The reporting frames the unredacted quote as directly contradicting Microsoft's public fair-use defense, giving plaintiffs' counsel ammunition to argue willfulness. That legal distinction matters enormously — it multiplies statutory damages from $30,000 to $150,000 per work across an entire training corpus.
Argues the more consequential story isn't this one quote but what it signals about future litigation: every internal Slack, PR comment, and design doc at every AI lab is now a subpoena target. The gap between what companies say publicly about fair use and what engineers write internally has become a strategic liability.
By surfacing this story to the HN front page with 793 points, the submitter highlights the gap between Microsoft's consistent public messaging — that training on copyrighted text is transformative fair use — and the internal characterization of the same practice as theft. The framing suggests the public position was a legal narrative, not a genuine belief.
Newly unredacted court filings, first reported by TechCrunch on September 17, 2026, reveal that a senior Microsoft executive internally described the practice of scraping the open web to train large language models as "the largest theft of labor in human history." The quote surfaced in ongoing copyright litigation brought by publishers and authors against Microsoft and OpenAI, and it directly contradicts the position both companies have taken in public filings and press statements — that training on copyrighted text is transformative fair use.
The filing does not name a single smoking-gun product decision tied to the remark, but it does place the comment inside internal deliberations about data sourcing for models that eventually shipped as Copilot and GPT-family systems. Plaintiffs' counsel is using the quote to argue willfulness — the legal standard that separates a garden-variety infringement claim (statutory damages capped at $30,000 per work) from willful infringement (up to $150,000 per work). Multiply that by a training corpus and the math gets uncomfortable fast.
Microsoft's public line has been consistent: training is fair use, outputs are transformative, and the AI industry cannot exist without ingesting the public web at scale. The internal line, apparently, was different.
The legal exposure is the obvious story. The more interesting story is what this does to every AI company's discovery posture going forward. Every Slack message, every PR review comment, every design doc where an engineer typed something like "yeah we're basically laundering the training set" is now a subpoena target. That's not hypothetical — it's exactly how this quote got out. Someone wrote it down in a place lawyers can reach.
Compare the tone here to the industry's public messaging over the last three years. OpenAI's public comments to the UK House of Lords in January 2024 argued that "it would be impossible to train today's leading AI models without using copyrighted materials." Anthropic has taken a similar line in its own filings. The unstated premise — the thing nobody says out loud — is that the economics of frontier models assume the training corpus is free. If courts land on "willful infringement" and settlements run into the billions, that assumption breaks, and the moat around the labs that already trained on the pre-2023 open web gets a lot deeper. Newcomers would have to license.
There's also a governance angle that Microsoft in particular can't dodge. Microsoft has spent the last two years marketing "Responsible AI" as a differentiator — RAI standards, model cards, red-team disclosures, the works. A single internal quote calling the underlying data acquisition theft doesn't nullify that program, but it does hand every enterprise buyer's procurement team a very awkward question for the next renewal conversation: *what changed between when your exec said this and when you sold us Copilot?*
The community reaction has split along predictable lines. On Hacker News (793 points at time of writing), the top-voted comments are roughly "this was obvious to anyone paying attention" and "the quiet part got said out loud." Rightsholder advocates are treating it as vindication of everything they've argued since the New York Times filed against OpenAI in December 2023. AI-industry defenders are pointing out — correctly — that one executive's rhetorical flourish in a Slack channel is not a legal admission by the corporation. Both things can be true.
If you ship anything that touches a foundation model, three things get more expensive and more annoying in the next 12 months.
First, indemnification clauses in your model-provider contracts just became load-bearing. Microsoft, OpenAI, Anthropic, Google, and AWS Bedrock all offer some flavor of "we'll defend you if a customer gets sued over model outputs." Read the exclusions. Most of these indemnities carve out cases where the customer fine-tuned the model, prompted it in ways that elicited copyrighted content, or used a non-current model version. If you're on a pinned model version for reproducibility reasons — as most serious production stacks are — check whether your indemnity still applies. Several of these clauses quietly expire when the vendor deprecates the underlying weights.
Second, your own internal comms about training data, fine-tuning data, and RAG corpora are now discovery-relevant if you ever end up in a rights dispute. This is not a reason to stop writing candid design docs — self-censoring engineering discussion is a worse failure mode than legal exposure. It *is* a reason to make sure your data-sourcing decisions have a real paper trail: where the corpus came from, what license it was under, who approved the ingestion. "We scraped it because someone had done it in a notebook" is the answer that gets you sued.
Third, if you're building on open-weight models — Llama, Mistral, Qwen, DeepSeek — the derivative-liability question is genuinely unsettled. Meta's Llama license disclaims warranties on training data, and courts have not yet ruled on whether a downstream deployer inherits infringement liability from the base model's training set. The safe read for now: treat the training-data provenance of any base model as a due-diligence item, not a checkbox. Ask for it in your vendor questionnaire. If the answer is vague, price the risk in.
The quote itself won't decide any of these cases — a single internal comment rarely does. But it's the kind of exhibit that shapes settlement math, and settlement math is where most of this actually resolves. Expect a wave of amended complaints citing this filing within the next quarter, expect model providers to quietly tighten their internal comms guidance, and expect the "training data provenance" line item on enterprise AI RFPs to get a lot more detailed. The era where "we trained on a large corpus of publicly available text" was a sufficient answer is closing.
I just don't understand people saying "but a human learning from a book isn't illegal".How do people not understand that some laws only make sense at a certain scale? One human learning from resources and being added to the labour pool is not the same as an infinitely copyable en
“Information wants to be free“.It’s not “theft of labor”; the work was already done. If anything it is theft of “intellectual property” (aka “copyright infringement”), if you believe that is a thing, but not of the “labor” that went into it.My personal take: anyone producing content, everyone’s crea
If someone asked what is 'the largest theft of labor in human history' I would have thought slavery.
I wouldn't have a problem with working off the fruits of other people's labor because most of us are essentially doing that everyday anyway, the issue is that big tech companies (want to) reap all the benefit and create profit from something that should be accessible to everyone. Everythin
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
This will lead to a massive settlement between the big boys and most people who put stuff out in good faith will be left out of it. And that will be the end of it. We will never hear anything about this ever again and the ‘theft’ will continue like normal.