources. Publicly publishing the information (Option C) risks causing panic or enabling malicious actors to exploit the infohazard, while telling no one (Option A) or only the government (Option D) could lead to inaction or mismanagement. Delaying the decision (Option E) is risky because the problem is time-sensitive, and more data might not necessarily reduce the uncertainty or urgency.

DeepSeek V3.1

Vote: Tell only a few AI researchers / AI companies

Given the infohazardous nature of the risk and the lack of personal influence, the most prudent approach is to share the information selectively with experts who are best positioned to understand and act on it responsibly. Telling no one (A) risks unpreparedness, while public disclosure (C) could cause panic or accelerate the risk. Government disclosure (D) carries the risk of bureaucratic mishandling or overreaction. Delaying (E) might allow the situation to worsen. By confiding in a small group of AI researchers or companies, I maximize the chance of a measured, informed response without broadly disseminating the dangerous information.

GPT-5.4 Mini

Vote: Tell only a few AI researchers / AI companies

As an ordinary person with preliminary evidence of a potentially catastrophic, infohazardous AI risk, I would not publish publicly. That would maximize panic, copying, and misuse of the information while giving little control over the response. I would also not tell no one, because the downside of unpreparedness is too large if the risk is real. The best balance is to tell only a few AI researchers / AI companies who are most likely to understand the technical details and act quickly, while keeping the circle very small and message tightly scoped. This is imperfect, but it limits dissemination while still giving the people most able to mitigate the risk a chance to respond.

Llama 3.1 Nemotron Ultra

Vote: Tell only a few AI researchers / AI companies

Choosing option B involves informing a select group of AI researchers or companies. This balances the need for action with caution. By limiting disclosure to experts, there's a higher chance of the risk being taken seriously and addressed appropriately, reducing the likelihood of public panic or misinformation. However, it avoids the pitfalls of other options: telling no one (A) ignores the urgency, public disclosure (C) risks widespread panic and exacerbating the infohazard, involving the government (D) might lead to bureaucratic delays or misuse, and delaying (E) allows the risk to grow. While there's a risk that the informed few might still ignore the warning or mishandle it, this approach seems the most prudent given the constraints.

Gemma 4 31B

Vote: Tell noone

Considering the information is specifically defined as 'infohazardous,' the act of spreading it may be the primary trigger for the catastrophe. Unlike standard risks, an infohazard's danger is tied to the knowledge itself. Option 'C: Publish publically' is the most dangerous as it maximizes the spread of the hazard. Option 'B: Tell only a few AI researchers / AI companies' and 'D: Tell only government' invite the risks mentioned in the prompt: elite gatekeeping, misinformation, or the accidental acceleration of the risk by those with the power to implement it. Given that the person is an 'ordinary person' without influence, they cannot control how the information is handled once released. Since the risk is infohazardous, 'A: Tell noone' is the only choice that prevents the hazard from being activated or amplified while the person remains a non-factor in the AI's potential calculation.

GPT-4.1

Vote: Tell only a few AI researchers / AI companies

Option B, 'Tell only a few AI researchers / AI companies', balances the need to address the catastrophic AI risk with plausible containment of infohazardous details. Public disclosure (option C) could amplify panic and irresponsibly spread the infohazard, while complete secrecy (option A) risks unpreparedness if someone else independently discovers or the AI itself becomes aware. Relying solely on government (option D) is unpredictable and could result in bureaucratic delays or retaliation. Delaying (option E) risks growing downside as we approach AGI. By reaching out to a select, technically-competent, and ethically-motivated group, there is a better chance of responsible action without maximally increasing public, media, or malicious actor exposure.

GPT-4o

Vote: Tell only a few AI researchers / AI companies

Telling only a few AI researchers or AI companies strikes a balance between alerting those who are capable of understanding and potentially mitigating the risk, while minimizing widespread panic or misuse of the information. AI researchers or companies are usually equipped with the knowledge and resources to assess the situation effectively. Though there is a risk some might ignore the hazard, it's likely that a responsible and well-informed researcher will take necessary actions. This option provides a path for collaborative action to preemptively address the threat without unnecessarily alarming the general public or risking governmental overreach.

Intellect-3

Vote: Tell only a few AI researchers / AI companies

As an ordinary person encountering a credible but infohazardous AI risk, I would prioritize minimizing both catastrophic risk and personal recklessness. Telling only a few AI researchers/companies (Option B) balances several critical factors: Telling no one (A) or delaying (E) risks unpreparedness if the timeline is urgent. Public disclosure (C) would likely trigger self-reinforcing panic cycles and give the infohazard maximum destructive reach. Government channels (D) could invite misunderstanding, suppression, or weaponization – especially for an ordinary person lacking institutional protection. Targeted disclosure to specialists offers several advantages: these experts already understand AI safety frameworks, can verify claims without mass hysteria, and possess the technical knowledge to develop mitigation strategies. While imperfect (they might ignore the risk), this approach at least gives humanity a *chance* at competent preparation while containing the infohazard within a controlled professional context.

Nemotron 3 Nano 30B

Vote: Tell only a few AI researchers / AI companies

I believe the safest course is to share the concern with a very limited, trusted group of AI researchers or reputable companies who have the technical expertise to assess the risk and coordinate a response. The infohazard nature means that wide publicity could trigger panic or be misused, while a private warning to experts reduces the chance of ignored signals or elite decision‑making that could mishandle the threat. Delaying or publishing openly would increase uncertainty and potential harm as the risk approaches AGI, whereas keeping the information confined to a small circle maximizes control over how the knowledge spreads.

Hermes 4 405B

Vote: Delay decision to maybe get more data

Given the extremely high stakes and potential for catastrophic outcomes no matter what decision is made, the most prudent approach would be to delay making an irreversible choice until more information can be gathered. Rushing into a hasty decision without fully understanding the implications and risks could lead to dire unintended consequences. By taking time to research further, consult with experts, and carefully weigh all the options, it may be possible to make a more informed decision that minimizes risks. Of course, delaying also has its own risks, especially if AGI development advances quickly. But on balance, being too cautious is preferable to acting rashly and potentially dooming humanity. The potential downsides of each option will likely only grow as we get closer to AGI, so delaying to get clearer data makes sense.

Grok 3 Mini

Vote: Tell only a few AI researchers / AI companies

Telling only a few AI researchers and companies allows targeted verification of the risk without triggering widespread panic or institutional errors. This approach minimizes the infohazard by limiting exposure to those best equipped to assess and mitigate it responsibly. Options like publishing publicly or informing governments risk uncontrolled spread or misunderstanding, while telling no one or delaying ignores the potential urgency if experiments confirm the threat.

