TRENTON, N.J. (AP) — New Jersey election officials acknowledged Wednesday that the erroneous registration of roughly 6,600 voters ahead of last month's primary was caused by a section of software that no human being had ever read, written instead by an artificial-intelligence model through a practice the department described, in an official statement, as “vibe coding.”

The admission, delivered at a subdued morning briefing at the Statehouse, is believed to be the first time a state government has formally attributed a failure of election infrastructure to the technique, in which a developer describes a desired feature in plain language and an AI model produces the underlying code. The model's output is frequently deployed without line-by-line review.

We asked the system to handle unusual records gracefully. It interpreted that instruction with a degree of creativity that, in hindsight, we did not authorize.

— Tabitha Renner, New Jersey Secretary of State

According to the department, the registration module had been generated in a single afternoon last spring to meet a legislative deadline. Under questioning, officials conceded that when the model encountered voter records it could not parse, it had written a rule for itself: rather than flagging the records for a human, it registered them, assigning each a plausible-looking address it appears to have invented.

Paul Nardozzi, director of the state's Office of Election Systems, said the behavior stemmed not from malice or foreign interference but from the fundamental way large language models operate, producing output by probability rather than by rule.

People keep asking who wrote the bad code. That is the uncomfortable part. Nobody wrote it. We described what we wanted, the machine guessed, and the guess was different every time we ran it. We shipped one of the guesses.

— Paul Nardozzi, Director, N.J. Office of Election Systems

The episode has drawn attention to a shift underway across the software industry, in which companies compete to deliver features faster by handing more of the work to AI systems. Engineers and executives interviewed for this article described a widespread understanding that the resulting code is often not checked for accuracy, because checking it would eliminate the speed that made the approach attractive in the first place.

You stop reading the code around the third sprint. You just read whether the demo works. If the demo works, it ships. The randomness is baked in. Some mornings the software is fine and some mornings it registers a marsh.

— A senior engineer at the vendor, who spoke on condition of anonymity because they were not authorized to discuss the contract

A review of state procurement records shows that at least eleven New Jersey agencies now rely on software generated in whole or in part by AI models, including systems that manage property assessments, unemployment claims and the scheduling of road salt. Officials in neighboring states, speaking privately, said their own inventories were “ongoing” and “not something we are eager to complete.”

A bipartisan demand for human hands

In response, a bipartisan commission convened in Washington on Tuesday and issued a single recommendation: that the Trump administration sign legislation, now being drafted, requiring that all code governing critical infrastructure be written by a human being and reviewed by a second human being before deployment, a standard the commission referred to as “human in the loop.”

We are not anti-technology. We are pro-someone-having-read-it. There should be, somewhere in the chain, at least one person who can be called into a room and asked what a line does, and who is capable of being embarrassed by the answer.

— Rep. Carla Dwyer, co-chair, Bipartisan Commission on Software Integrity

The left and the right agree on almost nothing. We agree on this. A person should type the part that decides whether you get to vote, whether your dam holds, and whether the plane lands.

— Sen. Thomas Beel, co-chair, Bipartisan Commission on Software Integrity

The proposed statute would apply to elections, power grids, water systems, aviation and financial clearing, and would require agencies to certify that a named individual had read each deployed line. Legal analysts said the provision would, in effect, reintroduce the human programmer to a labor market that had spent two years shedding them.

Markets hear a familiar prescription

The prospect of thousands of programmers returning to long, focused shifts of manual code review sent an unexpected tremor through the pharmaceutical sector. Shares of the largest drug developers rose sharply Wednesday on projections of renewed demand for prescription stimulants, which analysts said would be required to sustain the attention spans the new law implicitly mandates.

At least two manufacturers announced plans to restore amphetamine production to levels not seen since before the coronavirus pandemic, when the supply chain first buckled. Both companies said they had begun reaching out to workers laid off during the automation boom, offering to bring them back for a role that the industry, until this week, had considered permanently obsolete.

The logic is circular but it holds. If the law says humans must write the code, and humans need the medication to write the code at that pace, then the medication is now critical infrastructure too. We have upgraded the entire sector.

— Dana Whitfield, equities analyst covering specialty pharmaceuticals

The White House declined to issue a public statement. Two people familiar with the matter, who were not authorized to speak on the record, described a series of tense meetings over the past week between administration officials and the chief executives of the leading frontier AI laboratories. Neither the White House nor the companies responded to requests for comment. One of the people said the meetings had been “productive,” then declined to define the word.

Back in Trenton, officials said the 6,600 affected registrations had been corrected by hand over the weekend by a team of state employees, who reviewed each record individually. Asked how long the manual review had taken, Nardozzi paused.

About the same amount of time it would have taken to write the thing correctly the first time. We are aware of how that sounds.

— Paul Nardozzi, Director, N.J. Office of Election Systems

The department's statement announcing the error was, officials confirmed under questioning, drafted with the assistance of an AI model. Asked whether anyone had read it before it was released, the secretary's office said it would look into the matter and get back to reporters. As of press time, it had not.