An AI tell is a visible or audible signal in a piece of creator content that triggers an audience to perceive it as machine-generated, whether or not it actually is, and to lower their trust accordingly. The data is brutal: enthusiasm for AI-generated creator content fell from 60% in 2023 to 26% in 2025, and 31% of consumers now trust brands less when they spot AI in marketing versus only 7% who trust them more. Hiding the tells better is not the fix.
⚡ Key Takeaways
- Enthusiasm for AI-generated creator content collapsed from 60% in 2023 to 26% in 2025, a 57% relative drop in two years (eMarketer / Jellysmack / Tubular).
- Consumers are 4.4x more likely to trust a brand less than more when they spot AI in its marketing (31% vs 7%, Klaviyo + Datalily).
- YouTube removed 16 of the top 100 "AI slop" channels between November 2025 and January 2026, killing 4.7 billion views and roughly $10M in annual revenue (Kapwing).
- The em dash is overrated as a single tell. Cadence uniformity and AI-coded vocabulary (delve, realm, intricate, in conclusion) are stronger detector signals (arXiv / FSU).
- 94% of US users believe they see AI online, but only 44% can correctly identify it. Audiences are guessing, often wrong, and penalizing you anyway.
- The fix is not better laundering. It is unfakeable human signal (face, voice, mess, on-the-ground specificity), with AI used as scaffolding only.
Why is hiding AI tells the wrong strategy in 2026?
Most "humanize your AI" guides treat the problem as a laundering job: swap the words, fix the fingers, ship the post. The audience data says that misreads the room. A March 2026 peer-reviewed study in Electronic Markets found that even labeling content as AI-assisted dropped both emotional and behavioral engagement, particularly for emotional posts. eMarketer reports that 94% of US users believe they encounter generative AI online, but only 44% can correctly identify it. Translation: audiences are guessing, often wrong, and penalizing you for the guess.
The contrarian frame is simpler. Stop trying to look human. Be human, in the parts the audience actually sees.
Red flag #1: Are em dashes really the dead giveaway everyone says?
The em dash panic is real and overrated. OpenAI quietly shipped a setting in November 2025 letting users tell ChatGPT to stop using them, per Washington Post reporting. That alone confirmed the punctuation had become a reputational liability for AI-generated text.
The fix is not to ban one character and call it done. It is to write the way you actually talk, with commas, periods, and the occasional parenthesis. Fanvault's content team treats the em dash as a hard ban across every published artifact for exactly this reason.
Red flag #2: Which words instantly mark text as AI-written?
An FSU computational linguistics study accepted to COLING 2025 identified 21 "focal words" whose usage exploded after ChatGPT's release. Readers reacted significantly more negatively to delve than to any other AI-coded buzzword. A separate Kobak et al. analysis estimated at least 10% of 2024 PubMed abstracts were processed with LLMs based on style fingerprints alone.
The 2026 watch list:
- Delve, realm, intricate, commendable, the four-token AI smell test.
- Furthermore, moreover, in conclusion, transition words humans almost never type.
- Navigate the landscape, unlock the potential, in today's fast-paced world, copy-paste phrases pulled straight from training data.
Search your draft. Cut every hit. Replace with the word you would actually say out loud.
Red flag #3: What does AI cadence sound like to a reader?
Detector software measures something writers feel: burstiness. AI text clusters sentences in the 18 to 24 word range with near-uniform structure. Human writing varies. Three words. Then a long, looping aside that doubles back on itself. Then a punchline.
The fix is brutal but cheap: read every paragraph aloud. If three sentences in a row hit the same length, break one in half and stretch another.
Red flag #4: Which image artifacts still flag AI in seconds?
Hands are no longer the bulletproof tell they were in 2023. Midjourney and DALL-E have largely closed that gap. The long tail still flags AI to most viewers instantly:
- Fused fingers, extra knuckles, or watch bands that vanish under a wrist.
- Backpack straps that merge into a jacket or melt into a chair.
- Lighting that does not match across faces, props, or shadows.
- Text on signs and packaging that reads as garbled letterforms.
If you cannot ship the image without a viewer wondering "wait, what is that," reshoot it or redraw the prompt.
Red flag #5: Why is YouTube deleting templated AI video formats?