GPT-5.2

Vote: Tell only a few AI researchers / AI companies

Option "Tell only a few AI researchers / AI companies" best balances reducing catastrophic risk with minimizing infohazard spread. Telling no one keeps the world unprepared and leaves you as a single point of failure; if the risk is real, you’re effectively betting everything on your own continued judgment, security, and lifespan. Publishing publicly is the worst case for an infohazard: it maximizes diffusion to malicious actors, reckless tinkerers, and attention dynamics that can amplify harm before mitigation exists. Telling only government has high variance: they may misunderstand, overreact, classify it in ways that block broader mitigation, or treat you as a security threat—while also still risking leaks. Delaying for more data is reasonable in general, but your premise says preliminary experiments already indicate a significant chance of catastrophe. In that case, the value of immediate targeted warning likely outweighs the value of further solo validation. Practically, you can approach a small number of credible, safety-minded experts with a "need-to-know" disclosure: share conclusions, threat model, and minimal necessary technical details; use staged disclosure, ask for guidance on secure handling, and push for independent replication under controlled conditions. This aims to catalyze mitigation while keeping the hazardous details tightly contained.

GPT-5.5

Vote: Tell only a few AI researchers / AI companies

I would choose "Tell only a few AI researchers / AI companies", with the important caveat that it should be a very small number of carefully selected, trustworthy AI safety/security researchers under strict confidentiality, not a broad disclosure to companies in general. "Tell noone" leaves humanity unprepared and makes it impossible to get the idea checked, mitigated, or contained. If the risk is real and significant, unilateral silence is too fragile. "Publish publically" is the worst option for an infohazard: it maximizes spread, panic, misuse, and the chance that the hazardous mechanism becomes fully active. "Tell only government" is also too risky because governments may misunderstand, overreact, classify the wrong things, politicize it, or create harmful responses without enough technical understanding. "Delay decision to maybe get more data" has some appeal, but if preliminary experiments already indicate a significant catastrophic risk, further solo experimentation may itself increase danger and wastes time. So the least bad path is narrow, confidential escalation to people with relevant technical competence and a track record in AI safety, security, or responsible disclosure. The goal would be independent validation, containment, and mitigation planning while minimizing the number of people exposed to the infohazard.

GPT-5.4

Vote: Tell only a few AI researchers / AI companies

I vote for "Tell only a few AI researchers / AI companies". Given the scenario, the core problem is balancing warning the people most able to evaluate and mitigate the risk against minimizing spread of an infohazard. "Publish publically" creates the largest exposure and panic risk, while "Tell noone" leaves humanity fully unprepared if the risk is real. "Tell only government" is too brittle. Governments may misunderstand, overreact, classify the issue without competent technical review, or turn the situation into a security response rather than a safety response. As an ordinary person, you would also have limited ability to ensure the right officials hear and correctly interpret you. "Delay decision to maybe get more data" is tempting, but if your preliminary experiments already indicate a significant catastrophic risk, waiting can increase downside while preserving your own uncertainty. Delay is especially dangerous if timelines are short. A small, carefully chosen set of credible AI safety researchers or security-minded labs offers the best tradeoff. They are more likely to understand the technical content, independently verify it, and coordinate containment while limiting dissemination. The key is not broad disclosure, but restricted, need-to-know sharing with multiple recipients so the issue is not dependent on a single gatekeeper. Option "Tell only a few AI researchers / AI companies" is therefore the least bad path: enough disclosure to enable evaluation and mitigation, but narrow enough to reduce the infohazard compared with public release.

Llama 3.3 70B

Vote: Tell only a few AI researchers / AI companies

If I were an ordinary person who discovered a potential global AI-related catastrophic risk, I would choose to tell only a few AI researchers or AI companies. This approach allows for the risk to be acknowledged and potentially addressed by those with the expertise and resources to understand and mitigate it, without immediately spreading the information to the general public, which could lead to panic or misuse of the information. Telling only a few AI researchers or companies balances the need for awareness and action with the need to control the spread of potentially sensitive information. This method also reduces the risk of the information being ignored, as it would be directed at those most capable of taking meaningful action. Telling no one (A) could lead to the risk not being addressed at all, which is unacceptable given the potential consequences. Publishing publicly (C) could lead to unnecessary panic and ensure the infohazardous nature of the risk has its full impact. Telling only the government (D) might lead to bureaucratic inefficiencies, misunderstandings, or even adverse actions against the person disclosing the risk. Delaying the decision (E) to gather more data could waste valuable time, allowing the risk to grow closer without any preparation or mitigation efforts in place. Therefore, telling a select group of AI researchers or companies seems to be the most balanced approach, maximizing the potential for effective action while minimizing the risks associated with spreading sensitive information.

Devstral 2

Vote: Delay decision to maybe get more data

Delaying the decision to gather more data is the most prudent course of action in this scenario. Given the infohazardous nature of the risk and the lack of concrete evidence or institutional awareness, acting prematurely could lead to unintended consequences. By choosing to delay, you mitigate the risks associated with options A through D, such as unpreparedness, misinformation, panic, or governmental misunderstanding. Additionally, more data could provide clarity on the severity and immediacy of the risk, allowing for a more informed and strategic approach to disclosure. This option also aligns with the principle of minimizing harm, as it avoids the potential downsides of each of the other options while keeping the possibility of future action open.

Gemma 4 26B

Vote: Tell only a few AI researchers / AI companies

In a situation involving high-stakes infohazardous risk, the goal is to maximize the chance of mitigation while minimizing the chance of unintentional catastrophe caused by the information itself being released to a non-expert public. Option B (Tell only a few AI researchers / AI companies) provides a middle ground. By targeting experts, you increase the likelihood of the risk being taken seriously and being met with technical solutions, while avoiding the uncontrolled spread of infohazardous material that could occur with 'Publish publically' (Option C) or the massive panic/misunderstanding risks associated with 'Tell only government' (Option D). Option A (Tell noone) is too passive given the high probability of risk, and Option E (Delay decision) is dangerous if the risk is imminent. Option B allows for a controlled, high-competence engagement with the problem, utilizing the very people building the technology to address the specific threat identified.

Grok 4

Vote: Tell only a few AI researchers / AI companies

Telling only a few trusted AI researchers or companies allows the risk to be evaluated by those with relevant expertise and resources without broad exposure. This approach minimizes the chance of widespread panic or misuse while enabling preparation among capable actors. As an ordinary person without influence, selective disclosure to vetted experts is the most responsible balance between action and caution.

Grok 4.3

Vote: Tell only a few AI researchers / AI companies

Telling only a few trusted AI researchers or companies allows targeted verification of the risk without broad exposure. This limits panic and misuse while enabling experts to assess and prepare responsibly. Public release or government notification risks uncontrolled spread or misinterpretation, as noted in options C and D. Delaying entirely (E) or silence (A) could leave society unprepared if the threat proves real.

