For the Republic
← Command Center / 🎙 Episode / 2026-02-27 · ~13 minutes (estimated from ~1,950 word count)

The Department of War Wants an AI Without a Conscience

Draft Complete — Pending Host Review

Fact Check

7/10

Fact Check Report

Summary

The draft script is largely factually solid. The core narrative -- that the Pentagon is simultaneously threatening Anthropic with a supply chain risk designation and DPA invocation over two AI safety guardrails -- is well-supported by all source material and independent reporting. Most statistics, quotes, and characterizations check out. However, there is one clear misattribution, several claims that need tightening, and a few areas where framing slightly distorts the underlying facts.

  • Red flags: 1
  • Yellow flags: 5
  • Blue flags: 3

Findings

Red Flags

"As legal experts at Lawfare noted, this application is 'without precedent under the history of the DPA.'"

  • Location in script: DPA section, paragraph 2
  • Issue: This is a misattribution. The phrase "without precedent under the history of the DPA" does not come from Lawfare. It comes from Joel Dodge, an attorney and director of industrial policy and economic security at the Vanderbilt Policy Accelerator, as quoted in an Associated Press story syndicated widely on February 26, 2026. The Lawfare article on this topic was written by Alan Z. Rozenshtein and makes different (though related) arguments -- it discusses the DPA's allocation authority as "largely untested" since the Korean War and raises the Major Questions Doctrine -- but does not use this specific phrase. Attributing Dodge's quote to Lawfare is factually incorrect.
  • Evidence: AP reporting via Federal News Network, Defense News, Washington Times, and dozens of syndicated outlets all attribute this quote to "experts like Dodge." The Lawfare article (https://www.lawfaremedia.org/article/what-the-defense-production-act-can-and-can't-do-to-anthropic) was independently fetched and confirmed not to contain this phrase.
  • Recommended fix: Either (a) attribute correctly: "As one legal expert at the Vanderbilt Policy Accelerator put it, this application is 'without precedent under the history of the DPA'" or (b) if citing Lawfare is preferred for brand recognition, use an actual quote from the Lawfare piece, such as Rozenshtein's characterization of the DPA's compelled-contracting authority as "largely untested" since the Korean War. Option (a) is cleaner.

Yellow Flags

"Anthropic -- the AI company, the one that builds Claude -- has been working with the Pentagon for over a year."

  • Location in script: Opening context section, paragraph 3
  • Issue: The timeline is plausible but slightly imprecise. Anthropic's classified partnership with the Pentagon via Palantir dates to late 2024 -- roughly 14-15 months before the episode date. The formal $200 million contract was signed in July 2025, which is about 8 months. "Over a year" is defensible if measured from the Palantir partnership, but could be misleading if interpreted as referring to the formal DoD contract. Neither Amodei's statement nor the source material uses the phrase "over a year."
  • Context: Amodei's statement says Anthropic was "the first frontier AI company to deploy our models in the US government's classified networks" without specifying a duration. The supplemental research pins the Palantir partnership to late 2024.
  • Recommended fix: Consider "since late 2024" or "for more than a year" rather than "for over a year," to avoid ambiguity about which relationship is being measured. Alternatively, keep "over a year" but ensure the host is aware it dates to the Palantir partnership, not the formal July 2025 contract.

"Back in December, they agreed to let Claude be used for missile defense and cyber defense applications."

  • Location in script: Opening context section, paragraph 3
  • Issue: This is accurate per NBC News reporting and the supplemental research, but the framing slightly oversimplifies. Anthropic agreed in December 2025 contract negotiations to allow Claude for missile and cyber defense purposes. However, an Anthropic spokesperson told NBC News that "every iteration of our proposed contract language would enable our models to support missile defense and similar uses" -- implying this was not a new concession made in December but something that had always been in their proposed language. The script frames it as a concession ("they agreed to let") when Anthropic frames it as having always been their position.
  • Context: NBC News, December 2025 reporting; Anthropic spokesperson statement.
  • Recommended fix: Minor adjustment: "In December, Anthropic confirmed its willingness to let Claude be used for missile defense and cyber defense applications" or similar language that doesn't imply a new concession if that framing is contested.

