Microsoft exec called AI scraping 'the largest theft of labor in human history'

Is AI training “theft” or just copyright infringement?

• Many argue scraping for training is “the largest theft of labor”: models extract value from millions of works without consent or pay, then compete with the creators.
• Others insist it’s not “theft” in the strict sense: originals remain, copying is copyright infringement at most, and law traditionally protects copies, not “ideas/facts/patterns.”
• A recurring counterpoint: human learning from books is legal; opponents respond that a non‑rival human mind is not comparable to a massively scalable machine learner used commercially.
• Legal status is described as unsettled: some courts have found aspects of training to be fair use; fair use vs market harm remains central and contested.

Scale, automation, and double standards

• Commenters stress scale changes everything: one person pirating vs corporations hoovering the web, paywalled and even pirated datasets, then monetizing outputs.
• Many note hypocrisy: individuals have been heavily punished for piracy; AI labs doing similar acts at massive scale are rewarded with trillion‑dollar valuations.
• Labs’ complaints about model “distillation theft” are viewed as especially hypocritical, since their own models are trained on unlicensed material.

Impact on creators, culture, and labor

• Strong concern that AI undercuts the market for writers, artists, coders, researchers, musicians, etc., by selling cheap substitutes built from their work.
• Some open‑source and blog authors say they regret publishing publicly; they now see their unpaid work powering tools that may end their careers.
• Others argue aggregation and remixing is how culture has always progressed; copyright maximalism itself is framed as a historical aberration driven by large media firms.
• Fear that original web content will disappear as sites die (costs, bot load, loss of traffic), with knowledge effectively “trapped” in proprietary models and buried under AI-generated “slop.”

Commons vs property and redistribution

• One camp welcomes AI as the ultimate “information wants to be free” moment and a partial unwinding of overextended IP, especially for works that arguably should be in the public domain.
• Another camp insists IP is a legitimate right; if training on copyrighted work is allowed, then models and weights should also be treated as part of the commons.
• Proposed remedies include: heavy taxes or profit‑sharing on AI firms, creator royalty schemes tied to training data, or universal benefits funded by AI gains.

Regulation and structural proposals

• Ideas range from:
  • Require consent and licensing for training, with opt‑outs that actually work.
  • Mandate disclosure of training data and release of older model weights.
  • Declare model outputs uncopyrightable and potentially subject to original authors’ claims.
  • Cap company size or compute, or even ban certain proprietary frontier models.
• Some note enforcement will be politically hard, especially given geopolitical competition and corporate lobbying.

Open vs closed models and distillation

• Open‑weight models are cited as partial democratization, but critics say open weights trained on unlicensed data are still “laundered theft.”
• Several defend distillation as a “human right” to work with models trained on humanity’s corpus; others worry it further entrenches the initial uncompensated appropriation.

Broader societal concerns

• Threads explore:
  • Concentration of wealth and power if AI automates both mental and physical labor.
  • Environmental and infrastructure costs of huge data centers.
  • Analogies to surveillance: “public” data use becomes problematic when automated and centralized at extreme scale.