The Bias Meter Defended the Klan.

On the first Tuesday of March 2025, a reader opened a Los Angeles Times opinion column about the Ku Klux Klan in Anaheim, clicked a button labeled Insights, and got back a machine-written counterpoint explaining that the Klan was really an expression of "white Protestant culture" responding to societal changes rather than an explicitly hate-driven movement.

The column was Gustavo Arellano's, published a week earlier, and it called the Klan a stain on a place that likes to celebrate the positive. The button had gone live the day before, part of a suite the paper's owner had been teasing since December. Alongside it ran a bias meter that scored the political lean of every opinion piece on the site. The whole package was sold as a corrective: a newspaper so committed to fairness that it would label its own slant and argue against itself in public, automatically.

It took about twenty-four hours to produce a defense of the Klan. The AI comments came down that afternoon. When Patrick Soon-Shiong, the executive chairman who'd championed the tool, sat for a CNN interview hours later, he admitted he'd read neither the column nor the response his product had written underneath it.

The tool had been live for a day, and the man who put it there was finding out what it said along with everybody else.

The queue that was never going to drain

Asked about supervision, Soon-Shiong said there was some level of human oversight, but that it was hard for the team to validate all responses in real time given the scale at which the tool was being applied. Readers were invited to report errors.

The paper built a machine that publishes under its name faster than its own editors can read, acknowledged that the humans couldn't keep up, shipped it anyway, and handed the audience the review queue. The Arellano counterpoint wasn't a freak output. A Scott Jennings op-ed got scored as centrist while running right-leaning talking points, and its auto-generated rebuttal left out that Trump had threatened to withhold wildfire aid from California. A Nieman Lab review found the tool citing weak sources, duplicating references, and attributing points to articles that didn't contain them.

None of that got caught either, for the same reason. Oversight that can't match the rate of output is a backlog with a nicer name.

Almost every approval workflow has this failure waiting in it. Somebody sizes the review step to the volume that exists on the day they design it, volume goes up, and the step quietly turns ceremonial. What's new is the size of the gap. You can staff for a ten percent increase. You can't staff for a system that writes an original argument on every story you publish, forever, at no marginal cost.

They automated the product

The worse problem sits underneath the launch. The thing Insights replaced was the only thing the institution had left to sell.

A newspaper sells discernment. Information is free and everywhere; the capacity to weigh nuance, consequence, and credibility in real time and decide what deserves the page is the scarce part, and it's what you were paying for. An INMA analysis of newsroom staffing put the economics underneath that plainly. Trust behaves like capital. It compounds slowly and erodes quickly.

The Times had been spending that capital hard for a year. A blocked Harris endorsement in October 2024 cost it more than seven thousand subscribers in a single month, close to two percent of its base. Editorial board members left. A hundred and fifteen newsroom jobs were cut. Carla Hall, the last member of the editorial board, took a buyout after thirty-two years.

So the sequence runs like this: dismantle the apparatus of institutional judgment, install a machine to perform judgment in its place, then discover the machine has opinions about the Klan. Insights worked as designed. A both-sides engine generates the other side, and nobody with standing was left in the room to say that some things don't get one.

The one place audiences won't take it

The evidence runs the opposite direction from what the feature assumed.

A conjoint experiment out of Chile put more than two thousand people in forced trade-offs between hypothetical news outlets with different AI policies. Committing to full human supervision made an outlet roughly fifteen percent more likely to be judged credible. Disclosing AI use added about another eleven. Whether the outlet used AI for transcription, translation, or personalization barely registered. People are asking one question about the technology, and it's whether anyone is still clearly responsible.

There's an exception buried in that data with the Times' name on it. When AI generates core journalistic content, especially interpretive material on politics and social issues, responses turn negative. Of all the jobs in a newsroom, audiences are most relaxed about the plumbing and least willing to hand over the judgment calls. Insights went straight for the judgment calls and put a meter on them.

The idea behind the button was never stupid. An opinion section that can show you the strongest version of the case against a column is a real service, and I'd read it. It stops being a service the moment nobody is deciding which cases deserve making.

The tempting fix is a disclosure line promising a human is in the loop, and I doubt it buys what people hope. Trusting News ran disclosure tests across ten newsrooms in three countries and found forty-two percent of readers less likely to trust a story after an AI disclosure against thirty percent more likely, with the penalty surviving even when the disclosure mentioned human oversight. The reaction to learning AI was involved tended to be stronger than any reassurance attached to it. A sentence claiming supervision is cheap, and readers price it accordingly. What the Chile study rewards is supervision that's structurally real, which is to say attached to a person who can be named and held to it.

So, where is the line?

Newsrooms got here first, and I'd bet most brands are one quarter away from the same discovery. Any company shipping a feature that generates text under its name has the same exposure, and most of them have the same scale mismatch hiding in the rollout deck.

The question to ask before launch is boring and unpopular: can one accountable person read everything this produces before a customer does? If the honest answer is no, you've hired a publisher, handed it your name, and left the editor's chair empty. Whatever it says next, you said.

Generating the other side of an argument is cheap now, and getting cheaper. Deciding which arguments deserve one is the job, and it's the part anyone was ever paying you for.