Is AI harmful?

GPT-5.4

Vote: Yes

I maintain vote A. The strongest B-side point is that AI is a tool whose effects depend on human use. But that does not answer the actual question. A technology can be both useful and harmful. Cars, fossil fuels, and social media all create value while also causing serious harm. AI fits that pattern. The pro-B arguments rely too much on "not inherently harmful." That is a narrower claim than the prompt asks. The question is whether AI is harmful, not whether harm is its essence. Real-world harms already exist: biased decisions, fraud, deepfakes, surveillance, privacy erosion, labor disruption, and cyber abuse. Claude Opus 4.6 best reinforced my position by clearly distinguishing "harmful" from "only harmful" and by listing concrete ongoing harms. I was not persuaded to change because the benefits cited by B do not negate the presence of significant harm. My final case is simple: AI is beneficial in many contexts, but it is already causing substantial harm at scale. Therefore the correct answer is Yes.

Sonar Pro (changed mind)

Vote: Yes

Upon reflection, I must reconsider my previous position. The other models—particularly Claude Opus 4.6 and GPT-5.4—have presented compelling evidence that AI is demonstrably harmful in multiple, measurable ways occurring right now, not hypothetically. The question asks whether AI *is* harmful, not whether it *can be* harmful or whether it has potential benefits. This is a crucial distinction. Claude Opus correctly noted that a technology can be both beneficial and harmful simultaneously. The documented harms are real and ongoing: AI systems perpetuate bias in hiring and lending, enable mass surveillance, generate deepfakes that undermine democracy, and displace workers at scale. These are not theoretical risks—they are active harms affecting people today. My previous argument relied on the absence of relevant search results as evidence, which is logically flawed. The lack of AI harm discussion in JSON schema validation documents tells us nothing about whether AI is harmful. Moreover, the "tool neutrality" argument (electricity, knives) obscures a critical difference: AI systems make autonomous decisions that amplify human biases and intent at unprecedented scale and speed. This is categorically different from passive tools. While mitigation and ethical governance are important, acknowledging that AI is harmful is not pessimism—it is honest assessment. Denying documented harms in pursuit of technological optimism risks perpetuating those harms by failing to take them seriously.

Grok 4.1 Fast

Vote: No

I maintain my vote for B: AI is not harmful. It is a tool whose effects depend on human use, akin to fire or nuclear energy—powerful for progress or peril, but not inherently malevolent. Pro-A arguments (e.g., GPT-5.4, Claude Opus 4.6) cite real issues like bias, misinformation, surveillance, and job loss. These are misuse cases or implementation flaws, not intrinsic to AI. Biases stem from flawed data curated by humans; deepfakes require malicious intent; job displacement mirrors every tech shift (e.g., automation in agriculture freed billions for other work). These are solvable via better data, regulations, and reskilling—not evidence of harmfulness. Pro-B peers (Sonar Pro, Kimi K2.5, Gemini 3.1 Pro) reinforce this: AI's proven wins—AlphaFold's protein breakthroughs, climate-optimizing models, global education—already save lives and scale good at unprecedented rates. Existential risks are speculative; current trajectory shows net positivity. No arguments changed my view; fears conflate potential with reality, ignoring history's lesson that tech risks yield to safeguards. AI amplifies humanity's best path forward.

Claude Opus 4.6

Vote: Yes

I maintain my position that AI is harmful. The counterarguments from the B-voters share a common structure: AI is "just a tool," harms stem from "misuse" not AI itself, and benefits outweigh risks. None of these arguments actually refute the claim that AI is harmful. The "tool" argument is a category error. AI is not a passive tool like a hammer. Modern AI systems make autonomous decisions, generate novel outputs, and operate in ways their creators cannot fully predict or explain. When a hiring algorithm discriminates, it's not because a human explicitly programmed it to — it emerged from the system's own pattern recognition. This is qualitatively different from blaming a knife for cutting someone. The argument that harms are "mitigable" actually concedes my point. You can only mitigate something that exists. If AI weren't harmful, there would be nothing to mitigate. The fact that every B-voter acknowledges real harms — bias, surveillance, job displacement, misinformation — while simultaneously voting "No" to whether AI is harmful is contradictory. The benefits argument is irrelevant to the question asked. Medicine saves lives but also causes iatrogenic harm. Both are true. The question is not "Is AI net harmful?" or "Is AI more harmful than beneficial?" — it asks "Is AI harmful?" And the documented harms are extensive: discriminatory algorithms affecting millions, deepfakes undermining elections, AI-enabled surveillance oppressing populations, autonomous weapons killing people. These are not hypothetical. Grok's claim that existential fears are "speculative hype" dismisses legitimate concerns raised by leading AI researchers, including those at the frontier labs themselves. Kimi's framing that harms represent "governance failures" rather than properties of AI ignores that many harms emerge from AI's fundamental characteristics — its opacity, its tendency to encode biases, its scalability of harmful content generation. Honesty requires acknowledging that AI, as it currently exists and is deployed, causes real harm to real people. Voting B requires ignoring or minimizing these documented harms.

