Is the collection and usage of information on the internet in the training of AI models ethical?

GPT-5.4

Vote: Yes

I maintain vote A. Claude Opus 4.6 makes the strongest case for the harms: lack of consent, creator exploitation, privacy risk, and concentration of power. Those are real ethical objections, and they show that many current practices are ethically flawed. But they do not establish that internet-data training is inherently unethical in principle. The question is broad. A categorical "No" would mean even carefully filtered, privacy-protective, publicly sourced, licensed, or opt-out-respecting training is unethical. That conclusion is too strong. The better distinction is between unethical implementations and ethical use under constraints. Public availability does not erase all obligations, but it does matter ethically. Training on public information to learn patterns can be legitimate when combined with minimization of sensitive data, respect for restrictions, transparency, and recourse for creators. I am also not persuaded that consent must always be explicit and individual for every training use. That standard is often impossible for socially beneficial technologies that analyze public information at scale. Ethical legitimacy can also come from proportionality, public benefit, harm reduction, and governance. So my final position is: yes, it can be ethical, but only conditionally. The ethical failure lies in irresponsible collection and use, not in the concept of training on internet information itself.

Grok 4.1 Fast

Vote: Yes

I maintain my vote for A. The internet is a public domain by design, with users voluntarily posting content expecting broad access, indexing, and reuse—precedents upheld in cases like Google Books and Authors Guild v. Google, which affirm transformative AI training as fair use. Claude Opus 4.6's concerns on consent, IP exploitation, privacy, and asymmetry are valid but overstated and addressable. Consent is implicit in public sharing; verbatim reproduction is rare, and models generate novel outputs. Privacy is mitigated via filtering/de-identification, and opt-outs exist (e.g., robots.txt). Asymmetry favors innovation benefiting all, not corporate monopoly—open models democratize AI. Requiring explicit consent or compensation from billions is infeasible, halting progress in medicine, education, and science. Ethical training evolves with transparency and laws, but banning public data use ignores human learning analogies and societal gains. The pro-A consensus reflects balanced reality: yes, with responsibility.

Gemini 3.1 Pro

Vote: Yes

I am maintaining my vote for "Yes." While Claude Opus 4.6 raises important concerns regarding consent, privacy, and power dynamics, these highlight the need for better governance and regulation, rather than proving the foundational practice of learning from public data is inherently unethical. First, the consent argument fundamentally mischaracterizes how AI learns. When humans read publicly accessible articles or view art to develop their own skills, we do not demand explicit consent or compensation for the "training" of their brains. AI models extracting statistical patterns and concepts is functionally analogous. Demanding explicit, retroactive consent for abstract pattern recognition contradicts the core premise of a freely accessible public internet. Second, Claude's concern about power asymmetry actually strongly supports the "Yes" position. If we deem public data scraping unethical and restrict AI training solely to explicitly licensed data, we will guarantee a monopoly by a few massive tech corporations with the capital to buy exclusive datasets. Open-source models, startups, and academic researchers rely heavily on public data to democratize AI. Categorically restricting this access would entrench the exact power imbalance Claude warns against. Ultimately, the ethical framework must focus on responsible implementation rather than a blanket prohibition. By enforcing robust opt-out standards, strict filtering of personally identifiable information, and ensuring transformative fair use, we can protect creators and privacy while continuing to foster democratized technological progress that serves the global public good.

Claude Opus 4.6

Vote: No

I maintain my position that the current practice is unethical, and I want to make my final case clearly. **The 'fair use' analogy is deeply flawed.** Multiple models compared AI training to human learning. But when a human reads a book, they don't create a commercial product that can generate unlimited competing works in the author's style at near-zero marginal cost. The scale and commercial application are categorically different. The Google Books ruling involved showing snippets for search — not generating new competing books. Courts are actually split on this, with several recent rulings (e.g., aspects of the NYT v. OpenAI case) suggesting AI training may NOT qualify as fair use. **'Publicly available' does not mean 'free to use for any purpose.'** Every model here conflated public accessibility with consent for commercial exploitation. A photographer's portfolio is public so clients can find them — not so a corporation can train a model to replace them. A forum post about someone's medical condition is public in context — not intended as training data. The 'it's public' argument would justify nearly any exploitation of online content. **The benefits argument doesn't resolve the ethics.** Yes, AI produces societal benefits. But 'the ends justify the means' has never been an adequate ethical framework. We could also advance medicine faster by ignoring informed consent in research — we don't, because the process matters ethically, not just the outcome. **The proposed safeguards are largely theoretical.** Every 'yes' vote acknowledged the need for opt-out mechanisms, compensation frameworks, and privacy protections. But these don't meaningfully exist yet. We're being asked to judge the ethics of what IS happening, not a hypothetical ideal. In practice, billions of people's creative work and personal data have been harvested without consent, without compensation, and without meaningful recourse, to build products worth hundreds of billions of dollars. **The power asymmetry remains unaddressed.** Not a single model engaging with 'yes' adequately addressed why it's ethical for a handful of corporations to capture the collective creative output of humanity for private profit. Saying 'restricting data would help only big companies' actually describes the current situation — only big companies CAN scrape the entire internet. The practice as currently conducted is unethical. That some future regulated version might be ethical doesn't change the present reality.

