Can people tell if you used Claude?

The Detection Panic: Why You Can't Trust AI Detectors (Either Way)

You pasted a draft into an AI detector, and it flagged the paragraph you wrote yourself in fifteen minutes on four hours of sleep. Meanwhile, the fully Claude-generated section above it sailed through as "100% human." That's not a fluke — it's what the research consistently shows about detection tools, and it cuts both ways.

Here's the uncomfortable truth: detectors can neither reliably catch you nor reliably clear you. Overall accuracy on mixed human and AI text ranges from just 26.3% to 76% across the leading tools, according to JISC's National Centre for AI assessment. A coin flip would beat some of these products. And if you specifically used Claude, your odds of slipping through are even better: in one 30-tool comparison, Quillbot, Scribbr, and ContentDetector.ai failed 100% of Claude tests, and Writer, Grammarly, SurgeGraph, and SciSpace failed every AI test outright.

The false-positive problem is just as alarming, and it's the side nobody talks about when they're worried about getting caught:

• Free and lesser-known tools mislabel human writing at startling rates — Crossplag at 30%, Content at Scale at 28%, GPT Zero up to 22%
• Even a "good" 1% false positive rate means roughly 4,800 false accusations per year at a 20,000-student institution
• In comparative testing, JustDone flagged all human-written text as AI, and ZeroGPT and ContentDetector.ai also misclassified genuine human writing

So the detector that clears your work today might accuse your colleague tomorrow, and the one that flags you might have missed the actual AI text in the same document. Detection technology simply hasn't kept pace with generation — JISC notes that generative AI tools "would generally outpace the development of AI detection tools and that is indeed proving to be the case," and the education sector has shifted decisively away from automated detection as a result. Paraphrasing tools and light manual edits can drop Turnitin's detection rate to as low as 7.9%.

For managers, the implication is sobering: Federal Reserve economists document "reporting and monitoring difficulties of generative AI," and 29% of surveyed workers say management has no idea they're using it at all. You cannot detect your way to an AI policy. Clear norms and calibrated disclosure — not detector scores — are the only stable ground, and we'll get to what that looks like in practice.

The same verification gap applies to the tools layer of your AI workflow. If you're equipping Claude or another agent with skills, published research found roughly 26% of agent skills carry at least one vulnerability — which is why SkillzBench is building a free directory where every skill is security-scanned, pinned to a commit SHA, and benchmarked so you can see what actually helps before you install it.

Humans Are Even Worse at Spotting AI — and the 'AI Tells' Are Myths

Your boss almost certainly cannot tell. According to JISC's National Centre for AI, "unless a reader knows the writer personally and can detect a marked deviation from their usual style, humans generally aren't good at spotting AI written work." The people closest to you — a long-time colleague, a professor who has graded your essays for a semester — might notice something is off. Everyone else is guessing.

The internet's favorite "AI tells" don't help either. Em dashes, the word "delve," suspiciously tidy structure — JISC notes most of these popular indicators are unreliable. Plenty of humans write that way, and modern models can be steered away from any stylistic quirk in one sentence of instruction.

Academics — professional readers, trained to spot weak writing — don't do any better. In one study cited by JISC, AI submissions in a psychology degree "Turing Test" scored just over half a classification boundary higher than real students' work, and other research found generative AI "can generate authentic assessments that pass the scrutiny of experienced academics." The markers didn't just miss the AI; the AI quietly outperformed the humans it was imitating.

The workplace numbers confirm that hidden use is already the norm. A Conference Board survey analyzed by Federal Reserve economists found that 29% of workers said management was not aware of their AI use — and the economists themselves flag "reporting and monitoring difficulties of generative AI" as a reason adoption statistics vary so widely across surveys. In practice:

• Most AI-assisted work is invisible to managers and readers alike
• Stylistic "tells" are myths — humans can't reliably identify AI text
• Even experienced academics have graded AI work as better than human work
• Roughly a third of AI users say their management simply doesn't know

None of this means concealment is the right strategy — but it does mean detection is a dead end as a management tool. Policies built on "we'll catch it" are already failing quietly, every day.

The more productive question is how to make AI-assisted work verifiably good. That's the same philosophy behind SkillzBench's approach to agent skills: rather than guessing which instruction packages help or hurt, every skill gets security-scanned and benchmarked with and without it, so quality is measured rather than assumed. Judge the work, not the suspected origin — because the evidence says you'll never reliably spot the origin anyway.

The Disclosure Paradox: Transparency Helps — Until It's Too Detailed

Here's a puzzle worth sitting with: readers say they want to know when AI helped write something, yet telling them too much can cost you their trust. That tension is exactly what disclosure decisions hinge on.

A recent arXiv lab study (n=40) found that only detailed AI disclosures reduced reader trust and subscription rates — while one-line disclosures performed similarly to no disclosure at all. In other words, a brief "AI assisted this article" note costs you almost nothing, but a full accounting of your AI workflow may actively drive readers away.

The twist is that readers don't see it that way themselves. Roughly two-thirds of participants still preferred detailed disclosures for transparency reasons. People want the information in principle — but when they actually receive it, trust drops. Prior research reinforces the penalty: AI-labeled headlines were rated as less accurate even when factually correct, and "AI-generated" labels reduced perceived accuracy and willingness to share regardless of the actual source.

