In early July 2026, the University of Hawaii Economic Research Organization (UHERO) temporarily withdrew a widely cited energy policy report after its lead author acknowledged that a key data point had come from an AI assistant he had not checked. The episode is a concrete, recent illustration of a research-integrity failure mode institutions are increasingly having to formalize policy around: unverified generative-AI output entering a publication as if it were sourced fact.
What the report claimed
The report, “Hawai’i’s Electricity Future: Three Findings on Solar Reform, Enhanced Geothermal, and the JERA LNG Proposal,” was authored by UH economist and UHERO fellow Michael Roberts together with doctoral student Ethan Hartley. Released in late June 2026, it argued that O’ahu did not need a proposed new gas-fired power plant tied to a liquefied natural gas (LNG) supply proposal from JERA, and that expanded solar generation was a more affordable path forward for the island’s electricity system.
The findings drew immediate pushback. Hawaiian Electric Company (HECO) disputed the report’s solar cost figures, and Hawaii Governor Josh Green publicly dismissed the analysis as “baloney.” The state’s Hawai’i State Energy Office (HSEO), led by chief energy officer Mark Glick, was separately reported as characterizing the paper as offering nothing serious to the debate over O’ahu’s energy future; HSEO’s own formal review concluded the report’s conclusions “may be true only within the model’s assumptions” and raised 56 unanswered questions about its methodology.
How the error reached the page
Under scrutiny, Roberts began auditing the report after HECO flagged a labeling issue with solar-plus-storage contract data. That audit surfaced a more serious problem: in describing mainland solar-plus-storage contract pricing, the report had relied on an AI assistant to summarize a Lawrence Berkeley National Laboratory dataset. The AI-generated summary asserted that contract prices were “flat to falling” between 2018 and 2024 — the opposite of what the underlying data showed, where prices had roughly doubled over that period. Roberts had also drawn on outdated 2018-era solar pricing (roughly 8–12 cents per kWh) rather than 2026 figures (roughly 21–23 cents per kWh), compounding the distortion.
Roberts was direct about the cause: he said he had leaned on the AI assistant and did not check what it produced, because “it looked right” to him. He also noted that the errors, taken together, had made the case against the LNG proposal look stronger than the corrected data actually support.
Withdrawal, not retraction — and why the distinction matters
UHERO’s own description of the action, and the reporting around it, used the language of a temporary withdrawal rather than a permanent retraction: Roberts said he intended to correct the identified errors, complete a fuller review, take in additional stakeholder feedback, and republish a revised edition. That distinction is not a technicality. In standard research-integrity practice, a retraction permanently removes a work’s standing as part of the record, typically with a formal retraction statement explaining why; a withdrawal or an expression of concern is a lesser, often reversible, signal that a work’s reliability is in question while it is under review. Readers and outlets covering this story should be careful not to conflate the two — see CASRAI’s comparison of retraction vs. correction for how publishers and institutions typically draw that line, and the related distinction between a retraction and a redaction.
UHERO also said it would update its publication policies to more clearly distinguish independently authored working papers and reports — like this one — from positions that represent the institute or university itself, a governance gap this episode exposed.
Why this matters beyond one report
This case did not involve fabricated data in the classic research-misconduct sense; it involved a researcher trusting an AI tool’s summary of a real, findable dataset without verifying it against the source. That is precisely the failure mode that generative-AI disclosure and verification policies from journals, funders, and institutions have been written to prevent over the past two years — requiring authors to disclose AI tool use and independently verify any AI-assisted output before it enters a publication, rather than treating a plausible-looking summary as equivalent to a checked citation. See CASRAI’s guide to journal and publisher policies on generative AI in manuscripts for how that policy landscape has developed, and the dictionary entries on hallucination and AI-generated content for the underlying vocabulary. More broadly, the episode is a live example of the accountability questions CASRAI’s research integrity and research misconduct vocabulary exists to help institutions reason about clearly — even when, as here, the underlying conduct falls short of misconduct and lands instead in the more common territory of an unverified shortcut with real policy consequences.
Sources
- Honolulu Star-Advertiser, “UHERO temporarily withdraws energy report, citing errors” (July 8, 2026)
- Hawai’i Public Radio, “UHERO fellow admits error in energy report came from AI” (July 9, 2026)
- Honolulu Civil Beat, “UH Expert Withdraws Paper Calling For More Solar Instead Of Power Plants” (July 2026)
- Aloha State Daily, “State energy office slams retracted UHERO solar energy report” (July 9, 2026)
- Hawai’i State Energy Office, “HSEO Issue Brief: Review of the Hartley/Roberts Report on Hawai’i’s Electricity Future”
- University of Hawaii News, “Faculty author temporarily withdraws UHERO report on Hawai’i’s electricity future”
