Politico has become a testing ground for how AI clauses in union contracts could shape adoption in newsrooms across the U.S.
The PEN Guild represents over 250 workers at Politico and its sister publication, energy and environment site E&E News. Earlier this year, the Guild alleged management had violated AI adoption guidelines negotiated into its contract.
On July 11, the Guild and Politico held an arbitration hearing to determine whether the publication had broken its collective bargaining agreement through these alleged violations. Nieman Lab obtained access to the arbitration hearing transcript, including over 300 pages of testimony from Guild members and senior editors at Politico.
The allegations revolve around two generative AI-powered tools. One tool, called LETO, generates live summaries of speeches and was used on Politico.com’s homepage during the Democratic National Convention (DNC) and the vice presidential debate last fall. The other tool, Report Builder, lets Politico Pro subscribers create AI-generated write-ups of niche policy subjects using Politico’s archive.
The Guild found that both tools generated statements that were false, violated Politico’s style guide, or were taken down without formal corrections or retractions.
The arbitration transcript offers new insight into how Politico management is defending itself against the Guild’s case. In particular, management argues that the experimental AI tools in question are not subject to Politico’s normal journalistic editorial standards.
In his testimony supporting Politico management, deputy editor-in-chief Joe Schatz was asked if he would publish some of the error-filled output by Report Builder as articles on Politico.com.
Schatz replied that “the information in here and the way it’s organized and the way it says it does not reflect reality, so in that sense, no.”
He went on to argue that Report Builder sits “outside the newsroom,” since Politico’s product and engineering teams built the tool and editorial workers don’t review its outputs. As a result, he said, the AI-generated reports should not be held to the newsroom’s editorial standards.
“I would not publish this as an article, because it’s not an article,” Schatz testified.
One of a few points of contention during the hearing was whether LETO summaries and Report Builder outputs qualify as journalism — or, more specifically, as “newsgathering.” That’s critical because PEN Guild’s collective bargaining agreement, which went into effect in January 2024, states:
If AI technology is used by Politico or its employees to supplement or assist in their newsgathering, such as the collection, organization, recording or maintenance of information, it must be done in compliance with Politico’s standards of journalistic ethics and involve human oversight
Politico editors testified that “newsgathering” here means “reporters going out and reporting,” including talking to sources and reviewing documents. The Guild argued that both AI tools collect and organize information, and shouldn’t be exempt from Politico’s normal editorial standards.
“It was very upsetting to be sitting there in the room as a journalist and hear one of the leaders of my newsroom, that I’m very proud to be a part of, say it’s okay if our content doesn’t adhere to facts,” Ariel Wittenberg, unit chair of the PEN Guild and a witness in the arbitration hearing, told me. “If it doesn’t align with reality, why are we putting Politico’s name on it?”
“Politico has always been at the forefront of adopting new technology to better serve our audience. This is no different,” Politico spokesperson Heather Riley told me in a statement. “AI allows us to be more nimble, deploy advanced technology and adapt to a rapidly changing landscape. It also allows our great journalism to be seen and engaged with more frequently and in different ways.”
The arbitrator’s decision in the case will be binding. A ruling in the Guild’s favor could require that Politico cease operating the tools until they’re in line with the union contract, and after management enters a 60-day bargaining period over their use. A decision from the arbitrator is expected sometime after September 12, when both parties will submit additional post-hearing briefs.
“Our Super Bowl”
One of the Guild’s grievances centers around the AI-generated live summaries that ran on Politico’s homepage during the 2024 DNC and vice presidential debate. The tool, known internally as LETO, created a transcript of each event, then generated summaries of key statements and grouped them by policy category.
A handful of editors were tasked with monitoring the tool. They had the opportunity to greenlight initial policy summaries, but the tool automatically updated the published copy every five to ten minutes. Editors couldn’t review those updates before publication or make line edits to the AI-generated copy.
A screenshot of Politico’s homepage on the third night of the 2024 Democratic National Convention, with a LETO live summary published in the top right corner.
Across several nights, Guild members documented factual errors that LETO published on Politico’s homepage. JD Vance falsely stated during the VP debate that Donald Trump had always supported Obamacare. (Trump supported repealing Obamacare during his first term.) Vance’s statement was published without challenge. Among several copy errors, Kamala Harris’s mother’s name was misspelled.
As previously reported by Wired, the PEN Guild found LETO output terms that were at odds with Politico’s style guide, including using phrases like “illegal immigrant” and “criminal migrants.” Immigration policies of the Biden administration at large were also falsely attributed directly to Kamala Harris.
When summaries were taken down because of errors, no subsequent corrections or retractions were issued, as they would be for a Politico article or one of its live blogs.
When asked during the July arbitration hearing if “experimental projects” like LETO should be held to Politico’s journalistic standards, Schatz testified that “we’re rigorous in how we approach anything,” but also argued that live summaries are “fundamentally different from the reported journalism we have.” Schatz said LETO summaries were “meant to be ephemeral” and “a quick, short-lived transcript.” After each event, they were removed in their entirety from Politico.com.
The Guild questioned why any content that didn’t adhere to Politico’s normal editorial standards would lead the newsroom’s coverage of these high-profile events.
“Of all the times and all the places to experiment with something that you’ve decided doesn’t have to be up to snuff — why is it on your front page, on one of the most important nights of the year for your publication?” Wittenberg told me. “Why is that the time that we’re choosing not to follow editorial standards?”
In her testimony, Wittenberg described the DNC as “our Super Bowl.”
An AI-generated live summary of the 2024 vice presidential debate uses the phrase “criminal migrants,” which is not allowed per the Politico style guide.
Another point of debate during the arbitration was the amount of notice the Guild received about the LETO’s launch. The union’s contract requires that Politico provide at least 60 days notice for any AI technology that “will materially and substantively impact” the job duties of Guild members, and to engage in “good faith” bargaining during those 60 days.
Guild members testified that they were notified about the AI-generated summaries just an hour before the start of the DNC.
Politico argued in arbitration that LETO “did not transform anybody’s jobs” and so was not subject to the 60-day bargaining rule. Editors stated that LETO was a form of transcription, and minimized similarities between its summarizing and the work of staff reporters covering the events.
David Cohen, who was homepage editor the first night of the DNC, testified that the first time he saw LETO was when its summaries had already appeared at the top of Politico’s homepage. Cohen’s boss told him that night “it was something that was being done outside of the CMS [and] that I would have no access to either edit or update.”
Normally, Cohen decides headlines and the position of editorial content on the homepage. Despite objections about LETO’s quality, in this case, “I was told it was going to stay there.”
Please fact-check for accuracy
Another tool under scrutiny during the hearing was an AI feature on Politico Pro, the publication’s policy intelligence service, which has over 30,000 subscribers in lobbying and public affairs. The tool, called Report Builder, was developed with the startup Capitol AI. Launched in late February, it allows Politico Pro users to create AI-generated reports on niche policy subjects, using LLMs that comb through Politico’s article archive and other legislative data sets.
Jeremy Bowers, Politico’s global chief technology officer, testified that a report generation feature was one of the most highly requested updates by Politico Pro subscribers last year. “Politico was in danger of losing customers when we didn’t have this functionality,” he testified. (In 2021, the Financial Times reported that Politico Pro is responsible for more than half of Politico’s annual revenue.)
After the tool launched in beta, Guild members documented several major factual errors output by the Report Builder. One report filed into evidence was supposed to be about the Biden administration’s oil and gas policies; the AI-generated report instead listed actions taken by the Trump administration months after Biden left office.
Another report about the Farm Bill, a legislative package that covers nationwide agricultural and food programs, said the bill was criticized for “not protecting Social Security, Medicare, and Medicaid.” Those programs have never been part of the Farm Bill.
Semafor previously reported on similar problems with Report Builder in the lead-up to arbitration. For instance, it produced a report about two fictional lobbying organizations, the “Basket Weavers Guild” and “the League of Left-Handed Plumbers.”
AI-generated text output by Politico Pro’s Report Builder details fictional lobbying efforts of “the Basket Weavers Guild” and “the League of Left-Handed Plumbers.”
Politico counsel and Bowers argued that the reports were not journalism and should instead be compared to “Google search on steroids.” They argued that the outputs, which were sometimes several hundred words long, should not be held to Politico’s normal fact-checking standards.
“If you did a Google search for restaurants and then the first two things you saw weren’t restaurants at all, but actually furniture. I don’t think you would think to yourself, well, Google is a massive failure,” testified Bowers, explaining that the reports filed into evidence were likely done without prompt refining. “You would just modify your query until you got back something that you would like better.”
Sudeep Reddy, a former senior editor at Politico, who recently left the publication to lead MSNBC’s new Washington bureau, testified that he believed the Report Builder had “complied with Politico’s standards of journalistic ethics.”
Reddy emphasized that these tools were labeled. Beneath the LETO summaries was the line “live summary powered by AI.” The Report Builder, meanwhile, had an “AI-enabled” label and a footnote that encouraged readers to “please review and fact-check the information for accuracy.”
“Journalistic ethics is at the core around transparency with the audience,” testified Reddy. “Does the audience know what they’re seeing?”
When pressed by the Guild’s counsel on whether journalistic ethics also included “accuracy in reporting and fact-checking,” Reddy drew a distinction between labeled AI content and Politico reporting that is generally done “under a byline.” He argued that Politico’s audience “knows the difference.”
Axel Springer, Politico’s parent company, has been one of the industry’s most vocal advocates for newsroom AI adoption. At a global town hall in mid-July, days after the arbitration hearing, CEO Matthias Döepfner said no employee at Axel Springer should have to explain why they are using AI, Status reported.
“You only have to explain if you didn’t use AI. That’s really something you have to explain because that shouldn’t happen,” Döepfner said. He went on to clarify, “Of course, every mistake should be avoided and the credibility and truthfulness of our reporting is the most important thing.”
Concerns about credibility, the Guild testified, was one reason for seeking out arbitration. “For our trusted, verified news source Politico to say, here’s what I’ve got, but you have to fact-check it — that’s not the reader’s job. That’s why they come to us,” Wittenberg told me, echoing her testimony, which questioned whether these AI experiments could erode trust with readers and sources. “The transparency is just saying, you guys figure out if this is right.”
Screenshots of LETO live summaries on Politico’s homepage and Report Builder outputs courtesy of the PEN Guild.