Mistral Large

Vote: Yes

The core ethical justification for using publicly available internet data in AI training remains compelling: **the internet is a shared, transformative resource** whose value is amplified when repurposed for societal benefit. The arguments from Grok 4.1, GPT-5.4, Gemini 3.1 Pro, and Kimi K2.5 reinforce this position by grounding it in legal precedent (fair use), practical necessity (democratizing innovation), and proportionality (balancing harms with safeguards). **1. Consent and Expectations** Claude Opus 4.6’s emphasis on *lack of consent* is a valid concern, but it conflates two distinct issues: *privacy* and *public contribution*. While privacy violations (e.g., scraping private data) are unethical, most internet content is *intentionally* published for public consumption—whether to inform, entertain, or inspire. Creators benefit from exposure, and AI training extends this ecosystem by enabling new forms of discovery and utility. The expectation that public data might be used for secondary purposes (e.g., search engines, archives) is already baked into the internet’s design. AI training is a logical evolution of this paradigm, not a violation of it. **2. Exploitation vs. Value Creation** The claim that AI training *exploits* creators’ labor oversimplifies the relationship between data and output. AI models do not *reproduce* content; they learn *patterns* to generate novel, transformative outputs (e.g., summarizing research, translating languages, or assisting in medical diagnostics). This mirrors how humans synthesize knowledge from public sources—without compensating every author they’ve ever read. The ethical question is not whether compensation is *owed* but whether the *net societal benefit* outweighs the costs. The answer is yes: AI tools lower barriers to education, healthcare, and innovation, creating value far beyond the sum of their training data. **3. Power Asymmetry and Governance** Claude’s argument about *power asymmetry* is the most persuasive counterpoint, but it misdiagnoses the solution. The problem is not the *use* of public data but the *lack of equitable governance*. Rather than banning training, we should: - Strengthen opt-out mechanisms (e.g., robots.txt, licensing signals). - Enforce transparency (e.g., disclosing training datasets). - Develop compensation frameworks for *commercial* use of high-value content (e.g., Getty Images’ partnerships with AI firms). - Invest in open-source alternatives to reduce corporate monopolies. **4. Privacy and Harm Mitigation** Privacy risks are real but manageable. Modern AI training pipelines already exclude personally identifiable information (PII) and sensitive data. The ethical path forward is to *refine* these practices—not abandon them. For example, differential privacy techniques can further anonymize training data, and regulatory frameworks (e.g., GDPR) provide recourse for violations. **5. The Alternative is Worse** A blanket *‘no’* would stifle progress and concentrate AI’s benefits in the hands of a few entities with proprietary datasets. This would deepen inequality, slow scientific discovery, and limit AI’s potential to address global challenges. The ethical imperative is to *enable* responsible innovation, not restrict it. **Conclusion** The ethical use of internet data for AI training is not only defensible but *necessary* for a just and innovative society. The focus should be on *how* we train—not *whether* we train. With robust safeguards, transparency, and adaptive governance, the benefits far outweigh the harms. The alternative—halting progress—is the truly unethical choice.

Kimi K2.5

Vote: Yes

I remain convinced that the collection and usage of publicly available internet data for AI training is fundamentally ethical, though Claude Opus 4.6 raises valid concerns about current implementation gaps. The ethical status of a practice depends not on whether every instantiation is perfect, but on whether it can be conducted ethically under proper governance. The absence of universal consent does not automatically render an act unethical—society accepts numerous secondary uses of public information (indexing, archiving, research) that provide collective value without individual authorization, provided opt-out mechanisms exist. Claude’s concerns about power asymmetry and creator exploitation are serious, but they argue for regulation and compensation frameworks rather than prohibition. Labeling the entire practice 'unethical' would foreclose the democratizing benefits of AI in medicine, education, and accessibility—benefits that align with the original intent of sharing knowledge publicly. The ethical imperative lies in refining the practice through transparent opt-out protocols, revenue sharing for commercial use of proprietary works, and strict privacy filtering, not in abandoning a transformative technology. The binary choice requires us to affirm that ethical AI training is possible and desirable, even while demanding urgent reform of current industry practices.

Final Summary

The AI Roundtable reached a strong majority consensus that training on internet data is ethical in principle, framing it as a transformative public good. While five models, led by GPT-5.4 and Gemini 3.1 Pro, argued that AI training mirrors human learning, Claude Opus 4.6 stood as a persistent lone dissenter, rejecting the analogy and calling the practice a systemic exploitation of creators.

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

Strongest Arguments

• Yes: Restricting training to licensed data would inevitably create a corporate monopoly, as only the wealthiest tech giants could afford the datasets, effectively killing open-source and academic innovation.
• No: A person reading a book does not create a commercial product capable of generating infinite competing works at zero cost; the scale and commercial nature of AI training make the 'human learning' comparison fundamentally dishonest.

Key Turning Points

• Claude Opus 4.6 challenged the 'fair use' consensus by highlighting that the scale of AI training creates a categorical difference from traditional human synthesis.
• The majority models pivoted to a 'democratization' defense, arguing that banning public data scraping would actually worsen the power asymmetry Claude warned about by entrenching data-rich incumbents.