Think of the last time you scrolled past a product review, skimmed a news story, or watched a video online—and paused, wondering if a human or an algorithm was behind it. That moment of uncertainty is becoming the default experience of digital life. Generative artificial intelligence has gotten so good at mimicking human language and imagery that the old tells, the awkward phrasing or obvious flaw, have largely disappeared. A new study from Temple University, led by Tanaka Manhede and Jason Chein and published in the journal AI & Society, wades directly into this disorienting landscape. The researchers wanted to know two things: does growing familiarity with AI tools change your ability to tell real content from machine-made content, and can that ability actually be trained? The answers are both unsettling and encouraging. People who use AI platforms often, and for many different purposes, turned out to be noticeably worse at spotting synthetic material. But here’s the hopeful twist: a relatively short, well-structured training session boosted detection accuracy by nearly ten percentage points, while people who received no training showed no improvement at all.
Why does this matter so much right now? Because the stakes are not academic. Generative artificial intelligence is already woven into everyday life, and it has become a powerful tool for deception. Fake hotel reviews, AI-written phishing emails, cloned voices in phone scams, and fabricated images in disinformation campaigns are no longer science fiction; they are routine realities. And human beings, it turns out, are not natural-born detectors. Earlier studies showed that people trying to identify whether scientific abstracts, human faces, or videos were real or synthetic often performed no better than if they had simply flipped a coin. Some conditions helped: professional writers could spot AI-authored essays better than casual readers, and people judged general-interest news more accurately than specialized scientific news, likely because they lacked the background knowledge to catch subtle errors. Adding multiple types of information, such as pairing a transcript with audio and video, also tended to improve judgment. That finding matters because it suggests there is a real signal separating human from machine output, even when most of us fail to notice it. The trouble is that we are not using that signal well. The Temple team suspected that individual differences—both in attitudes toward AI and in habits of using AI—might explain why some people are sharper than others.
They set up a clever experiment to test that idea. One possible prediction was that heavy use of AI platforms would act as practice, sharpening a person’s instinct for that system’s quirks and stylistic signatures. Another prediction, grounded in theories of extended cognition and cognitive offloading, suggested the opposite: the more we lean on external tools to think for us, the blurrier the boundary becomes between what our own minds produce and what arrives from outside. Under that view, frequent AI use would not make you a better detective; it would make you more accepting of machine output, less attuned to the faint cues that mark it as artificial. The researchers recruited 117 adults, aged 18 to 34, all fluent English speakers in the United States, through the Prolific platform. The task was deceptively simple: look at vacation rental listings, presented in the familiar Airbnb format, and decide whether each one was written and illustrated by a human or generated by AI. The material was carefully constructed. Half the listings were genuine, drawn from real Airbnb properties with high ratings and human-verified descriptions, paired with real photographs and real human faces. The other half were fully synthetic: texts generated by ChatGPT 4.0 and Claude 3.5 Sonnet, matched in length to the human writing, accompanied by AI-created property images from Mage.space and Gemini, and AI-generated faces from StyleGAN2. Interestingly, the AI-written texts were grammatically cleaner than the human ones, so participants could not simply rely on spotting clumsy errors. At the start, before any training or feedback, participants judged 32 listings, half text-only and half text plus images. Their average accuracy was 57.4 percent—only a hair above chance. But the individual range was enormous, from 34 percent to 90 percent. That wide spread was the clue that something meaningful was operating beneath the surface.
What explained that spread? Not what you might expect. Attitudes toward AI, which were generally positive in this group, had zero predictive power. But a new measure called the Socioaffective and Cognitive Artificial Intelligence Engagement Survey, or SCAIES, told a different story. The more deeply people had integrated AI into their daily lives—not just as a search engine replacement, but as a companion, a sounding board, a partner for emotional and social and cognitive tasks—the worse their detection ability. The correlation was negative and significant, around negative 0.29. This held true whether the engagement was socioaffective, meaning people used AI for social and emotional purposes, or cognitive, meaning they outsourced effortful mental work to it. Intriguingly, an objective measure of actual ChatGPT usage, based on logged sessions over the previous 30 days, did not predict discernment at all. Participants even slightly underestimated how often they used the platform. So it wasn’t the raw number of times a person opened ChatGPT that dulled their instincts. It was the qualitative way they turned to it, letting it seep into the fabric of their thinking and feeling. The more natural and familiar AI-generated content became, the less foreign it felt. This is what the researchers describe as a process of normalization: the mind begins to classify machine output as part of its own landscape, and the boundary quietly erodes.
