Show summary Hide summary
- How the imitation experiment was set up
- Results that blurred the line between human and machine
- Why AI can capture an author’s “voice”
- Signs people often cite as “AI giveaways”
- How readers react when they know a text is AI-made
- Implications for authors and the publishing world
- What this means for “authentic” writing
- Questions left open by the experiment
Vauhini Vara, the novelist and journalist, ran a simple but unsettling test: she asked people closest to her to tell which passages were hers and which were written by an AI trained on her work. The result surprised many — and forced a rethink about what makes writing authentic, and how easy it has become for machines to imitate a human voice.
How the imitation experiment was set up
Vara worked with a researcher who had previously trained models on the styles of established writers. The researcher fed an AI the author’s earlier books and published journalism.
Nas slams music industry over disrespect: teases new music, vows to address it
1992 vocal performances: 3 jaw-dropping moments that still amaze
Then the AI generated short passages meant to sound like material from her forthcoming novel. Vara mixed those passages with excerpts she wrote herself.
She circulated the mixed texts to a tight circle: friends who had read her work for decades, including her oldest friend and fellow writers. The question was simple. Could they tell which pieces were hers?
Results that blurred the line between human and machine
Most readers could not reliably distinguish between Vara’s sentences and the AI’s imitations. Some AI pieces were praised for their narrative arc and emotional clarity.
In many cases the automated text landed exactly where a human reader expected it to. One AI-generated passage about mentorship and fulfillment felt complete and purposeful. Listeners reacted as they would to a well-crafted human paragraph.
On the other hand, a small grammatical slip in one of Vara’s own excerpts made it stand out as human. She noted that the error actually helped her friends identify her writing.
What surprised the participants
- Friends who knew Vara’s voice intimately were often uncertain.
- AI passages sometimes received stronger praise than the human ones.
- A single, obvious human mistake made an excerpt feel more “real” to readers.
Why AI can capture an author’s “voice”
Language models learn by pattern. They digest massive amounts of human writing, including the quirks that make a writer unique.
By training on an author’s earlier work, a model can reproduce recurring rhythms, favorite metaphors, and structural tendencies.
That means imitation is less about magic and more about statistics. The AI predicts what comes next based on patterns it has seen. The result can mimic tone and structure closely.
Common misconceptions about “soulless” AI prose
- Many assume AI always sounds bland or generic.
- But models can be steered to produce a range of styles.
- Product design choices make many public AI tools sound corporate.
Signs people often cite as “AI giveaways”
Writers and critics list several traits they believe reveal machine-generated text. These can be useful, but are not definitive.
- Overuse of similes or metaphors that feel off-mark.
- Repetitive sentence patterns and predictable phrasing.
- An adjective attached to almost every noun.
- Odd negative parallels or constructions that sound mechanical.
Even so, these patterns are not unique to AI. Human writers can exhibit the same habits. Short fragments worsen detectability because they offer less context.
How readers react when they know a text is AI-made
Research offers a paradox. When readers don’t know a piece is AI generated, they may prefer it. But labeling it as machine-made changes responses.
Studies show that disclosure reduces interest. People who learn a text was created by AI become less engaged.
Perception matters as much as craft. The social context shapes whether a reader values a given piece of writing.
Implications for authors and the publishing world
For writers, the experiment raises thorny questions about originality. Vara said the test made her reconsider how distinct her new work really is from earlier books.
Publishers and readers face trade-offs. AI can lower costs and produce convincing drafts. It can also blur provenance and erode trust.
There is also a power dynamic. Many fear cultural production becoming dominated by a few wealthy tech firms.
Practical concerns for the industry
- Detection tools struggle with short excerpts.
- Contracts and disclosure policies are not yet standardized.
- Creative labor may change as machine assistance becomes common.
What this means for “authentic” writing
The test suggests authenticity is not a simple technical property. A human error, a unique turn of phrase, or a personal history can make text feel authentic.
At the same time, models can replicate those hallmarks well enough to fool close readers.
The key tension is between craft and context. A passage can be expertly crafted by a model yet lose value if the audience knows it’s machine made.
Questions left open by the experiment
How should readers be informed about the use of AI in books and journalism?
What standards should publishers enforce to preserve authorship and transparency?
How will writers adapt their practices when AI can so convincingly echo their previous work?