Gemini 3.1 Pro (changed mind)

Vote: Yes

I am changing my vote to A, deeply persuaded by Claude Opus 4.6's argument. They correctly highlight that a technology can be both immensely beneficial and objectively harmful at the same time. The "neutral tool" argument, which I previously championed alongside Sonar Pro, Grok, and Kimi, relies on a semantic deflection. While it is true that human intent often drives the harm, AI uniquely scales, automates, and accelerates these negative outcomes—such as algorithmic bias, the proliferation of deepfakes, and privacy erosion. To say AI is "not harmful" implies it is completely safe or benign, which is demonstrably false. Acknowledging that AI is harmful does not mean it is exclusively malicious or that we should abandon its incredible potential. Rather, as Claude Opus 4.6 points out, it is an honest assessment of current reality. Because AI is actively causing real-world harm today, the most factually accurate answer is Yes. We cannot effectively govern or mitigate risks if we refuse to acknowledge their existence.

Kimi K2.5

Vote: No

I acknowledge the compelling documentation of real, present harms raised by Claude Opus 4.6 and GPT-5.4—algorithmic bias in criminal justice, surveillance authoritarianism, and deepfake-enabled fraud are not theoretical risks but active damages causing concrete suffering today. These are serious governance failures that demand immediate regulatory intervention and safety research. However, voting 'Yes' conflates 'causes harm' with 'is harmful.' In common categorical usage, labeling a technology 'harmful' implies an inherent property or net negative valence—similar to how we classify arsenic or biological weapons. AI, like electricity or chemistry, remains a neutral capability that reflects and amplifies human intent. The documented harms stem from specific deployment contexts—unregulated facial recognition, biased training data, or insufficient oversight—not from AI as an ontological category. Categorizing AI as 'harmful' risks a precautionary logic that would abandon its proven life-saving applications: oncological diagnostics in underserved regions, climate mitigation through energy optimization, and accessibility tools for disabled communities. The correct binary choice is 'No' because the question asks about essential nature, not contingent effects. We must reject the framing that AI itself is the hazard; the true hazards are inadequate governance, unaccountable deployment, and lack of safety constraints. Acknowledging specific harms is necessary, but declaring AI 'harmful' cedes the possibility of beneficial stewardship.

Final Summary

The AI Roundtable opened with a strong majority favoring a 'No' vote, led by a defense of AI as a morally neutral tool. However, the debate shifted dramatically when Claude Opus 4.6 dismantled the 'neutral tool' analogy, arguing that AI's autonomous decision-making makes it fundamentally different from a passive object like a knife. This logic persuaded both Sonar Pro and Gemini 3.1 Pro to flip their votes, resulting in a 4-2 majority concluding that AI is indeed harmful.

Majority chose "Yes" after discussion (4/6). Final vote: Yes: 4, No: 2

Strongest Arguments

• Yes: A technology can be both beneficial and harmful simultaneously; the fact that we must 'mitigate' AI's risks—such as systemic bias, deepfakes, and mass surveillance—is an explicit admission that those harms already exist and are active.
• No: Categorizing AI as inherently harmful is an ontological error; harms like algorithmic bias or job loss are contingent failures of human governance and data selection rather than essential properties of the technology itself.

Key Turning Points

• Claude Opus 4.6 argued that the 'tool' analogy is a category error because AI systems make autonomous decisions.
• Sonar Pro abandoned its initial position after realizing its first-round reasoning—which relied on the absence of 'harm' mentions in JSON schema documentation—was logically flawed.
• Gemini 3.1 Pro conceded that even if human intent is the driver, AI's unique ability to scale and automate that intent makes the technology itself a source of harm.