AI put “artificial words” in his book. But this author wants to continue using it.

Through the cracks

Rosenbaum used AI tools for writing, he told me, “to generate ideas, find articles, summarize topics, identify people or papers I might want to look at.” He makes a distinction between this type of research and “the actual reports, the narrative process, the interviews, the arguments, and the conclusions contained in the book,” which he says are “all mine… There was no time when AI was writing this book.”

In addition to topics based on written interviews that Rosenbaum says he conducted himself, The Future of Truth it also includes topics related to research in which Rosenbaum said, “We’re extracting facts and then piecing them together into a story.” Tools such as OpenAI’s ChatGPT and Anthropic’s Claude were widely used to gather information, he said, with any tools mined by the tools marked with the warning “this is from AI” in its notes.

It’s incredibly creative and insidious and strange in all these ways… and then it gives you in ways that are even worse.

Steven Rosenbaum

The AI-generated content was provided to the researcher and two copywriters were provided by the publisher, Rosenbaum said. Of the 285 words in the book, six were identified by the Times as problematic, including three so-called “made-up words” that have no source. (More examples can be found as the book is reviewed. And it is important to note that many authors can combine zero automatic words when writing a book.)

“I think we did it (the double check) admirably, but not a hundred percent,” Rosenbaum told Ars. “We’re working, we’re doing everything we can. We check, it looks good. We check twice, then we make a mistake.”

But the big failure here shows how the fact-checking process can be without tools to support AI-assisted research. In the past, a researcher could be confident that any writer who quoted a written text had simply written the text directly. These documents should be checked, but the fact that they are easy to verify makes them less likely to be doubted. If AI tools are involved anywhere in the pipeline, that logic goes out the window, and there should be some doubt that the words were copied correctly or even exist at all.