We Tested Our Team on Spotting AI-Generated Work. They Failed.

Can employees tell if content is AI generated?

Mostly, no. When we quietly tested our own team on a mixed batch of AI- and human-made text and images, accuracy barely beat a coin flip. That's not an outlier. Peer-reviewed studies put human detection accuracy for AI-generated content anywhere from 19% to 76%, depending on format and test conditions, and most of that range sits uncomfortably close to random guessing.

If your AI governance plan assumes employees can eyeball the difference, the data says otherwise. Here's what we found, what the research confirms, and what to build instead.

What happened when we ran the experiment on our own team?

We pulled together a stack of work samples: short blog passages, internal memos, product images. Half were AI-made, half were made by our own writers and designers. Nobody labeled which was which. We handed the stack to the team, asked everyone to mark each item AI or human, and asked for a confidence rating on every guess.

Nobody knew a test was happening. We wanted honest first-instinct calls, not careful detective work, because that's how people actually meet AI content on the job. Nobody forensically analyzes a Slack message or a stock photo before using it.

Overall accuracy landed close to a coin flip. Text was easier to call than images, but not by much, and the pieces that fooled the most people were the ones raters were most confident about.

How did the guesses break down between text and images?

A few patterns stood out once we tallied the results:

• Overall accuracy sat just above half. Slightly better than random, but not a margin anyone should build a policy on.
• Text outperformed images. Raters did noticeably better spotting AI-written passages than AI-generated images, mirroring the pattern in the commercial research on this exact question.
• Confidence did not track with correctness. The items rated "obviously AI" or "obviously human" with the highest confidence were wrong almost as often as items people flagged as unsure.
• A handful of images fooled almost everyone. AI-generated images with human faces got misclassified far more than ones showing objects, food, or landscapes, which lines up with what larger studies have found.

No one on the team beat the group average by much. This wasn't a case of a few careless guessers dragging down a sharp majority. Nearly everyone landed in the same narrow, unimpressive band.

Are we alone, and what does the published research say?

Our informal test isn't rigorous science, but it tracks with a growing body of research landing on the same conclusion: humans are not reliable detectors of AI-generated content, and the accuracy range is wide enough to be nearly useless as a governance strategy.

• A peer-reviewed study using healthcare-journal writing found human raters correctly classified AI vs. human text only 19% of the time, worse than the 20% expected from pure chance.
• A large ACM study with 1,276 participants under near real-world social-feed conditions found overall detection accuracy of 51.2%, with image-only detection the worst at 49.4%.
• A Microsoft AI for Good Lab analysis of roughly 287,000 image judgments from more than 12,500 participants found an overall success rate of 62%, described by the researchers as "slightly above chance."
• A Kellogg School study with over 50,000 participants found people correctly identified AI images 76% of the time, but individual accuracy ranged from just 32% to 100%.
• A commercial survey from Nexcess found respondents correctly identified AI-generated copy 57% of the time and AI-generated images only 53% of the time.

The range is wide because the tests differ: some use faces, some use landscapes, some give unlimited time, some give one second. None of them land anywhere close to the confident, near-perfect detection rate that "just look closely" policies assume.

Why can't people detect AI-generated content?

A few mechanics explain why detection is this unreliable.

1. People default to believing content is real. In the ACM study, participants correctly identified genuine content 64.6% of the time but only caught synthetic content 38.8% of the time. The default assumption is authenticity, not suspicion.
2. Faces are a blind spot. Detection accuracy for AI images drops to 46.6% when a human face is present, versus up to 54.7% for non-face subjects like animals or landscapes.
3. Viewing time matters more than instinct. In the Kellogg study, accuracy on AI images rose from 72% at one second of viewing to 80% at ten seconds. Snap judgments are measurably worse than slow ones.
4. Detection tools don't agree with each other, let alone with humans. Free AI-text detectors like ZeroGPT and Grammarly can rank escalating AI involvement reliably, but their agreement with one another was as low as an ICC of 0.57, and some mislabeled fully human text as up to 6.5% AI and fully AI text as up to 92.5% human.