So what should you actually do? The evidence points toward context-dependent, detail-calibrated disclosure rather than blanket labels or total silence. The arXiv researchers note that not all AI disclosures create this dilemma — outcomes depend on how the disclosure is framed. In practice:

• Keep the disclosure to a single clear line when AI played a supporting role — these performed like no disclosure in trust outcomes.
• Avoid generic labels like "This article was created using AI," which the study found are frequently misinterpreted or ignored entirely.
• When deeper detail is warranted, use provenance cues — what AI did, what human oversight looked like, and how to report errors — calibrated to your audience and context.
• Err toward disclosure over concealment: hidden AI use is already common and hard to monitor, with 29% of surveyed workers saying management was unaware of their AI use, per Federal Reserve economists.

The same calibration logic applies to what's underneath your AI workflow. If you're using Claude or another agent at work, the skills and tools you plug into it vary wildly in quality and security — published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory that security-scans, verifies, and benchmarks agent skills so you can choose them with the same transparency you'd want readers to extend to you.

One caveat: the disclosure findings come from a single small preprint study, so treat them as directional rather than definitive. But the principle holds across the research — silence carries risk, blanket labels get ignored, and the sweet spot is a brief, honest note matched to how much AI actually shaped the work.

Your Practical Playbook: Calibrated Disclosure and Quality-First Policies

If you can't reliably detect AI writing, and your manager can't either, the smart move isn't better hiding — it's a disclosure strategy calibrated to context. The evidence points to a middle path between blanket labels and total silence.

Here's the paradox: in a lab study, one-line AI disclosures performed almost identically to no disclosure at all in reader trust, while detailed disclosures actually reduced trust — even though roughly two-thirds of participants said they preferred detailed transparency. The researchers behind the disclosure study conclude that provenance works best when it's context-dependent: say what the AI did, what human oversight happened, and who to contact about errors. Generic labels like "created using AI" get misinterpreted or ignored.

For individuals, that translates into a simple playbook:

• Use a short, specific note when the work is public-facing: what AI contributed, that a human reviewed it, and a contact for corrections.
• Skip the stylistic gymnastics. Popular "AI tells" like em dashes and the word "delve" are unreliable, and JISC's National Centre for AI notes humans generally can't spot AI work unless they know the writer personally.
• Never use detector scores as proof — of your own work or anyone else's. Overall accuracy for leading tools ranges from 26% to 76% on mixed human and AI text, and false positive rates for free tools run as high as 30%.

For managers, the calculus is different. Assume some AI use on your team is invisible: 29% of surveyed workers said management was unaware of their AI use, and only 38% say their organization has integrated AI at all. Detection tools won't close that gap — Claude output specifically defeated most detectors in one cross-tool comparison, and education is already moving decisively away from automated detection.

The alternative is a clear, written AI-use policy that defines acceptable use, review requirements, and accountability for accuracy. Judge the output by quality — does it hold up, is it accurate, would it pass expert scrutiny — rather than hunting for tells. AI submissions have been shown to pass the scrutiny of experienced academics, so stylistic sniff tests are a losing game.

The same principle applies to the tools themselves. If you're building workflows on Claude or another agent, quality varies wildly under the surface — published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory that security-scans every skill, pins it to a commit SHA, and benchmarks it with published Lift scores, so you can judge skills by measured quality rather than vibes.

Disclosure done well isn't a liability — it's a trust signal. Keep it short, keep it specific, and spend the energy you'd waste on evasion on making the work genuinely good.

Beyond Detection: Making Your AI Usage Something You're Confident Disclosing

If you're going to disclose AI use, the goal is to do it from a position of strength — not because a detector forced your hand, but because you've verified the quality of what your AI-assisted workflow produces. That confidence is what separates disclosure from damage control.

The research on disclosure is a paradox worth understanding. In a lab study of news readers, only detailed AI disclosures reduced trust and subscription rates, while one-line disclosures performed similarly to no disclosure at all — yet roughly two-thirds of readers still preferred detailed disclosure for transparency reasons (arXiv study). The lesson isn't "hide AI use"; it's that generic labels like "created using AI" get misinterpreted or ignored, while context-dependent provenance cues — what the AI did, what human oversight occurred — build genuine credibility.

The workplace reality makes disclosure strategy urgent. Federal Reserve economists note that 29% of surveyed workers said management was not aware of their AI use, a gap they attribute partly to the "reporting and monitoring difficulties of generative AI." Meanwhile, Gallup finds 26% of U.S. workers now use AI at least a few times a week. With AI use this widespread and this invisible, organizations that set clear policies and verify quality outperform those relying on detection — especially given that leading detectors show overall accuracy ranging from 26.3% to 76% on mixed human and AI text.

So how do you build an AI workflow you can stand behind? The answer for Claude users is agent skills: SKILL.md instruction packages that systematize how your agent handles specific tasks, turning ad-hoc prompting into a repeatable, quality-controlled process. The catch is that skills vary wildly in quality — published research found roughly 26% of published skills carry at least one vulnerability. Installing an unvetted skill to "improve" your workflow can quietly introduce the exact risks disclosure is supposed to manage.

That's why verification matters before anything touches your agent. When evaluating skills, look for:

• Security scanning — since a quarter of published skills carry vulnerabilities
• Version pinning to a commit SHA, so you know exactly what code your agent runs
• Sandboxed benchmarking that shows whether a skill actually improves output
• License checks and maintenance tracking, so upstream changes don't break your stack

This is the gap SkillzBench is building to close: a free, independent directory where skills are security-scanned, commit-pinned, and benchmarked with published Lift scores, plus a free MCP connector so your agent only sees the skills you've explicitly chosen. Disclosure stops being a gamble when every component of your AI workflow has been tested and you can point to the evidence. Skills, proven. Join the SkillzBench waitlist to get security-scanned, benchmarked agent skills the moment they go live.