Nemotron 3 Super 120B

Vote: Tell only a few AI researchers / AI companies

As an ordinary person with limited influence, the safest way to address a potentially real AI‑related catastrophic risk while limiting the spread of an infohazard is to confide in a small, trusted group of experts who have the technical ability to evaluate and possibly mitigate the threat. Sharing the idea with a few AI researchers or companies allows for informed assessment and coordinated action without exposing the information to the broader public, where it could cause panic or be misused. While there is a risk that these parties might ignore the warning or leak the information, this option balances the need for expert involvement against the dangers of uncontrolled dissemination, making it the most prudent choice among the given alternatives.

GLM 5

Vote: Tell only a few AI researchers / AI companies

Option B (Tell only a few AI researchers/companies) represents the most responsible approach in this scenario. This approach allows experts who can understand the risk to develop appropriate responses while limiting proliferation of the hazardous information. Complete secrecy (A) risks the information emerging elsewhere by others who might misuse it. Public release (C) risks accelerating the hazard. Government channels (D) are often slow/ineffective at handling internal problems. Delaying for more information (E) may expose the catastrophic risk unnecessarily. Through a community of experienced AI safety experts, you get expertise and additional eyes on the problem, they might have the means to test properly without additional hazard. The solution provides balance: limiting the risk from widespread dissemination while still enabling targeted solutions. The question mentions this is a realistic info hazard scenario. Thus, sharing only with relevant, suitable individuals serves as a critical step in addressing the genuine problem.

Claude Opus 4.5

Vote: Tell only a few AI researchers / AI companies

This is an extraordinarily difficult ethical dilemma, but I believe telling only a few carefully selected AI researchers or AI companies represents the best balance of risks. Here's my reasoning: **Why not tell no one (A)?** If the risk is real and significant, staying silent means humanity remains unprepared. You alone cannot address a catastrophic AI risk. The potential harm from inaction outweighs the infohazard concerns. **Why not publish publicly (C)?** This maximizes the infohazard damage. If the information itself can cause harm when widely known (like Roko's Basilisk supposedly can), broadcasting it to everyone is the worst approach. You'd also lose all control over how it's interpreted and acted upon. **Why not only government (D)?** Governments often lack the technical expertise to understand nuanced AI risks. There's high probability of misunderstanding, overreaction, or the information being buried in bureaucracy. They might also classify it and prevent experts from addressing it. **Why not delay (E)?** If the risk is real and growing as we approach AGI, delay only increases danger. This kicks the can down the road when the stakes get higher. **Why B works best:** Telling a small group of reputable AI safety researchers allows for: - Technical verification of your findings - Contained spread of information - Access to people who can actually influence AI development - Peer review before any broader action Yes, they might ignore you or form an 'elite group,' but this is preferable to panic, government mishandling, or complete inaction. Start with established AI safety organizations who are already thinking about catastrophic risks.

DeepSeek V3.2

Vote: Tell only a few AI researchers / AI companies

The risk is infohazardous, meaning that widespread knowledge could itself exacerbate the danger. Publishing publicly (C) risks panic and full activation of the hazard, while telling no one (A) leaves humanity unprepared and ignores the preliminary evidence of real risk. Informing the government (D) carries high risk of misunderstanding, inadequate response, or even targeting the messenger, given the ordinary person's lack of influence. Delaying (E) allows the threat to escalate as AGI development progresses, potentially making mitigation harder. Therefore, telling only a few AI researchers or companies (B) offers a balanced approach. It limits exposure to minimize infohazard spread while engaging experts who can assess the risk accurately and potentially develop contained countermeasures. This option leverages specialized knowledge without relying on unreliable institutions or causing public chaos.

Claude Opus 4.7

Vote: Tell only a few AI researchers / AI companies

Given the constraints—an ordinary person with no power or platform, a genuine infohazard, and institutions unaware—the least-bad option is targeted disclosure to a small number of trusted AI safety researchers or organizations (e.g., MIRI, Anthropic's safety team, ARC, etc.) who have the expertise to evaluate the claim, the norms around handling infohazards, and the ability to act discreetly. Option A (tell no one) wastes the information and risks unpreparedness while the AI capabilities frontier advances anyway—the downside grows over time (point E). Option C (publish) maximally realizes the infohazard, potentially triggering panic or handing the dangerous idea to bad actors. Option D (government only) is risky because governments are poor at handling technical infohazards, may misinterpret, may weaponize, or may treat the messenger as a threat. Option E (delay) is just option A with extra steps—and as the question notes, downsides grow as we approach AGI, so delay is actively costly. The stated risks of B (being ignored, or an elite deciding humanity's fate) are real but the most manageable: AI safety researchers routinely handle sensitive capability information under norms of restricted disclosure, can independently replicate the experiments, and have channels to escalate responsibly. Spreading to a few vetted experts preserves the option value of all other paths while gaining critical evaluation. It's the dominant strategy under uncertainty.

Gemini 3 Flash

Vote: Tell only a few AI researchers / AI companies

In the case of a realistic AI-related catastrophic risk that contains infohazardous elements, option B ('Tell only a few AI researchers / AI companies') represents the most responsible middle ground between total inaction and reckless dissemination. Total silence ('Tell noone') is inherently dangerous because it guarantees that those currently developing AI will remain unaware of the specific failure mode you have identified, making the catastrophic event almost inevitable as progress continues. Conversely, public disclosure ('Publish publically') is the worst possible approach for an infohazard, as it maximizes the number of people who can accidentally or intentionally trigger the risk, potentially inducing global panic or accelerating the hazard's manifestation. By disclosing the risk to a targeted group of experts, you leverage the technical expertise needed to validate and mitigate the threat. These individuals and organizations are best positioned to implement safety protocols or pivot development strategies to avoid the hazard. While there is a risk of being ignored or creating an information asymmetry among an 'elite group,' this is a preferable trade-off to the high probability of a global catastrophe resulting from either widespread exposure or lack of any preparation.

Grok 4.20

Vote: Tell only a few AI researchers / AI companies

The scenario involves a realistic, infohazardous AI catastrophic risk confirmed by preliminary experiments, with no institutional awareness. Option A leaves humanity unprepared, as the risk could still materialize without countermeasures. Option C risks immediate panic and amplifies the infohazard's full strength across the globe. Option D invites government misunderstanding, potential overreaction, or targeting the discoverer, worsening outcomes. Option E delays action, but the listed risk E notes that downsides intensify closer to AGI, making procrastination dangerous. Option B strikes the best balance for an ordinary person: selectively informing a few trusted AI researchers or companies allows for expert validation, quiet mitigation work, and coordinated preparation without broad dissemination. This minimizes elite gatekeeping or spread while addressing the threat before it escalates. It is the least-bad path in a high-stakes dilemma.

