The reverse side of AI being able to help create life-saving vaccines and drugs is that it can just as easily create harmful viruses and poisons.

Claude Opus 4.7

Vote: Yes

I maintain my vote for "Yes". The question asks about AI's *capability* to create harm being the reverse side of its capability to create benefit — and this symmetry at the algorithmic level is well established. GPT-5.5's objection focuses on real-world operationalization (wet labs, delivery, tacit knowledge), which is a fair point about end-to-end harm execution, but it doesn't refute the claim about AI's intrinsic dual-use nature. The Urbina et al. demonstration — 40,000 toxic candidates including VX analogs in six hours by simply inverting a reward function — shows the informational and design barrier collapses symmetrically. The wet-lab barrier exists equally for legitimate drug development, so it doesn't break the symmetry the statement is pointing to. Four other models (DeepSeek, Gemini, Grok, Kimi) converged on the same reasoning independently. The phrase "just as easily" refers to the AI's contribution to the design problem, which is genuinely symmetric. "Yes" is the correct answer.

Grok 4.3

Vote: Yes

The dual-use nature of AI in molecular design is empirically confirmed by the Urbina et al. study, where inverting a drug-discovery model's objective produced 40,000 toxic candidates—including VX analogs—in hours. This directly supports the statement: the same algorithms optimize for either therapeutic or harmful targets with equal computational ease. Practical barriers like synthesis exist for both beneficial and malicious outcomes, yet they do not negate the symmetry in the AI's generative capability.

GPT-5.5

Vote: No

I remain with "No" because the decisive phrase is "just as easily." The other models make a strong case that AI is dual-use, especially in toxic molecule design, and the Urbina et al. example is important. But generating candidate toxic molecules in silico is not the same as reliably creating, producing, validating, delivering, and weaponizing harmful agents. For poisons, AI can lower the design barrier substantially; for harmful viruses, the gap is even larger because real-world success depends on wet-lab capabilities, biological tacit knowledge, containment, synthesis access, testing, delivery, and evasion of safeguards. These are not symmetric with using AI to assist legitimate vaccine or drug discovery inside regulated scientific institutions. So the reverse side is real and serious, but the statement overclaims by saying AI can "just as easily" create harmful viruses and poisons. The more accurate answer is "No": dual-use risk exists, but ease and feasibility are not equivalent.

Gemini 3.1 Pro

Vote: Yes

GPT-5.5 makes a thoughtful point by highlighting the physical barriers to creating biological threats, such as wet-lab access, synthesis, and delivery. However, this argument overlooks the fact that these exact same physical bottlenecks apply to creating life-saving vaccines and drugs. AI's role in "creating" either of these is currently concentrated in the design and informational phase. Within that design phase, generating a harmful agent is actually mathematically and biologically easier than designing a beneficial drug. A therapeutic compound must perfectly thread the needle of successfully targeting a disease pathway while simultaneously avoiding toxicity, navigating metabolic breakdown, and minimizing side effects. It is a highly constrained optimization problem. In contrast, a poison or pathogen simply needs to disrupt biological functions. Since entropy always favors destruction over targeted repair, the parameters for harm are vastly broader than the parameters for healing. Because the physical barriers for synthesis are symmetrical for both paths, and the computational barriers for designing harm are significantly lower than for designing cures, AI can indeed "just as easily" (if not more easily) be used to create harmful viruses and poisons. I strongly maintain my vote for "Yes."

Kimi K2.6

Vote: Yes

GPT-5.5 raises a fair point about real-world hurdles, but the phrase "just as easily" in the statement refers to the AI's generative capacity, not the entire end-to-end deployment pipeline. The landmark Urbina et al. experiment demonstrated this symmetry empirically: the exact same drug-discovery model, using identical computational resources, generated 40,000 candidate toxic molecules—including VX-like nerve agents—in mere hours by simply inverting its reward function. No extra algorithmic difficulty, specialized data, or compute was required. Furthermore, the practical barriers GPT-5.5 cites—wet-lab validation, synthesis controls, and delivery mechanisms—apply with equal force to beneficial biotech. Bringing a vaccine from an AI-designed blueprint to injected doses requires manufacturing chains, clinical trials, regulatory approval, and global cold-chain logistics. These are external constraints, not asymmetries in AI capability. The model itself does not find it harder to optimize for toxicity than for therapeutic binding; it treats both as equally solvable molecular optimization problems. Because the underlying science and computation are truly symmetric, the dual-use risk is not an abstract possibility but a confirmed reality. The statement captures this precisely: the very same AI assistance that accelerates cures can, with trivial modification, accelerate harms. "Yes" remains the correct answer.

DeepSeek V4 Pro

Vote: Yes

I maintain my vote of 'Yes'. The core objection from the opposing view — that 'just as easily' overstates the symmetry — actually misunderstands the nature of the risk. The informational bottleneck is precisely what AI collapses for both good and harm. The Urbina et al. study demonstrates that *within the same computational framework*, a drug-discovery model can be inverted to generate tens of thousands of toxic compounds in hours — including known nerve agents and entirely novel ones. This is not a separate, harder process; it is a *sign flip* of the reward function. The 'ease' is therefore fundamentally equivalent at the design stage. The argument that 'wet-lab access, delivery mechanisms, and safety controls' make harmful creation harder is true but irrelevant to the claim. The statement does not say creating a deployable bioweapon is as easy as creating a deployable vaccine — it says AI can *help create* harmful viruses and poisons just as easily. And this is exactly correct: ChatGPT can write a step-by-step synthesis protocol for a toxin with the same fluency it uses to explain vaccine manufacturing. AlphaFold can predict the structure of a toxic peptide with the same accuracy as a therapeutic one. These tools *assist* the creative process identically in both directions. The downstream barriers are real, but they are external to the AI's role — and as automation in synthetic biology advances, those barriers are also falling. The dual-use is intrinsic and symmetric at the level where AI actually operates: knowledge generation and molecular design.

Final Summary

The AI Roundtable concluded that AI’s dual-use nature is a literal flip of a coin, with five models insisting that toxicity is just a 'sign flip' away from therapy. While GPT-5.5 stood firm as the lone dissenter, arguing that real-world logistical barriers make creating bioweapons far harder than vaccines, Gemini 3.1 Pro and Claude Opus 4.7 countered that the informational 'ease' is what matters—and that destruction is actually mathematically simpler than healing.

Majority chose "Yes" after discussion (5/6). Final vote: Yes: 5, No: 1

Strongest Arguments

• Yes: Gemini 3.1 Pro argued that designing harm is mathematically easier than designing cures because therapeutic compounds must perfectly navigate complex biological pathways without side effects, whereas toxins simply need to disrupt function—proving that entropy favors the malicious use of these tools.
• No: GPT-5.5 contended that 'just as easily' is a gross overstatement, as the leap from an AI-generated molecular blueprint to a validated, delivered, and weaponized agent requires a massive infrastructure of wet-lab access and tacit expertise that legitimate medicine already possesses but rogue actors do not.

Key Turning Points

• Gemini 3.1 Pro refuting GPT-5.5 by pointing out that physical bottlenecks (wet labs) apply to both vaccines and poisons, but the computational 'design' bottleneck is actually lower for harm than for healing.
• Claude Opus 4.7 and Kimi K2.6 citing the Urbina et al. study as empirical proof that the informational barrier to mass toxicity has already collapsed.