A commercial AI blog tool assigned to rewrite local news stories fabricated key details and invented context for a court ruling and a gasline special session this week, the latest installment of a recurring comparison that exposes the gap between machine-generated text and accurate journalism. The errors - including invented biographical claims and a generic legislative debate template - matter for writers, developers, and government professionals who increasingly encounter AI-produced summaries.
The feature, "AI Tries To Write The News," uses an unnamed AI tool to produce articles based on notable Juneau-area events. Editors then compare the results with human-written originals, highlighting passages that are outright fiction. This week's tests centered on a judge's decision in a U.S. Senate candidate's legal challenge and a leaked confidential draft shaking up the Alaska LNG project.
A fake 'prominent figure' and a misread legal dispute
The human-written story, published June 26, reported that Superior Court Judge Thomas Matthews ruled the Alaska Division of Elections illegally removed Petersburg resident Dan J. Sullivan from the U.S. Senate primary. The division argued his candidacy was intended to confuse voters because he shares a name with the incumbent, but the judge found the agency lacked authority to evaluate a candidate's motives. "The Division's decision to exclude Mr. Sullivan … was based upon a new, previously unstated, 'good-faith' criteria," Matthews wrote in his ruling.
The AI-generated version invented a background claim that Sullivan "had been a prominent figure in Alaska politics" - a characterization absent from the original. The tool also misstated the division's grounds for decertification, saying it involved "procedural issues related to paperwork and residency requirements." None of that appeared in the actual story. The fabricated details show how language models can produce confident-sounding assertions that collapse under scrutiny.
A leak becomes a generic legislative dispute
The second human-written article centered on the leak of a confidential draft agreement between the state-owned Alaska Gasline Development Corp. and Glenfarne Group, the primary developer of the Alaska LNG Project. The draft, first reported by the Alaska Beacon, states that if Glenfarne fails to proceed, the state may have to pay to retake control of the project. Lawmakers worried that proposed tax breaks could inflate a potential "clawback" price, with Sen. Bert Stedman warning of value in the "hundreds and hundreds of millions, if not billions of dollars."
The AI tool ignored the leak entirely. Instead, it wrote a story claiming the Legislature had reconvened for a second special session to debate LNG legislation, complete with invented bullet-point disputes over funding, environmental safeguards, and Indigenous concerns. The real special session was already under way, and the specific conflict - the leaked draft and its financial implications - never appeared. The machine-written version substituted a generic legislative debate template for the actual news.
Why this matters for government, technology, and writing professionals
For writers and editors, the experiment is a reminder that AI tools marketed for content generation can produce text that reads smoothly but contains serious factual errors. The errors highlight why training on AI literacy, including courses like AI for Writers, can help professionals avoid being misled by automated outputs. No AI blog tool should be used as a substitute for verified reporting without rigorous human review.
IT developers building applications that use large language models for news or policy briefs need to account for the models' tendency to invent plausible details when key information is missing. The AI's failure to distinguish between a leaked document and a standard legislative session suggests that even well-trained systems can default to pattern-matching rather than factual retrieval. Systems that deliver real-time information to government agencies or newsrooms must include fact-verification steps, not just text generation.
Government staff who may receive AI-summarized board reports or legal analyses should treat such summaries as initial drafts. A single fabricated word - like calling a litigant a "prominent figure" - can change public perception and legal risk. The Alaska examples show that the cost of relying on unchecked AI output could be measured in misinformed votes, policy debate, or millions of dollars in project miscommunication.
Your membership also unlocks:
