AI Transcription May Save Time, but Still Requires Fact Checking - Generative AI in the Newsroom

AI Transcription May Save Time, but Still Requires Fact Checking

A spate of journalists have been caught out publishing quotes made up by generative AI. Despite this, experts advise that AI tools can still be useful and we can avoid the same mistake by always checking quotes on the original source.

At the beginning of May, The New York Times issued a correction after its Canada bureau chief, Matina Stevis-Gridneff, incorrectly quoted a Canadian politician as calling his colleagues “turncoats.” An editors’ note at the bottom of the article states: “This article was updated after The Times learned that a remark attributed to Pierre Poilievre, the Conservative leader, was in fact an AI-generated summary of his views about Canadian politics that AI rendered as a quotation.”

Michelle Cyca wrote in the Walrus that copying quotes produced by generative AI is a “troubling” error. “This example makes the problem self-evident: generative AI programs — ChatGPT, Gemini, Claude, Grok, and the like — hallucinate, a term referring to their tendency to present fabrications as facts. Fabrications used to be a mortal sin in journalism.”

The New York Times spokesman gave me more detail on how exactly Stevis-Gridneff made the mistake: “Our reporter used A.I. to locate the most recent public remarks by Pierre Poilievre. The tool provided links to a video of a speech as well as purported transcribed quotes from that speech. The remark we initially published was, in fact, an A.I.-generated summary of Poilievre’s comments incorrectly rendered as a transcript. The reporter should have checked the words attributed to Poilievre against the video of his speech.”

Journalists can avoid this mistake by using a robust fact-checking process, advises journalism lecturer Paul Bradshaw. “The source of this claim isn’t the summary; the source is the original video. So you check it against that,” he advises. He adds that the word turncoat is central to the story, and this teaches us that “factual claims in the article that are central to it or are potentially legally or ethically more problematic — you would pay more attention to those.”

He adds that there is still value in using AI tools to speed up the process of finding quotes. “It will get you to the right place of that document, or the most useful part of that document, quite quickly, but you still need to look at that part. NotebookLM is particularly useful for that. It’s quicker for you to find where it got this information from.”

Co-founder of the Canadian AI Journalism Network, Rignam Wangkhang, has trained hundreds of CBC journalists to use AI tools accurately — and he sees different levels of risk in different uses of AI. “AI has been in transcription for a while now and has proven its worth. I think it’s a part of every journalist’s workflow; no one is transcribing by hand. Summarization is a different use case. Still useful but comes with slightly more risk, but can be very helpful.”

He suggests journalists “treat AI output as an unverified source.”

Errors in AI summaries

In March, experienced journalist and editor Peter Vandermeersch published made-up quotes on his Substack Newsletter. One of the people misquoted, Director Tow Center for Digital Journalism Emily Bell, said on Bluesky that he had attributed things to her that she didn’t say. Vandermeersch wrote in his “Press and Democracy” Substack newsletter, “I am admitting my mistake,” and explained the events leading up to the mistake.

“I acknowledge that, when I started my newsletter in the last quarter of last year, I used AI language models such as ChatGPT, Perplexity, and Google Notebook while writing. I was enthusiastic about the possibilities these tools offered and wanted to experiment with them extensively. Even I — with all my years of experience and knowledge — fell into the trap of hallucinations. I summarised reports using AI tools and worked from those summaries, trusting they were accurate… I wrongly put words into people’s mouths…”

Vandermeersch reminds us that it can be easy to forget to check when the AI-generated output is plausible: “These language models are so good that they produce irresistible quotes you are tempted to use as an author,” he wrote. He added what he should have done to avoid the mistake: “Of course, I should have verified them. The necessary ‘human oversight,’ which I consistently advocate, fell short…”

‘Sloppiness’

Get Clare Spencer’s stories in your inbox

Join Medium for free to get updates from this writer.

In May, a book called “The Future of Truth, How AI Reshapes Reality” was reviewed by The New York Times. The Times found more than a half-dozen misattributed or fake quotes in the section they reviewed. The book’s author, Steven Rosenbaum, is quoted in the Times as saying that the episode “serves as a warning about the risks of A.I.-assisted research and verification; that is why I wrote the book.”

Rosenbaum wrote in The Future of Truth’s acknowledgment section that ChatGPT, Claude, NaturalReaders, ProWritingAid, and Grammarly had helped “refine and polish the presentation of ideas,” reports Wired.

Whether Rosenbaum used AI to write or edit, Bradshaw said it’s clear that “whatever the reasons for arriving at those misquotes, it’s a mistake either way… It’s sloppiness. Some step has been missed in terms of checking, whether that’s checking his own notes or checking the AI version.”

‘Journalistic sin’

In February Ars Technica Editor-in-Chief Ken Fisher announced they had “published an article containing fabricated quotations generated by an AI tool and attributed to a source who did not say them. That is a serious failure of our standards. Direct quotations must always reflect what a source actually said. That this happened at Ars is especially distressing. We have covered the risks of overreliance on AI tools for years, and our written policy reflects those concerns.”

The journalist who wrote the retracted article, Benj Edwards, apologized on Bluesky. He described pasting a blog post into ChatGPT and “inadvertently ended up with a paraphrased version”. He told me “ChatGPT spit out the paraphrased version of the post, and somehow I confused that with the original blog post when I got back to writing because I had a fever and COVID — was literally working from bed.” The published article included three fabricated quotes.

Bradshaw told me this mistake could have been prevented if the Ars Technica team had made a checklist of facts to check. He said that when journalists use AI tools and agents, it becomes even more important to have a checklist. So he has added more content to his journalism course this year to encourage students to systematically break down an article into its claims.

He tells journalism students to check:

- What’s the source of this claim?

- What’s the evidence for each claim?

- Have you clearly attributed the claim?

“A checklist would hopefully have just forced him to take that step back and go ‘okay, I need to just go through this one more time. That quote, let me go back to the blog post. Is that the quote?’”

“We have very well-established workflows in journalism for addressing human flaws. A lot of our checks have eroded over time because of the speed of news. Those steps have been removed or weakened, and we haven’t yet built a new series of steps to address all the non-human errors that the AI can introduce,” added Bradshaw.

While we continue to learn together how we should and shouldn’t use AI tools, the way we handle mistakes is important. The journalists here who have come under scrutiny and reacted by sharing their exact workflows, no matter how embarrassing, have allowed all of us to learn from them and, hopefully, avoid the same mistakes.

The person who was misquoted was complimentary of how Ars Technica responded. Software engineer Scott Shambaugh wrote in his blog: “Ars Technica’s debacle is actually an example of these systems working. They understand that fabricating quotes is a journalistic sin that undermines the trust their readership has in them and their credibility as a news organization. In response, they have taken accountability and issued initial public statements correcting the record… This is exactly the correct feedback mechanism that our society relies on to keep people honest.”