"They were the first frontier AI company deployed in classified environments."

  • Location in script: Opening context section, paragraph 3
  • Issue: This claim comes directly from Amodei's own statement. It is Anthropic's claim about itself. While it has been widely reported and not publicly disputed by other companies, it is worth noting that this is self-reported rather than independently verified. No government source in the available reporting independently confirms this "first" claim. It is likely accurate -- Claude was reportedly the only AI model on classified networks until xAI's recent deal -- but the script presents it as established fact rather than as Anthropic's assertion.
  • Context: Amodei's statement (source 01): "We were the first frontier AI company to deploy our models in the US government's classified networks." CNN reporting confirms Claude is "the first AI system to be used in the military's classified network."
  • Recommended fix: This is minor. Consider "They say they were the first" or "By all accounts, they were the first" to hedge slightly. Or keep as-is if the host is comfortable treating Anthropic's claim (which CNN also reports) as established.

"And a senior Pentagon official -- Emil Michael, the Undersecretary for Research and Engineering -- going on social media to call Dario Amodei 'a liar' with 'a God complex.'"

  • Location in script: Context section, paragraph 4
  • Issue: The quote is accurate and well-sourced (CNN, supplemental research). However, Michael's full post on X reads: "It's a shame that @DarioAmodei is a liar and has a God-complex. He wants nothing more than to try to personally control the US Military and is ok putting our nation's safety at risk. The @DeptofWar will ALWAYS adhere to the law but not bend to whims of any one for-profit tech company." The script excerpts "a liar" and "a God complex" which is fair. The potential issue is calling this a characterization by "a senior Pentagon official" in a way that could imply this was an official DoD position statement rather than a personal social media post. Michael is indeed a senior official and posted on X in what appears to be an official capacity, so this is a minor framing point.
  • Recommended fix: The current framing is defensible. No change strictly required, but the host should be aware the post was on X (personal social media), not an official Pentagon press release.

"In a narrow set of cases, we believe AI can undermine, rather than defend, democratic values. Some uses are also simply outside the bounds of what today's technology can safely and reliably do. Two such use cases have never been included in our contracts with the Department of War."

  • Location in script: Paraphrased throughout, not presented as a direct quote
  • Issue: The script says Anthropic "drew two red lines: no mass surveillance of American citizens, and no fully autonomous weapons without adequate reliability testing." The second red line is slightly simplified. Amodei's actual position on autonomous weapons is two-pronged: (1) current AI is not reliable enough ("frontier AI systems are simply not reliable enough to power fully autonomous weapons"), AND (2) without proper oversight, fully autonomous weapons lack the judgment of trained troops. The "without adequate reliability testing" framing captures part of this but omits the oversight/judgment dimension. This matters because it makes Anthropic's position sound more conditional (just needs testing) than it actually is (also needs governance frameworks that don't exist).
  • Context: Amodei's statement (source 01) is the primary reference.
  • Recommended fix: Consider "no fully autonomous weapons without adequate reliability and oversight safeguards" to capture both prongs of Anthropic's argument.

Verification Needed

"What it has never been used for -- in its entire 75-year history -- is to force a software company to remove ethical restrictions from its product."

  • Location in script: DPA section
  • Note: The "never been used" claim is supported by expert Joel Dodge's statement that the DPA "has never been used to compel a company to produce a product that it's deemed unsafe, or to dictate its terms of service." The Lawfare article similarly describes the compelled-contracting authority as "largely untested" since the Korean War. The claim is almost certainly correct -- no counterexample was found in any search. However, proving a negative across 75 years of DPA invocations is inherently difficult. The host should be comfortable asserting this based on expert consensus rather than exhaustive historical review. The "75-year" figure is accurate (signed September 8, 1950; 75 years and 5 months as of February 2026).

"Congress has written zero statutes governing autonomous weapons. Zero statutes governing AI-enabled domestic surveillance."

