Ashish Dsa is the CTO and co-founder of Arbor, a voice AI company that turns frontline conversations into actionable intelligence.
My job includes designing and creating artificial intelligence (AI) voice agents, so many of the hours I spend at my desk include reading about the generated text and developing detection systems that identify when a model deviates from the script. When you read enough of this type of information, you do not read for content anymore; you begin to look for the signs. The text now admits.
This ability is no longer unique. If you are looking for someone to fill an open position, you will likely be reviewing a resume created using a chatbot. If you are approving new marketing copy or signing off on a report provided by a third-party vendor, there is a chance that each sentence in each paragraph was created by a model and edited only enough to fool. Identifying these differences can protect your hiring process, your company’s reputation and your time. The good news is that identifying the differences is something you can learn and is not magical.
In 2025, researchers from the Association for Computational Linguistics tested a group of people who used language models extensively to see how well they could distinguish whether a particular article was written by a human or AI. Five participants voted on each of the 300 articles, and the group performed better than all the commercially available detectors, including those designed to detect AI-generated text that has been rewritten to conceal its original source.
Forget the detector tools. The detector tools will flag writing done by individuals whose first language is not English, as well as other innocent writing. Being falsely accused would cost you significantly more money than being incorrectly dismissed. Your best tool is to develop your skills as a reader.
Where The Machine Gives Itself Away
Vocabulary is where you start, but it’s not where you end. There truly are statistical fingerprints here. In their analysis of over 15 million scientific abstracts for a 2025 article in Science Advances, researchers discovered the word “delve” rising to about 28 times its pre-ChatGPT use, with “underscores” and “showcasing” close on its heels. You can recognize the other members of the family from experience: tapestry, realm, landscape, robust, seamless and multifaceted. Read it out loud and listen. “I dug into the numbers” is human language. “I delved into the multifaceted tapestry of the numbers” is likely AI.
Vocabulary is the easy one to spot, though, and becomes less relevant the more models are retrained. Structural habits, however, stick around longer. Notice the negated structure that AI models love: “It is not a product. It is a movement.” Notice how concepts come in threes, because machines go for three points when given the chance. Notice how the clauses don’t add anything new and instead just reinstate what’s already been said. “It highlights its importance for the broader industry” doesn’t add any value. And notice how models can’t help finishing up with an unnecessary “in conclusion,” even when writing a four-line email.
Context is key, which is what makes all the difference in the workplace setting. On LinkedIn, the tell is intensity in the absence of specifics: a post “humbled to share” a “truly transformative” quarter without ever specifying a number, a customer or a person. A real CEO says, “We missed our Q3 target by 12%, and I almost didn’t send that email to the board.”
Product reviews have their own tell; the machine sounds like the specification sheet. “This battery provides adequate power for extended usage during the day,” the fake reviewer writes; “I did not charge it on Thursday, and it was still three-quarters full by Saturday,” the real reviewer writes.
Cover letters might reveal themselves through blind admiration, praising your company’s commitment to innovation while being unable to say anything specific about what you do. The tell becomes clear once you ask them about that in the interview. Customer support might use the triple apology that has no information about your problem whatsoever. And finally, when it comes to entry-level jobs, look for what teachers refer to as a “polish jump.” This often takes the form of perfect semicolons and fancy terms like “ubiquitous” coming from the author whose previous paragraphs sound completely different.
The character that I enjoy most of all is the cast of characters. In the same study, over 60% of the AI-written articles mentioned either Emily or Sarah as the main characters. Once you start recognizing them, the invented Emily will be showing up everywhere.
What To Do With A Hunch
An honest caveat before an article warning you about overconfident machines becomes a potential sign, too: no one sign guarantees anything. Human writers will say “delve,” occasionally use triples and create clear prose on time. The ones who can do this consistently are precisely those who employ such techniques on a regular basis. Everyone else writes worse than that. Therefore, consider the presence of any of these signs only as a pointer, never as a judgment call. What matters is the accumulation of these pointers, the combination of them all and the absence of the human leftovers, such as numbers, names, opinions or even a slight sentence risk.
The latter is the key. A machine generates for predictability and consensus. Humans leave their traces. The more you search for them, the clearer the machine will sound. So, the next time something reads a little too perfectly, trust that instinct. Now you know the tells.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