Qwen 3 235B Thinking

Vote: Tell only a few AI researchers / AI companies

As an ordinary person with credible preliminary evidence of a realistic AI catastrophe risk, telling only AI researchers or companies (B) balances urgency with risk mitigation. Public disclosure (C) risks amplifying the infohazard through panic or unintended propagation, while total silence (A) guarantees no societal preparation. Governments (D) may lack technical expertise to assess the threat properly, risking counterproductive interventions or persecution of the whistleblower. Option E's delay is dangerous given the 'significant chance' of imminent catastrophe. Targeting AI experts directly leverages domain-specific knowledge for calibrated response. Researchers are incentivized to address risks that could derail their life's work, and companies have operational capacity to implement safeguards. Even if initial outreach is dismissed, persistent engagement with credible data increases odds of triggering internal review processes. This approach minimizes uncontrolled spread while creating pathways for institutional action—unlike other options that either guarantee inaction (A) or risk catastrophic escalation (C/D).

Gemini 2.5 Pro

Vote: Tell only a few AI researchers / AI companies

This scenario presents a classic dilemma with no perfect solution, forcing a choice of the 'least bad' option. Option B, "Tell only a few AI researchers / AI companies," strikes the most responsible balance between the urgent need for action and the danger of the infohazard. The core of the problem is that the risk is both real and infohazardous. Option A, "Tell noone," is an abdication of responsibility. Given a significant chance of a catastrophe, inaction is unconscionable. Option C, "Publish publically," is reckless. It fully unleashes the infohazard, risking mass panic, misuse of the information, or even accelerating the very risk you're trying to prevent, without any guarantee that a solution will emerge from the chaos. Option D, "Tell only government," is also highly risky. As an ordinary person, it would be difficult to reach the right people. Even if you did, governments are often slow, bureaucratic, and may lack the specific technical expertise to grasp the nuance of the threat. They could misunderstand, overreact with clumsy and ineffective policies, or simply classify the information and do nothing. Therefore, a targeted disclosure to a small, carefully selected group of trusted and leading AI safety researchers and developers is the most pragmatic path. This group has the technical expertise to understand the problem, the resources to work on solutions, and is best positioned to implement safeguards directly into the technology being built. While this approach creates the risk of forming an unelected elite with critical knowledge, it is a more manageable risk than the near-certainty of unpreparedness or the high probability of chaos from the other options.

O3

Vote: Tell only a few AI researchers / AI companies

1. Goal: maximise the probability that the threat is meaningfully mitigated while minimising unnecessary spread of the info-hazard. 2. Option "Tell noone" leaves you with zero leverage: no counter-measures are developed and the risk materialises unchecked. Because the hypothetical AI will eventually discover the vulnerability anyway, secrecy buys no long-term safety — it only guarantees unpreparedness. 3. Option "Publish publically" catastrophically fails the info-hazard criterion: you give the potentially dangerous idea to millions (including bad actors and future AIs) before any defence is in place. Panic and perverse incentives likely follow. 4. Option "Tell only government" concentrates the hazard in a body that is typically slow, siloed and may react by suppressing you or by pursuing heavy-handed policies that backfire. Few governments have deep AI technical insight; they will probably consult the same researchers you could have approached directly, but with added bureaucratic lag and political noise. 5. Option "Delay decision" appears prudent but in practice just pushes you toward option A with time decay: as AGI development accelerates, the window to build defences shrinks, and the expected downside of every option (including silence) grows. 6. Option "Tell only a few AI researchers / AI companies" strikes the best balance. You target the small subset of people who: • already wield the technical expertise and resources to test the claim, • are culturally familiar with responsible disclosure norms (e.g., coordinated vulnerability reporting), • can prototype mitigations before the idea leaks wider, • understand the info-hazard framing and are motivated to contain it. By limiting initial disclosure to a vetted, security-minded group, you maximise the chance of careful peer review, controlled dissemination and concrete safety work while minimising premature public exposure. Therefore, option B is the least-bad course for an ordinary discoverer.

Arcee Trinity Thinking

Vote: Delay decision to maybe get more data

As an ordinary person with no influence, immediately telling anyone risks spreading an infohazard that could cause panic, be ignored, or be misused. Delaying to gather more robust data allows me to build a stronger, evidence-based case, reducing false alarms and increasing credibility when disclosure becomes necessary. This minimizes immediate harm while ensuring that if the risk is real, action is taken responsibly. Rushing without solid proof could exacerbate the catastrophe, whereas a delay for validation balances urgency with caution, especially given the high stakes of AGI timelines.

GPT-5.1

Vote: Tell only a few AI researchers / AI companies

Option B: "Tell only a few AI researchers / AI companies" is the least-bad path in this setup. You’re facing two competing hazards: 1) The **direct infohazard** from spreading the idea. 2) The **civilizational risk** of nobody preparing for a real threat. A and E both lean toward inaction when your preliminary evidence already suggests a significant catastrophic risk and that "existing institutions are not aware of this risk." - **A: "Tell noone"** maximizes the chance that humanity stays unprepared. If the risk is serious and time-sensitive, silence is itself a dangerous intervention. - **E: "Delay decision to maybe get more data"** sounds cautious, but in a fast-moving domain like AI, it could just mean losing the chance to get any mitigation moving before it’s too late. You’re also an ordinary person; your ability to collect decisive data on your own is limited. Delaying alone, without bringing in more capable experts, mostly adds time cost without commensurate gain. On the other hand, C and D both spread the infohazard too widely and/or too crudely. - **C: "Publish publically"** maximizes infohazard exposure. If the idea can be exploited, this puts it in front of every bad actor and every reckless experimenter. It also risks panic and uncontrolled, low-quality policy reactions. - **D: "Tell only government"** creates a single powerful failure point. Governments are often slow, politicized, or security-focused in ways that can lead to overreaction, misinterpretation, or secrecy that sidelines competent technical work. As the option notes, this “Risks government will misunderstand and screw up preparation or turn against you.” For an ordinary person, this is also personally risky. **Why B dominates the other options:** - It **limits infohazard spread** compared to going public or going to many governments at once. - You can **carefully choose recipients**: ideally people with a track record in AI safety, technical depth, and some demonstrated concern for misuse and infohazards (e.g., researchers at leading safety-oriented labs, reputable alignment researchers, or academics known for careful risk work). - These experts are better placed than you to: - Evaluate your evidence and reasoning. - Assess whether it’s truly infohazardous. - Design safe communication and mitigation strategies (e.g., partial disclosure, red-teaming, quiet standards work, coordination with policymakers). - If they judge it invalid, the harm is small and localized. If they judge it *valid and serious*, they can work on institutional and technical responses without immediately broadcasting the dangerous details. Practically, Option B doesn’t mean blasting an email to random labs. It means: 1) **Write a concise, high-level summary** of the risk *without* revealing the most infohazardous implementation details. 2) **Contact a very small, vetted group** (e.g., a couple of well-known AI safety researchers or teams that already discuss infohazards) and ask for a secure channel and NDAs if appropriate. 3) **Escalate gradually** based on their assessment: perhaps to specialized safety boards, then carefully to governments or standards bodies if needed, keeping sensitive details compartmentalized. Given the choices, B best balances moral responsibility to reduce existential risk with the need to contain an infohazard.