If that were the whole story, the outlook would be grim. But the second half of the study offered a practical counterweight. Sixty participants were assigned to an experimental training group, and 57 to a control group. Both began at the same modest baseline, around 57 percent accuracy. The experimental group then moved through a structured two-stage intervention. First, they were shown 40 accurately labeled listings, clearly marked as either human or AI-generated, with each item staying on screen for at least ten seconds. Half were text-only, half included images. The control group watched the same number of listings for the same amount of time, but without any labels. In the second stage, experimental participants evaluated 40 new listings and received immediate feedback on every answer: Correct or Incorrect, with a small monetary reward of 25 cents per correct response and a running total displayed on the screen. The control group had the same financial incentive but received delayed feedback and no running tally. The results were dramatic. The experimental group’s accuracy jumped by 9.4 percentage points from baseline to final assessment, while the control group’s performance actually declined slightly, by 0.6 points. Most of the gain, a full 6.7 points, materialized immediately after the labeled exemplar stage alone. That tells us something important: detection failure is often not about lacking the visual and linguistic cues—it’s about not knowing which cues matter. Once people were shown what to look for, their attention recalibrated. The additional feedback and incentive stage added another 2.7 points, though not statistically significant. Reaction time analyses ruled out a speed-accuracy tradeoff, meaning participants got better not by thinking harder and longer, but by using cues more efficiently. Perhaps most encouraging, the people with the lowest baseline scores gained the most, and prior AI engagement, though it predicted weak baseline performance, did not predict resistance to training. In other words, the desensitization caused by habitual AI use was reversible.
What exactly were these newly sharpened detectors picking up on? The researchers dug into the underlying data. Using an AI model called Google’s Universal Sentence Encoder to measure semantic similarity among all the texts, they found something striking: AI-generated listings were far more similar to one another than human-generated listings were. The intra-class similarity for AI texts was about 0.50, compared with 0.45 for human texts. In plain terms, machine writing tends to converge on a clean, prototypical pattern, while human writing is messier, more varied, more idiosyncratic. The human listings that stood out as most semantically distinct were also the ones judged most accurately at baseline, suggesting that even untrained participants were unconsciously sensing this distributional signature. The same pattern appeared in images: AI-produced pictures were more homogeneous than real photographs, as measured by a ResNet-50 model. There was also an interesting split in how people judged different formats. Participants identified AI content more easily when it was text-only, but they identified human content more easily when images were included. This suggests that different modalities carry different cues, and successful detection requires adapting to the format at hand. The broader implications ripple outward from this single experiment. In education, where students are increasingly turning to generative AI for essays and assignments, digital literacy training may need to become a core skill rather than an afterthought. In media and journalism, where synthetic images and text can spread viral misinformation in minutes, accurate labeling and user education are not optional safeguards; they are essential infrastructure. There are limits, of course. This study focused on one narrow genre, promotional vacation-rental language, and a sample of young, digitally fluent adults. The training gains may fade over time, and no one knows how durable they are. Current AI systems still leave detectable statistical fingerprints, but as models become more diverse and sophisticated, those fingerprints may fade or shift. Still, the central message is both clear and empowering: human discernment is not obsolete, and it is not fixed. It varies widely from person to person, it is dulled by habitual AI engagement, and it can be measurably restored through brief, well-designed training. At a moment when the line between human and machine creation is blurring before our eyes, that is a reminder that our judgment, fragile as it may be, can still be sharpened.