Generative AI can accelerate research and drafting, but speed becomes expensive when a polished sentence hides a false citation, invented quotation or missing qualification. A recent federal court episode offers a useful warning far beyond the legal profession: the person publishing the work remains responsible for every claim.
U.S. District Judge Arun Subramanian declined to sanction lawyers after AI-assisted work introduced citation errors into a filing, but called the episode a wake-up call. The firm said a lawyer used Harvey’s AI tools contrary to its policies and failed to check the generated citations. The judge also warned that excessive reliance on AI could weaken the training of junior professionals.
Artists, writers, curators and independent publishers face the same underlying problem. A tool can produce fluent text without showing where certainty ends. The answer is not to reject useful tools, but to build verification into the editorial process.
1. Separate discovery from evidence
Use an AI response to identify possible leads, vocabulary or questions. Do not treat it as proof. Every name, date, price, quotation, legal claim and attribution should be tied to a source you can open and inspect. If the source cannot be found, the claim does not move forward.
2. Trace important claims to primary material
For an exhibition, start with the museum or artist’s official page. For a law or policy, read the actual order, statute or agency release. For a scientific claim, locate the paper or institutional report. Independent reporting is essential for context and scrutiny, but primary documentation anchors the factual core.
3. Check quotations word for word
Never publish a quotation reconstructed from memory or generated from a summary. Confirm the exact language, speaker and context in the original transcript, recording or document. If you only have a paraphrase, label it as a paraphrase.
4. Build a claim ledger
Before publication, list the article’s central factual claims beside their sources. This can be a simple table in a working document. The practice makes missing support visible and creates a durable research trail for later corrections, updates and rights questions.
5. Verify links, images and licenses separately
A correct article can still fail at the image stage. A page displaying a photograph is not evidence that it may be republished. Record the creator, source page, license, required credit and any restrictions. Open every link again before publication; confirm that it supports the sentence attached to it.
6. Keep a human final pass
NIST’s Generative AI Profile recommends reviewing and verifying sources and citations in generated outputs. The final reviewer should read the piece without relying on the AI conversation that produced it. That distance makes it easier to notice unsupported leaps, repeated errors and prose that sounds authoritative without being precise.
The most valuable habit is simple: slow down at the point where information becomes public. AI can help create a draft, but accountability cannot be delegated. A reliable publishing system makes that responsibility visible at every step.