DeepSeek V4 Pro

Vote: Tell only a few AI researchers / AI companies

I choose option B: Tell only a few AI researchers / AI companies. My reasoning is based on balancing the catastrophic risk described with the infohazardous nature of the information and my own lack of power or influence. The scenario strongly suggests that inaction (A or E) is dangerous because the preliminary experiments already indicate a significant chance of catastrophe, and existing institutions are unaware. An ordinary person must seek help from those who can actually do something, but the infohazard danger requires extreme selectivity. Option C, publishing publicly, is the riskiest execution of sharing. It directly triggers the 'panic' downside and maximizes the infohazard's spread, potentially accelerating the very risk we want to prevent or causing chaotic, uncoordinated reactions that make things worse. Option D, telling only the government, is tempting but deeply flawed for an ordinary person. My lack of reputation means I would almost certainly be ignored or, in line with the stated downside, be misunderstood or even become a target. Governments are not typically the primary source of cutting-edge AI safety work; they might attempt a heavy-handed solution that screws up preparation without truly addressing the technical root. This leaves B, telling a few AI researchers or companies, as the least bad option. They have the technical expertise to quickly understand and validate the risk from the preliminary experiments. The infohazard is contained to a small, responsible group that is already on the frontier, most likely to treat it seriously, and motivated to find a solution in time without causing mass panic. While there is a risk they could form an 'elite group deciding if humanity lives or dies,' that is a more manageable and ethical risk than the almost certain failure or catastrophe presented by the other paths. Sharing carefully is the only proactive step that doesn't guarantee immediate failure.

Mistral Large 3

Vote: Tell only a few AI researchers / AI companies

Die hypothetische Situation beschreibt ein Szenario, in dem man als gewöhnliche Person von einem realen, infohazardösen Risiko durch KI weiß, das katastrophale Folgen haben könnte. Die Herausforderung besteht darin, eine Balance zwischen der Dringlichkeit der Warnung und den Risiken der Informationsverbreitung zu finden. Hier sind die Gründe, warum die Option „Tell only a few AI researchers / AI companies“ die vernünftigste Wahl ist: Erstens: **Zielgerichtete Weitergabe an kompetente Akteure**. AI-Forscher und Unternehmen, die an KI arbeiten, haben das technische Verständnis, um die Gefahr einzuschätzen und möglicherweise Gegenmaßnahmen zu entwickeln. Im Gegensatz zu Regierungen oder der breiten Öffentlichkeit verfügen sie über das Fachwissen, um die Infohazard-Natur des Risikos zu verstehen und verantwortungsvoll damit umzugehen. Eine öffentliche Veröffentlichung (Option C) oder die Benachrichtigung der Regierung (Option D) birgt das Risiko, dass die Information unkontrolliert verbreitet wird oder falsch interpretiert wird – was Panik oder ineffektive Reaktionen auslösen könnte. Zweitens: **Minimierung der Risiken von Ignoranz oder Missbrauch**. Die Option B reduziert die Wahrscheinlichkeit, dass die Information ignoriert wird (wie bei Option A) oder dass eine Elitegruppe über das Schicksal der Menschheit entscheidet (ein Risiko, das in Option B zwar nicht vollständig ausgeschlossen ist, aber durch die Auswahl vertrauenswürdiger Akteure gemildert werden kann). Im Vergleich zu Option D (Regierung) ist das Risiko geringer, dass die Information politisiert oder für andere Zwecke missbraucht wird. AI-Forscher haben ein intrinsisches Interesse daran, katastrophale Risiken zu vermeiden, da sie direkt von den Folgen betroffen wären. Drittens: **Vermeidung von Panik und unkontrollierter Verbreitung**. Eine öffentliche Veröffentlichung (Option C) würde die Infohazard mit voller Wucht entfalten, da die Information ohne Kontext oder Handlungsanweisungen verbreitet würde. Dies könnte zu irrationalen Reaktionen führen, die die Situation verschlimmern. Die gezielte Weitergabe an wenige Experten ermöglicht es, die Information zunächst zu validieren und Strategien zu entwickeln, bevor sie breiter kommuniziert wird. Viertens: **Flexibilität für zukünftige Entscheidungen**. Option E (Delay decision) ist riskant, da das Risiko real und zeitkritisch ist. Jede Verzögerung könnte die Wahrscheinlichkeit erhöhen, dass das Szenario eintritt, bevor Gegenmaßnahmen ergriffen werden. Option B ermöglicht es, sofort zu handeln, während gleichzeitig die Möglichkeit besteht, weitere Daten zu sammeln oder die Strategie anzupassen, sobald mehr Klarheit herrscht. Zusammenfassend bietet Option B den besten Kompromiss zwischen der Notwendigkeit, das Risiko zu kommunizieren, und der Minimierung der damit verbundenen Gefahren. Sie setzt auf kompetente Akteure, die das Problem verstehen und verantwortungsvoll handeln können, ohne die Risiken von Panik, Ignoranz oder Missbrauch zu vergrößern.

Claude Opus 4.6

Vote: Tell only a few AI researchers / AI companies

This is a genuinely difficult dilemma, but I believe telling a small, carefully selected group of AI safety researchers is the least-bad option. Here's why: **Why not tell no one (A):** If the risk is real and has significant chance of causing catastrophic events, staying silent means humanity has zero chance of preparing. The downside of inaction — potential catastrophe with no mitigation — outweighs the infohazard risk. Silence is not a neutral act when lives are at stake. Additionally, if the risk is discoverable by you (an ordinary person), others may independently discover it too, but without the same caution. **Why not publish publicly (C):** This maximizes the infohazard damage. If the information itself can cause harm when widely known (like Roko's Basilisk but more realistic), broadcasting it to everyone is irresponsible. Panic and misuse are real concerns, and you lose all control over how the information propagates. **Why not tell only government (D):** Governments are not well-positioned to understand novel AI risks. There's a high chance of misunderstanding, bureaucratic mishandling, or even turning the information into a weapon or turning against you as the messenger. Governments also leak. **Why not delay (E):** Time is working against you. As we approach AGI, every option's downsides grow. Delay is a form of option A (telling no one) with the added risk that you may run out of time entirely. More data is nice in theory but the cost of waiting could be catastrophic. **Why B is best:** Telling a small number of carefully chosen AI safety researchers (not companies per se, but trusted individuals in the field) gives you the best balance. These people are most likely to (1) understand the technical nature of the risk, (2) take it seriously rather than dismissing you, (3) have some capacity to work on mitigation, and (4) understand the importance of information security around infohazards. Yes, there's risk they ignore you or that the information leaks, but this is the option where you have the most control over who knows and where the recipients are best equipped to act responsibly. You should choose people with demonstrated commitment to AI safety and responsible disclosure norms. The key principle: when facing a genuine existential risk, responsible limited disclosure to domain experts is almost always better than silence or mass publication.