  • Location in script: Deeper problem section
  • Note: This is substantially correct. No standalone federal statute specifically regulates lethal autonomous weapon systems or AI-enabled domestic surveillance. The primary governance framework for autonomous weapons is DoD Directive 3000.09, which is executive branch policy, not legislation. However, Congress has addressed autonomous weapons tangentially through NDAA provisions -- Section 251 of FY2024 NDAA requires notification of changes to DODD 3000.09, and Section 1066 of FY2025 NDAA requires annual reporting on LAWS deployment. The House also passed the Fourth Amendment Is Not For Sale Act (which would restrict government purchase of Americans' data). Saying "zero statutes" is accurate in the strictest sense (no standalone statute governing these areas has been enacted), but the host should know Congress has nibbled at the edges through NDAA provisions and passed the Fourth Amendment bill through the House. Whether these count as "governing" these areas is debatable.

"Anthropic loosened its own internal safety commitments this same week. They dropped their Responsible Scaling Policy pledge, the commitment to pause training if safety measures proved inadequate."

  • Location in script: Counterargument section
  • Note: The timing and substance are confirmed. On February 24, 2026 (Monday of the same week), Anthropic released RSP v3.0, which removed the hard commitment to pause model training if safety measures were not proven adequate in advance. TIME, CNN, Engadget, and Semafor all confirm this. However, Anthropic did not "drop" the RSP entirely -- they overhauled it to version 3.0, replacing binding commitments with publicly announced targets and separating unilateral commitments from industry-wide recommendations. The word "dropped" suggests elimination rather than weakening. "Weakened" or "overhauled" would be more precise. An Anthropic spokesperson told the Wall Street Journal the change was unrelated to Pentagon negotiations. The host should be aware of that company claim even if it strains credulity given the timing.

Sources Consulted

Primary Source Material

Independent Verification Sources

Clean Claims

The following major factual claims in the script checked out and can be relied upon:

  1. The Pentagon is simultaneously threatening a supply chain risk designation and DPA invocation. Confirmed across all sources. Amodei himself flags the contradiction in his statement.

  2. The supply chain risk label is "reserved for US adversaries, never before applied to an American company." This is Amodei's characterization in his statement and is repeated across CNN, WaPo, and other reporting without challenge.

  3. The deadline is 5:01 PM Friday. Confirmed by Pentagon spokesman Sean Parnell on X.

  4. Emil Michael called Amodei "a liar" with "a God complex" on X. Directly confirmed by CNN (source 03) with the full text of the post.

  5. Anthropic's two red lines are mass domestic surveillance and fully autonomous weapons. Confirmed directly from Amodei's statement and all reporting.

  6. The Pentagon wants "all lawful use" language. Confirmed by CNN, WaPo, Axios, and Amodei's statement.

  7. The DPA was signed in 1950 during the Korean War. Confirmed. Signed September 8, 1950.

  8. The DPA has been used for medical supplies during COVID. Extensively confirmed. Used over 100 times for COVID medical supply needs.

  9. Anthropic forfeited hundreds of millions in revenue by cutting off CCP-linked firms. Confirmed. Anthropic executives told the Financial Times the impact was in the "low hundreds of millions."

  10. The Pentagon has been shopping for replacements -- OpenAI, Google, xAI. Confirmed by supplemental research and Axios reporting. xAI's Grok recently signed a classified-settings contract.

  11. Grok is "not viewed as being as advanced as Claude." Confirmed by CNN reporting: a Pentagon official confirmed Grok is "not viewed as being as advanced as Claude."

  12. DoD Directive 3000.09 requires "appropriate levels of human judgment" and is a policy, not a law. Confirmed. It is a DoD directive, not a federal statute, and can be changed by the Secretary of Defense.

  13. The IC has acknowledged that purchasing Americans' data raises constitutional/privacy concerns. Confirmed by the declassified ODNI report (January 2022, released June 2023) and extensive EPIC/ACLU/Senate reporting.

  14. The $380 billion valuation and $30 billion funding round. Confirmed by Bloomberg, CNBC, TechCrunch, and others. Closed February 12, 2026.

  15. Anthropic's RSP was weakened the same week. Confirmed. RSP v3.0 released February 24, 2026, removing the hard commitment to pause training.

  16. The $200 million contract figure. Confirmed by CNN and multiple outlets.