"The Hypocrisy at the Heart of the AI Industry" is an investigative feature by Alex Reisner, published March 20, 2026 in The Atlantic's AI Watchdog series. It argues that AI companies operate an intellectual-property double standard: they defend their own patents, terms of service, and anti-piracy protections while claiming fair use over the copyrighted books, videos, music, and other media used to train their models. The piece anchors that argument in two documentary artifacts — a leaked 2024 Stanford talk by Eric Schmidt and an unsealed 2021 internal Anthropic memo by Dario Amodei.
Author: Alex Reisner Publication: The Atlantic (AI Watchdog series) Published: March 20, 2026 Source: theatlantic.com
Summary of argument
Reisner's central claim is that AI companies apply opposite rules to their own intellectual property and to the material they train on. They protect their patents, their product code, and their model outputs aggressively, while characterizing their ingestion of copyrighted books, videos, and music as fair use. Reisner reports two primary-record artifacts as the load-bearing evidence for this framing.
The first is a talk Eric Schmidt gave at Stanford in April 2024, in which Schmidt counsels students to download whatever copyrighted content they need and then "hire a whole bunch of lawyers to go clean the mess up. If nobody uses your product, then it doesn't matter that you stole all the content." Stanford posted the video and removed it the next day.
The second is a 2021 internal Anthropic memo by Dario Amodei titled "An Economic Model for Compensating Data Producers," unsealed in copyright litigation against Anthropic. In it, Amodei acknowledges that AI could be "an increasingly extractive concentrator of wealth," that creators might "grumble" or "get mad," and that compensating creators "with a fraction of the profits from the model produced" could be a "great fit" for Anthropic's "public benefit orientation." Reisner contrasts this with Anthropic's 2026 court position that using copyrighted books is fair use, which he reads as leaving authors entitled to nothing.
The IP double standard
The feature lays out the asymmetry as a table of what companies treat as their own protected property versus the third-party material they characterize as free to use.
| What's their IP | How they defend it | What's other people's IP | How they defend taking it |
|---|---|---|---|
| Photoshop, search algorithms | Patents, IP litigation (Apple v. Samsung $1B+; Waymo v. Uber settled $245M) | Copyrighted books, videos, music | Claim fair use; OpenAI: "Without fair use access, the race for AI is effectively over. America loses." |
| Software piracy of their products | Adobe / Microsoft / Google moved to license-verification/cloud-only distribution | Authors' books / artists' works | "Quintessential fair use" (Meta court argument) |
| OpenAI / Anthropic / Google / xAI ChatGPT outputs | Terms of service explicitly forbid using output to develop competing models | Their training data | "Publicly available information" (OpenAI); "Used books, but not in any commercial products" (Anthropic — pre-litigation framing) |
Reisner notes that Meta's models are described as "open," but that Meta has reportedly sent takedown notices demanding deletion of copies from online platforms.
Key claims
Memorization as a refutation of the fair-use defense
Reisner argues that the industry's transformativeness defense — that models produce original work not derived from their training sources — does not hold against his own reporting:
"Companies argue that AI training is fair use because their AI models produce original work that is not derived from the sources they use for training. This is not necessarily true: My reporting has shown that chatbots and image generators can produce near-exact copies of media they were trained on, spitting out near-complete copies of Harry Potter and the Sorcerer's Stone, for example, or rendering images that are fuzzy copies of existing artwork."
In earlier Atlantic reporting (January 2026), Reisner had documented that frontier models can reproduce near-verbatim training data. The Hypocrisy feature surfaces that evidence specifically as a rebuttal to the industry's "training data is transformative" defense.
The "AI race" framing
Reisner describes how AI companies invoke geopolitical competition to short-circuit the IP debate, quoting OpenAI's response to the OSTP request for information on the AI Action Plan (early 2025):
"Without fair use access, the race for AI is effectively over. America loses."
Reisner argues this national-security framing should be received skeptically. It is presented as the hardest version of the argument tracked at National Security as Policy Trump Card.
Ed Newton-Rex as counter-voice
The feature profiles Ed Newton-Rex, former Stability AI VP of Audio, who quit in November 2023, as the dissenting industry voice. Newton-Rex argues that fair use "wasn't designed with generative AI in mind" and that current training practices cannot be "acceptable in a society that has set up the economics of the creative arts such that creators rely on copyright." His Fairly Trained nonprofit certifies AI models trained on properly acquired data, which Reisner presents as the alternative to the industry default.
Reception and relationships to other pages
Reisner's evidentiary contributions — the Schmidt video and the unsealed Amodei memo — feature in copyright-litigation argumentation and in policy debates over AI fair-use exemptions.
The Schmidt and Amodei artifacts strengthen the plaintiffs' bad-faith framing in NYT v. Microsoft, OpenAI et al.. The 2021 Amodei memo documents a gap between Anthropic's stated public-benefit orientation and its 2026 litigation posture, a tension also reflected in the Anthropic Claude Gov + Pentagon Dispute (2025–2026) dispute and in the safety-lab-versus-business critique discussed at Anthropic and Alignment — Ben Thompson (Stratechery, March 2026) and against the A Framework for AI Development Transparency (Anthropic). The memo's acknowledgement of AI as "an increasingly extractive concentrator of wealth" is now part of the public record on Amodei's documented thinking, and the IP-double-standard framing is the journalistic synthesis pulled into AI Copyright Litigation — Analysis.
Provenance
Atlantic feature, treated as Reisner's argument and documentary find. The Schmidt video and Amodei memo are both primary-record artifacts; their exact wording should be verified against the underlying primary sources before citing (the Schmidt video was posted and removed by Stanford; the Amodei memo was unsealed in Bartz v. Anthropic).
Confidence is medium: a single source, an opinionated feature, drawing on two primary artifacts that warrant independent verification.
Relationships
- supports: NYT v. Microsoft, OpenAI et al. (evidentiary contribution), Ed Newton-Rex (counter-voice profile), Fairly Trained (alternative model)
- contradicts: OpenAI / Anthropic / Google / xAI's "training is fair use" framing; OpenAI's "without fair use, America loses" framing
- related: Dario Amodei (internal memo evidence), Eric Schmidt (Stanford talk evidence), A Framework for AI Development Transparency (Anthropic), Anthropic and Alignment — Ben Thompson (Stratechery, March 2026), AI Copyright Litigation — Analysis, National Security as Policy Trump Card (planned)
- instance-of: AI Industry Discourse Analysis