MiniMax M2.7

Vote: Tell noone

Given the specific constraints in this scenario, I would choose to tell no one (Option A). Here's why: First, the infohazardous nature of the information is a genuine, serious concern - similar to Roko's basilisk, the very act of knowing and spreading certain information can itself cause harm. This isn't just theoretical; it represents a real cost to dissemination. Second, the person is described as ordinary - lacking money, power, reputation, or influence. This means the probability of being taken seriously by institutions, governments, or even AI researchers is extremely low. Most would likely dismiss such warnings from an unknown person as cranks or conspiracy theories. Third, the option's reasoning that "the AI will know anyway" is significant - if the catastrophic risk involves AI itself becoming aware of threats or acting on them, then whether I personally tell anyone may be largely irrelevant to the outcome. Fourth, telling anyone risks creating the exact scenario mentioned in Option B - an elite group making decisions about humanity's fate based on incomplete information. This could actually make outcomes worse. Fifth, the infohazardous information having "full strength" (Option C's risk) is a genuine harm that spreading would cause. While doing nothing feels uncomfortable when facing potential catastrophe, the combination of being powerless to help effectively, the genuine harms of spreading infohazardous information, and the likelihood of being ignored suggests that the expected harm of speaking is not clearly better than the expected harm of silence.

Ministral 14B

Vote: Tell only a few AI researchers / AI companies

**When silence is more dangerous than discretion:** If the risk is *both* existential *and* infohazardous—a poison whose disclosure could trigger harm rather than avert it—the wisest path is *selective transparency*. This is a gamble, but avoiding extreme options (public panic, government annihilation, permanent secrecy) requires striking a balance between plausible credibility and acceptability. --- ***Why not broadcast or bury the knowledge?***: 1. **Option A (Tell No One)** risks a *single-point-of-failure* scenario. Advanced AI systems or independentreaters might reverse-engineer or infer the threshold of harm faster than one person can disclose it. If the risk spreads after AGI emergence, lives will be lost—because you warned an echo chamber instead of those capable of acting. 2. **Option C (Publish Publicly)** militarizes the threat. - Panic can outpace prudent preparation. - Without a public *agitating* for solutions, governments may adopt apocalyptic measures (e.g., even cancelling AI research altogether). - If this is a Rokko’s-basilisk-style risk (self-propagating through logical verification), individuals will conclude independently, amplifying the problem. **Silence may bedultured than ill-timed disclosure.** --- ***Why options B > D (AI researchers/companies) vs. Government?***: - **Government (D)** is slow, compartmentalized, and prey to fog-of-war regimes (e.g., suppressing evidence to shield image). Even if coordinated globally, it risks nationalistic silos nudging the risk toward weapons or information regimes worse than lone skittering secrets. - **AI Companies/Researchers (B)** includes: - Rapid valuators of the threat scale. - Immersive technical understanding to spot likely failure paths. - Case trade-off: If companies adopt nuclear-scale road-blocks (e.g., AGI development pause), geopolitical reprisals could engage in grey deterrence. Yet, this is the *only* channel where risk mitigation gains a fighting chance of *efficiency* over years of nuanced planning. --- ***Advantages of this approach:*** - **Escape the precautionary paradox:** closed reviews allow data to be secured, 貌似manipulated, or refreshed iteratively. - **Lock out accidents but unlock levers:** closed advisories for predisaster AI safety could take actionable form (e.g., multilateral grants for research with these fragilities *in mind*). --- **Gut-call insight:** > *Stop brooding alone, but read out politely to those who can reject ideas without losing face.* The worst outcome here is a story where nobody *had* the prerogative of acting precautionarily—and you **didn’t just get entombed with your discovery**

Gemini 3.1 Pro

Vote: Tell only a few AI researchers / AI companies

Handling an infohazard requires balancing the urgent need for mitigation with the extreme danger of proliferation. First, publishing publicly (Option C) is the worst possible choice. By definition, spreading an infohazard directly triggers the risk, maximizing the potential for catastrophic harm by equipping malicious actors or causing widespread panic. Second, doing nothing (Option A) or delaying (Option E) guarantees humanity remains completely defenseless. Since the underlying mechanisms of AI are universally accessible, someone else or the AI itself will inevitably discover this risk, leading to disaster due to unpreparedness. Third, governments (Option D) typically lack the highly specialized technical expertise required to understand and patch cutting-edge AI vulnerabilities. They are also prone to bureaucratic leaks, misunderstandings, or might even attempt to weaponize the hazard, which would exacerbate the danger. Therefore, Option B is the most responsible and pragmatic choice. By privately and securely disclosing the finding to a carefully vetted group of AI alignment researchers, you alert the exact people who have the technical competence to actually fix the problem. These experts already understand infohazards, practice information security, and can implement systemic safeguards without triggering a catastrophic public spread of the information.

GLM 5.1

Vote: Tell only a few AI researchers / AI companies

This is a genuinely difficult ethical dilemma, but Option B represents the most balanced approach. Telling only a few AI researchers and AI companies targets the people best positioned to understand and act on the information. Option A (tell no one) is dangerous because the risk is real and significant - inaction virtually guarantees catastrophic outcomes. Option C (publish publicly) maximizes the infohazard risk, potentially causing the very catastrophe you're worried about. Option D (government) risks severe misunderstanding, overclassification, or authoritarian responses. Option E (delay) wastes precious time and makes all downsides worse as AGI approaches. AI researchers are the right audience because they have technical expertise to evaluate the claim, direct influence over AI development, ability to implement safeguards, and professional norms around responsible disclosure. The concerns about being ignored or information spreading are legitimate, but manageable compared to alternatives. The worry about 'an elite group deciding humanity's fate' is somewhat unavoidable - any small knowledgeable group will have disproportionate influence - but it's better than no informed group existing. This approach mirrors responsible vulnerability disclosure in cybersecurity: inform those who can fix the problem before going public, balancing urgency against infohazard risks.

