External articles with editorial commentary. 17 links curated so far.
China's patent authority issues a formal warning against AI agent tools for patent drafting, naming OpenClaw specifically. The concerns — hallucination defects, information leakage, bad faith applications — are the same ones American lawyers raise about AI in practice. The difference is that a government is now naming specific tools as threats to practice integrity. The U.S. has not done this.
A family law practice guide built on Heppner. The headline says ChatGPT but the ruling involved Claude — the brand name has become generic shorthand, which is itself a small act of misrecognition. The substantive warning is sound: every AI chat about a legal issue that happens outside the attorney-client relationship is a potential exhibit. The authors predict AI-related discovery requests are coming. They are already here — Mobley v. Workday class notice went out in February.
The EEOC removed all AI employment guidance from its website in January 2025. Four states have since written their own laws, each with a different liability standard — from California's disparate impact framework to Texas's intent-only requirement. When the federal government withdraws standards, the gap fills with divergent state regimes and vendor self-assessment. That vacuum is exactly where confident misstatement thrives.
ProPublica's analysis of two decades of FOIA data: 23,000 criminal cases dropped in six months, including 5,000 drug cases, 1,300 terrorism cases, and 900 fraud investigations — all to resource 32,000 new immigration prosecutions. A 28-year DOJ veteran says the building blocks of successful fentanyl prosecutions were pulled out. A retired AUSA who describes himself as a Trump supporter worries about emboldened union corruption. This is the resource environment in which legal technology vendors are selling tools to public defenders. The offices that need AI most are the ones the system is defunding by other means.
An industry executive with 40 years in legal tech and no law license argues lawyers should become Iron Man by embracing Claude Code. The piece opens with two anonymous anecdotes — unnamed founder, unnamed company, unverifiable results — then attributes a "10X productivity" claim to Zack Shapiro that Shapiro did not make. The vocabulary is revealing: "transformational superpower," "superhero powers," "wear the suit." This is not analysis. It is the vendor hallucination phenomenon in published form — confident, fluent, structurally ungrounded. Read it as a specimen, not a source.
FOIA data: 63% of ICE arrests in Minnesota's enforcement surge were of people with no criminal record. The White House called all of them dangerous criminal illegal aliens. The characterization was confident, fluent, and contradicted by the government's own data — the same structure as a vendor telling you their product does something it does not. The data tells a different story than the press conference. It usually does.
Northwestern surveyed federal judges: 62% have used AI in their work. ChatGPT leads at 28.6% while Claude registers at 0.9%. Nearly half received no training. The irony: judges are overwhelmingly choosing the chatbot with the weakest privilege protections, and the landmark ruling that should inform those choices — Heppner — involved Claude. The system is adopting the technology faster than it is learning the rules.
CDT's deep policy analysis of AI-generated police reports — Axon's Draft One, Truleo, CopEntry, and others. The findings are damning: an AI hallucinated that an officer transformed into a frog (Disney movie in the background), vendors market limited audit trails as a feature to "reduce FOIA exposure," and it is often impossible to tell which parts of a finalized report were written by an officer versus generated by AI. For defense attorneys, this is the prosecution-side mirror of the hallucination problem. When an AI confabulates in a police report, it becomes evidence. You cross-examine the officer. The officer signed off on text a machine wrote. The machine heard a cartoon.
Judge Rakoff rules that 31 documents a securities fraud defendant generated using consumer Claude are neither privileged nor work product. The platform is not your lawyer, the ToS destroy confidentiality, and nobody directed the defendant to do it. The court left one door open: attorney-directed AI use might be treated differently. For public defenders, this is the clearest articulation yet of why local-first architecture is not a preference but an obligation. The privilege your client needs cannot survive a round trip through someone else's servers.
Furlong argues post-AI lawyers will be valued for judgment, not knowledge. Correct thesis, but it assumes a client who can afford that judgment. The keynote never names criminal defense.
EFF filing for transparency on a Medicare AI pilot. The privilege and transparency arguments parallel what defense attorneys face with evidence processing AI.
Academic treatment of AI in public defense. Maps the gap between what commercial legal tech builds and what indigent defense actually needs.
Addresses the core architecture question: can you run inference locally? Yes, and the cost is lower than most attorneys assume. In criminal defense, privilege makes this mandatory.
Official ABA preview. Nilay Patel keynote confirmed. The programming tilts toward firm-side adoption — the public defense track is notably absent.
Miami-Dade's public defender office adopting AI for case organization and research. A peer office in Florida working on the same problem from a different entry point.
Berkeley's catalog of AI tools in use in public defense, including body camera transcription. A reference point for what exists — and what doesn't.
The ABA's framework for AI confidentiality under Model Rule 1.6. This opinion makes local-first architecture not a preference but an obligation when the client's liberty is at stake.