Put plainly: the tells that used to give AI content away, robotic phrasing, warped hands, uncanny symmetry, are disappearing faster than people's instincts are adapting.

Is the real risk embarrassment, or shadow AI?

A failed internal quiz isn't the actual danger. The danger is what's already happening at scale while detection stays this weak.

Employees aren't waiting for IT-approved tools. A Cybernews survey of more than 1,000 US employees found 59% admit to using AI tools their employer hasn't approved, and 85% of employees who do have approved tools still use unapproved ones anyway, often feeding them sensitive company or customer data. That's the exact behavior covered in The "Free Claude" Problem: Why Shadow AI Is Your Next Governance Nightmare: employees routing around approved tools and pasting in data nobody signed off on. UpGuard's 2025 State of Shadow AI report corroborates this at even higher rates: 81% of employees and 88% of security leaders admit to using unapproved AI tools themselves.

That's the compounding problem. If employees can't reliably tell AI output from human work, and they're also routing around approved tools to generate that output in the first place, "review it carefully before it ships" isn't a control. It's a hope. Teams that have run into this gap with code have written about it in Shadow AI Governance: How to Secure the Agents Employees Build, and the same logic applies to text and images: you can't govern what you can't see, and you can't rely on people spotting it after the fact.

What actually helps people spot AI content?

Generic "stay alert" advice doesn't move the needle. The research is specific about what does.

University of Aberdeen and ANU researchers trained participants on six concrete perceptual cues for spotting AI-generated faces: symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability. Baseline accuracy sat around 40%. After roughly an hour of structured training, it rose to 80%, with a few participants reaching near-100% accuracy.

That's the difference between telling people to "be careful" and giving them a checklist. Slowing down helps too: the same viewing-time effect that pushed Kellogg's participants from 72% to 80% accuracy shows a few extra seconds of deliberate attention beats a confident snap judgment every time.

Awareness campaigns without structure don't work. Specific, trained criteria and slower review do.

What should an AI use policy include?

Given all this, the policy question isn't "how do we get better at spotting AI content." It's "how do we stop needing to." A workable acceptable-use policy shouldn't depend on anyone's ability to catch a fake. It should make disclosure and review structural.

A solid AI use policy needs five pillars, adapted from the standard acceptable-use-policy framework used across the industry:

1. Scope and roles. Define who the policy covers and who owns enforcement.
2. Authorized tools and prohibited use cases. A specific list, not a vague standard, of what's approved and what's off-limits.
3. Data-handling rules. What can and can't be pasted into an AI tool, and where that line sits for customer and financial data.
4. A governance and change process. A defined path for vetting new tools, drawn from the same accountability model laid out in How to Manage AI Agents in the Enterprise (Before They Manage You), instead of an informal free-for-all.
5. Concrete safe-use guidance, including a requirement that AI-generated content is verified before it ships, not assumed accurate.

Layered on top of that, borrow the graduated hierarchy from the healthcare-writing study: treat light AI editing (grammar, spelling) as low concern, heavy AI editing as a grey area requiring discernment, and fully AI-generated text produced from bullet points or a prompt as the highest-scrutiny tier, requiring disclosure by default. The same tiers apply to images: a lightly retouched photo isn't the same risk as a fully generated one used in customer-facing material.

Disclosure only works if it's built into the workflow, not remembered after the fact. That's the same reason it makes more sense to own the systems that track this than to bolt governance onto rented tools after the damage is done, which is the whole argument behind platforms like Remy: if you own the infrastructure, you can log who touched what and when, instead of hoping someone remembers to flag it.

What is the bottom line?

Our team failed the same test the published research keeps failing: detecting AI-generated text and images is close to a coin flip, and confidence doesn't predict correctness. That means "look closely" isn't a control. Here's what to do instead:

• Stop writing policy that assumes employees can catch AI output by eye.
• Build disclosure into the workflow, with a graduated tier system, not a binary ban.
• Assume shadow AI use is already happening at rates near 60% to 88%, and build oversight around that reality, not around the hope that it isn't.
• If you invest in training, make it specific and structured, not a generic "be vigilant" memo.

Detection isn't the control. Process is.