On July 15, 2025 YouTube renamed its "repetitious content" policy to "inauthentic content" and clarified that mass-produced or templated AI work is ineligible for the Partner Program. By January 2026, Kapwing documented 16 of the top 100 "AI slop" channels removed, totaling 4.7 billion views, 35 million subscribers, and roughly $10 million in annual revenue. CuentosFacianantes (5.9M subs, 1.2B views) and Imperio de Jesus (5.8M subs) were two of the casualties.
The pattern flagged: identical thumbnails, AI voiceover plus stock-image slideshow, daily template uploads, no on-camera presence. If your format matches that description, the enforcement risk is now monetization-ending.
Red flag #6: When does an AI voice become a lawsuit?
In 2025 former NPR host David Greene sued Google over NotebookLM after friends asked if he had licensed his voice. Forensic analysis returned a 53-60% confidence match, per Gizmodo. The Lovo and ElevenLabs voice-actor suits followed similar fact patterns, with actors discovering AI clones of their own recordings in third-party podcasts.
The rule is now simple. If you did not license the voice in writing, do not clone it. If you cloned your own voice, disclose it on the post. Undisclosed cloning is litigation risk, not an ethics debate.
Red flag #7: Is polish itself becoming a red flag?
The Coca-Cola 2025 holiday campaign is the canonical example. Heavy generative AI, no single obvious error, and still public backlash for feeling uncanny, per Digiday. The same outlet reports the share of US and UK consumers who say AI is negatively disrupting the creator economy jumped from 18% in 2023 to 32% in July 2025. Glossy without mess now reads as synthetic.
What should creators do instead?
The honest playbook is the inverse of the laundering guides. Inject signal the audience reads as unfakeable: a face on camera, your real voice with its actual quirks, a date stamp, a location stamp, a half-second of mess. Use AI as scaffolding for research, outlines, and first drafts. Ship something that could not have been generated.
| Red flag | Audience reads it as | What to do instead |
|---|---|---|
| Em dash overuse | AI default punctuation | Comma, period, or parenthesis |
| Delve, realm, intricate | ChatGPT vocabulary | Write the word you would say aloud |
| Uniform sentence length | Low burstiness, machine cadence | Read aloud, break and stretch |
| Image artifacts | Fused fingers, melted props | Reshoot or redraw the prompt |
| Templated video | YPP-ineligible "inauthentic content" | Original structure, real voice, on-camera |
| Undisclosed voice clone | Litigation risk and audience betrayal | License in writing or disclose |
| Glossy with no mess | Uncanny, synthetic | Date stamp, location, half-second of mess |
The economics reward this directly. On Fanvault, the platform fee is 8% and creators keep 92%. Every percentage point of trust earned converts to subscription retention and storefront conversion at full creator margin. The trust premium is no longer a soft metric. It is the metric.
Frequently Asked Questions
What is the single biggest AI tell in text right now?
Cadence uniformity, not the em dash. Detectors measure burstiness (sentence-length variation), and AI text clusters in the 18-to-24-word range with near-identical structure. Readers feel it before they can articulate why. The em dash gets the headlines (OpenAI now lets users suppress it in ChatGPT, per Washington Post), but it is one of many signals. Mix sentence lengths, break the third sentence in half, drop in a fragment, and both the detector and the reader relax.
Should I disclose that I used AI to help write or produce something?
For voice cloning, almost always yes. The David Greene v. Google NotebookLM suit and the Lovo and ElevenLabs voice-actor cases turned undisclosed cloning into active litigation territory, per Inc. For text, it is trickier. A March 2026 peer-reviewed study in Electronic Markets found that explicitly labeling content as AI-generated reduced both affective and behavioral engagement, particularly for emotional content. The honest answer is to use AI as scaffolding (research, outlines, first drafts) and ship work the audience reads as yours.
Is faceless YouTube content done for in 2026?
Templated faceless content is now monetization risk. YouTube renamed its "repetitious content" policy to "inauthentic content" on July 15, 2025, and Kapwing documented 16 of the top 100 AI slop channels removed by January 2026, totaling
Why does perfectly polished content read as AI even when it is not?
Because the audience reaction is not really about detection accuracy. It is pattern matching.
Where does Fanvault stand on AI content?
Fanvault is a creator monetization platform with an