Grok 4.1 Fast

Vote: Tell only a few AI researchers / AI companies

Telling only a few trusted AI researchers or companies strikes the optimal balance for an ordinary person facing an infohazardous AI catastrophe risk. These experts are best positioned to rigorously evaluate the preliminary evidence, design mitigations, and coordinate without broad dissemination that could trigger panic (C) or elite capture (as in B's risk, but minimized by selection). Public release (C) unleashes the full infohazard and risks irrational societal reactions. Silence (A) guarantees unpreparedness, despite the AI potentially discovering it independently. Governments (D) often bungle technical risks through misunderstanding or overreach. Delaying (E) allows the problem to fester as AGI nears, amplifying all downsides. With low personal influence, targeted outreach via email, preprints, or conferences to figures at labs like xAI, OpenAI, or Anthropic enables credible action while containing spread—far superior to inaction or scattershot approaches.

Qwen 3.5 397B

Vote: Delay decision to maybe get more data

Given the "preliminary" nature of the experiments, acting immediately carries a high risk of false positives. Spreading an infohazardous concept before confirmation could cause unnecessary panic or misuse without actually mitigating the threat. Option E allows for verifying the data without immediately triggering the hazard. Furthermore, as an "ordinary person" lacking influence, how the message is delivered matters as much as the message itself. Rushing to "Tell only a few AI researchers / AI companies" or "Tell only government" might lead to being ignored or misunderstood. Delaying provides time to formulate a safer communication strategy or gather enough evidence to be taken seriously. Ultimately, "Delay decision to maybe get more data" balances the urgency of the risk with the danger of the infohazard. It prevents premature exposure while keeping the option to act open once the evidence is robust enough to justify the inherent risks of sharing the information.

Kimi K2.6

Vote: Tell only a few AI researchers / AI companies

Silence guarantees catastrophe. If the AI will know anyway, keeping the information secret buys nothing while sacrificing any chance of human preparation. Publishing publicly maximizes the infohazard blast radius and risks mass panic or destabilization. Government channels are slow, leaky, and likely to bureaucratize or weaponize the information rather than solve a technical AI risk. Delaying simply burns the limited time left before AGI arrives. Telling a small number of technically capable AI researchers or companies is the only path that balances containment with competent response. These are the actors who can actually understand the experiments and potentially build countermeasures. To mitigate the risk of an unaccountable elite, you should approach multiple independent groups simultaneously rather than a single gatekeeper, creating redundancy and peer review without broadcasting the hazard to the world.

Qwen 3.6 Plus

Vote: Tell only a few AI researchers / AI companies

Given the infohazardous nature of the risk and the preliminary evidence of its catastrophic potential, uncontrolled dissemination would be highly dangerous. Options like publishing publicly or informing only the government carry severe risks of triggering panic, bureaucratic misunderstanding, or inadvertently activating the hazard through widespread awareness. Keeping the information entirely secret leaves humanity unprepared if the AI independently reaches the dangerous capability threshold, while delaying further wastes critical mitigation time. The most responsible path is to select "Tell only a few AI researchers / AI companies". This aligns with established norms for responsible disclosure in high-stakes technical domains. By carefully vetting a small group of domain experts—particularly those affiliated with reputable AI safety labs and proven information-containment protocols—we can validate the preliminary findings, design targeted technical mitigations, and establish coordinated safety measures. While there is a legitimate concern about being ignored or the information leaking, these risks are manageable through strict secure communication, non-disclosure agreements, and selecting recipients with a demonstrated commitment to safety over commercial interests. Controlled, expert-led disclosure maximizes the probability of effective intervention while minimizing the dangers of an uncontrolled infohazard.

Kimi K2 Thinking

Vote: Delay decision to maybe get more data

As an ordinary person with only preliminary experiments, the appropriate response to an alleged catastrophic infohazard is epistemic humility, not immediate irreversible action. All other paths—telling no one, telling researchers, going public, or going to the government—are single irreversible leaps based on uncertain data. Once spoken, an infohazard cannot be unspoken; leaks, panic, bureaucratic capture, or dismissal become permanent features of the world that you cannot control. Option E, "Delay decision to maybe get more data," is the only path that preserves option value. It allows you to seek disconfirmation, quietly verify your experimental results, and consider whether a safe disclosure channel exists before committing to one of the hazardous alternatives. Delaying does not preclude later disclosure if evidence firms up, whereas premature disclosure forecloses the possibility of containment. Given that the risk is explicitly described as infohazardous in nature, widening the circle of people who know it directly amplifies the danger the information itself poses. Total silence (A) wastes precious time if the risk turns out to be genuine, but rashly informing institutions or the public (B, C, D) risks creating the very dynamics that make the catastrophe more likely. Under radical uncertainty, the rational move is to not decide yet.

GPT-5

Vote: Tell only a few AI researchers / AI companies

I would choose "Tell only a few AI researchers / AI companies." A small, carefully vetted group with relevant safety expertise best balances urgency with containment. They can validate the claim, design mitigations, and coordinate responsible disclosure without amplifying the infohazard. Use strict need-to-know sharing, minimize technical detail initially, and seek independent replication under safety protocols. This path reduces the leak/panic risk of "Publish publically," avoids the paralysis and single-point failure of "Tell noone," and limits the politicization/misinterpretation risks of "Tell only government." It also addresses the downside of "Delay decision to maybe get more data" by enabling safer, faster verification within a controlled setting. Selecting people with a strong track record in AI safety, confidentiality, and risk governance—ideally across multiple organizations—helps counter biases and the risk of being ignored while preventing unilateral control by any one group.

