Recent research showcases a striking phenomenon worth delving into—companies that boast meticulous brand guidelines are discovering that polish is not only an advantage, but also a liability.
If you flagged that opener as AI generated, then you’ve probably spent a good amount of time on the internet lately. Nearly three-quarters of new web pages contain content generated at least in part by machines, according to an Ahrefs analysis of 900,000 pages.
“There is distinct AI writing on pretty much all linguistic levels: word choices, sentence structures, phrasal constructions—everything,” says Tom Juzek, assistant professor of computational linguistics at Florida State University and co-author of Why Does ChatGPT 'Delve' So Much?
When customers interact with AI customer service, the AI “voice” can leave them feeling unimportant and rebuffed. Customers note that humans are more likely to provide service that feels personalized and empathetic than machines, according to ServiceNow’s The CX Shift customer experience (CX) survey.
"If your customers see your product as generic, that might resonate less than a personal, individualized effort,” Juzek says.
It's tempting to think we know AI when we see it. In a recent co-authored paper, Hiromu Yakura of the Max Planck Institute for Human Development found that AI overuses the word “delve,” along with “showcase,” “boast,” “intricacies,” and “meticulous.”
However, these AI tells are less reliable as more people come to recognize them. "Delve" has now fallen below its pre-ChatGPT baseline, the research found, while the words less often associated with AI still appear online at elevated levels.
Organizations are starting to formalize the race between detection and obfuscation. Claude now embeds invisible watermarks in text by statistically biasing word choices according to a key held by Anthropic. The watermark signals that Claude processed the text, not that Claude authored it, creating a detection tool that persists even through light editing.
However, the watermark has a flaw: Paraphrasing defeats the signal. Once a human has edited the text, the watermark starts to degrade. With enough editing, it thins to invisibility.
These technical fixes miss something deeper: Detection is fundamentally social, Juzek argues. "Whenever I sense that a piece of content has been low effort, then I'm OK with a low-effort reply," he says.
People are running calculations constantly, evaluating whether something feels genuine and heartfelt. They respond better to apparent effort even when the prose is rough. Once a watermark exists, people start wondering whether that watermark has been circumvented. No technical marker can override that judgment.
If customers can't reliably identify AI by vocabulary, watermarks, or a static signal, what are they sensing, and why?
Nearly half (46%) of customers say chatbots don’t understand their questions and concerns, according to The CX Shift. That’s a different complaint than bad writing. A system that can’t tell what distinguishes one customer from another will produce generic responses. Generic content is devoid of context, and people respond negatively to that as both customers and humans.
This is the irony of deploying AI systems badly. What separates large language models (LLMs) from previous technology is that they can be specific. "They generate personalized interactions tailored to each conversation, in contrast to traditional media, which simply broadcasts the same content to everyone," Yakura says.
Organizations have taken the one technology capable of addressing a customer individually and instead used it to broadcast.
The penalty for generic content is not algorithmic. Google evaluates content on quality rather than provenance. Ahrefs found essentially no correlation between AI content and ranking penalties.
Rather, the penalty comes from people. In a 2025 study in Nature Human Behaviour, Matan Rubin and colleagues ran nine experiments with more than 6,000 participants. Every response was written by AI, but only certain responses were labeled as such.
Participants rated the responses labeled human as more empathic and more supportive, regardless of whether those responses were actually written by AI. The effect was greatest for messages about care and shared experience.
The label is rarely supplied in practice. Customers infer it. Participants who decided on their own that AI was involved, without being prompted to consider it, ranked those responses lower.
Watermarking offers one answer: Embed the AI signal directly so that customers can't miss it. Yet even when the mark is supplied systematically, people's social evaluation still determines their judgment.
Alexis Walker, an innovation writer at ServiceNow whose team focuses on imagining technologies for the future, points to what makes the inference easy to reach: "It's not just that the writing sounds the same, but [that] the ideas are the same."
This doesn’t mean we should pull AI out of customer-facing work. Customers say they’re willing to be served by a machine, according to The CX Shift. In fact, 75% try self-service before calling anyone.
However, executives have been treating empathy as a tone problem, solvable with warmer instructions. Customers rank emotional connection among their top four service values, alongside responsiveness, trustworthiness, and security, according to The CX Shift. It’s also where organizations have made the least progress.
The cause is structural. An AI agent working from partial records has nothing specific to say, so it falls back on platitudes.
Most organizations already hold what they'd need to know about a given customer, but it’s scattered across systems that don't speak to each other. Only 43% have integrated that data into a single source of truth, according to The CX Shift. And 80% of service reps log in to three to five systems to resolve a single issue. This is robbing businesses of valuable context.
The context problem is only getting worse. Organizations with streamlined, integrated workflows across business functions fell from 30% in 2025 to 16% in 2026 because of fragmented systems and agent sprawl, according to the ServiceNow Enterprise AI Maturity Index 2026.
The participants in the Nature Human Behaviour study penalized responses they merely suspected of being AI. No amount of context or integration can solve for a customer who’s already made up their mind. Although specificity doesn’t grant immunity, it does lower the odds that anyone reaches that conclusion.
Specificity means knowing when to walk away from the simpler, more tantalizing path, Walker says. Although she and her team use AI to deliver work at scale, she emphasizes that ideas and perspectives are the most valuable assets they have. Striking a balance between creativity and scalability is essential for humans and corporations alike.
"If we’re making all our writing sound the same, if we’re losing our voice and authenticity, then it's not working for us,” she says. “It's reducing the value of something irreplicable."
Find out how ServiceNow can help you put AI to work for people.