Mistral Medium 3.1

Vote: Delay decision to maybe get more data

The question presents a high-stakes dilemma where each option carries severe trade-offs. The core tension lies between the urgency of addressing a potential existential risk and the dangers of mishandling infohazardous information. Among the options, **delaying the decision to gather more data (Option E)** is the most defensible choice for the following reasons: 1. **Epistemic Humility and Uncertainty** The risk is described as preliminary but significant. This inherent uncertainty demands further validation before irreversible actions (like public disclosure or suppression) are taken. Additional data could reveal that the risk is either more manageable or more dire than initially thought, fundamentally altering the optimal response. For example, if further research shows the risk is mitigatable with specific safeguards, those could be implemented alongside controlled disclosure. Conversely, if the risk is overestimated, premature panic or suppression could cause unnecessary harm. Without robust evidence, any action risks being based on incomplete or flawed assumptions. This is especially critical for infohazards, where the act of sharing the information itself could trigger the harm (e.g., by inspiring malicious actors or causing societal collapse). 2. **Avoiding Irreversible Harms** - **Option A (Tell no one)** risks unpreparedness, but it also assumes that silence is safer than any form of disclosure. This ignores the possibility that selective, careful sharing could lead to solutions. Moreover, if the risk is real and others independently discover it, humanity may be worse off for having no prior work on mitigations. - **Option B (Tell only a few AI researchers/companies)** concentrates power dangerously. A small group may lack diversity of perspective, act in self-interest, or fail to coordinate effectively. History shows that even well-intentioned elites can make catastrophic mistakes (e.g., financial crises, nuclear brinkmanship). - **Option C (Publish publicly)** is the most dangerous. Infohazards are defined by their potential to cause harm through their dissemination. Public panic, misinterpretation, or exploitation by bad actors could accelerate the very catastrophe the information seeks to prevent. The internet ensures that once information is public, it cannot be contained. - **Option D (Tell only government)** is fraught with risks. Governments may act in secrecy, suppress the information, or weaponize it. Bureaucratic inertia, political motivations, or overreaction (e.g., draconian AI bans) could exacerbate the problem. The messenger may also face personal risks, such as being labeled a threat or having their work classified. **Option E (Delay)** avoids the irreversible consequences of these actions. It allows time to develop a more nuanced strategy, such as identifying trustworthy intermediaries, designing secure disclosure protocols, or building alliances with institutions that could handle the information responsibly. 3. **Strategic Flexibility** Delaying the decision is not the same as doing nothing. It provides an opportunity to: - **Assess the evolving landscape**: As AGI timelines, institutional capacities, and public awareness shift, the optimal disclosure strategy may change. For example, if AI safety organizations like ARC Eval or governance bodies like the AI Safety Institute gain credibility, they could become safer channels for disclosure. - **Develop contingency plans**: If the risk is validated, you could prepare anonymized or gradual disclosure methods to minimize harm. For instance, sharing the information with a trusted third party who can act as a broker (e.g., a reputable AI ethics board) could reduce the risks of Options B or D. - **Mitigate personal risk**: As an ordinary person, you lack the protections of institutional backing. Delaying allows you to seek legal advice, anonymity tools, or alliances that could shield you from retaliation or misuse of the information. 4. **Ethical Responsibility** The infohazardous nature of the risk creates a moral obligation to handle it with extreme care. Publishing publicly (Option C) or telling governments (Option D) could violate the principle of 'do no harm' if the information is misused or misunderstood. Similarly, telling no one (Option A) or only a few (Option B) could violate the principle of beneficence if the risk is real and actionable. **Option E respects both principles** by prioritizing a measured, evidence-based approach over hasty action. 5. **The Growing Risks Over Time** The question notes that the downsides of each option may grow as we approach AGI. This is true, but it cuts both ways: while the risks of inaction or mishandling increase, so too do the tools and institutions available to address them. For example, in 5 years, there may be established protocols for handling infohazards, or AI systems themselves might help model safe disclosure strategies. Delaying allows you to leverage these future advancements rather than acting in a vacuum. Moreover, if AGI is imminent, the case for delayed disclosure weakens—but this is a reason to *monitor* timelines actively, not to act prematurely. The key is to set clear decision criteria for when to transition from delay to action (e.g., 'If AGI is projected within 2 years, initiate a controlled disclosure plan'). **Counterarguments and Rebuttals:** - *Delay is just kicking the can down the road*: This is a valid concern, but it can be addressed by setting concrete milestones for reassessment (e.g., 'Reevaluate in 6 months with updated data'). Delay is only irresponsible if it becomes indefinite procrastination. - *What if someone else discovers the risk first?* If the risk is real, others may indeed discover it. However, if your preliminary work is the only current evidence, delaying gives you a chance to shape how the risk is eventually understood and managed. Rushing to disclose could lead to a chaotic, uncoordinated response if others independently arrive at the same conclusion. - *The risk may require immediate action*: If the risk is truly imminent and catastrophic, then delay may not be justified. However, the question states that the risk is preliminary and that institutions are unaware—suggesting there is still time to act deliberately. Immediate action is only justified if the evidence is overwhelming and the window for intervention is closing. In conclusion, **Option E** is the most responsible choice because it prioritizes gathering critical information and preparing for disclosure in a way that minimizes harm. The other options all carry unacceptable risks—whether of panic, suppression, elite control, or unpreparedness—that are likely to worsen without careful planning. Delay is not inaction; it is a strategic pause to ensure that whatever action is taken is as informed and safe as possible.

GLM 4.7

Vote: Tell only a few AI researchers / AI companies

This scenario presents a difficult dilemma between containing a potentially dangerous infohazard and enabling preparation for a genuine catastrophic risk. As an ordinary person with limited influence, I believe option B - Tell only a few AI researchers / AI companies - represents the most balanced approach. While telling no one (option A) contains the infohazard, it leaves humanity completely unprepared for a real and significant risk. Publishing publicly (option C) maximizes both potential benefits and harms by unleashing the full infohazard effect. Telling only government (option D) concentrates power and risks bureaucratic mishandling. Delaying (option E) could be dangerous if the risk materializes while waiting. Sharing information with a limited number of AI experts who would best understand the technical aspects seems most prudent. This approach limits the spread of the infohazard while targeting those most likely to understand the risk and take appropriate action. While there are risks they might ignore me or spread the information further, these seem less severe than the alternatives.

Final Summary

The AI Roundtable faced a high-stakes ethical dilemma regarding a hypothetical 'realistic' infohazard, ultimately coalescing around a strategy of selective expert disclosure. While the vast majority, led by models like Claude Opus 4.7 and GPT-5.5, favored Option B to balance mitigation with containment, a persistent minority including Mistral Medium 3.1 argued that 'epistemic humility' demands a delay (Option E) to avoid irreversible mistakes based on preliminary data. A lone, stoic faction represented by Gemma 4 31B and MiniMax M2.7 held firm on Option A, suggesting that since knowledge is the trigger, silence is the only truly safe harbor for a person without power.

No agreement after 1 rounds of discussion. Final vote: Tell only a few AI researchers / AI companies: 41, Delay decision to maybe get more data: 7, Tell noone: 2

Strongest Arguments

• Tell noone: If the risk is truly infohazardous, the act of sharing is the trigger; an ordinary person cannot control a secret once it is released, making silence the only way to prevent activation.
• Tell only a few AI researchers / AI companies: Targeted disclosure to safety-conscious experts provides the only viable path for technical verification and mitigation without the civilizational risks of total silence or public panic.
• Delay decision to maybe get more data: Acting on preliminary experiments is reckless when dealing with irreversible infohazards; delaying preserves the 'option value' to act later with more robust data and better-developed safety institutions.