A shocking photo in your feed. A product review that reads slightly off. A news article that is unusually smooth. Which of these were written by a person? As generative tools became ordinary, reading with suspicion turned into a basic literacy skill.
This guide has two halves. The first covers visible tells. The second covers verification, which matters far more — because tells keep disappearing while verification keeps working.
Signals in AI-written text
- Uniform smoothness: almost no grammatical errors and even sentence rhythm, but no specific lived experience or anecdote.
- Both-sides endings: repeated conclusions like balance is important or there are pros and cons, which commit to nothing.
- Volume without specifics: long text that is oddly short on dates, place names, named people, or checkable figures.
- Inflated register: phrases like a new frontier or a paradigm shift appearing far more often than the content warrants.
- Structural sameness: many articles on one site sharing near-identical length, section count and tone.
These are circumstantial. Humans write this way too, and AI drafts edited by a person are essentially undetectable. That is why the next section matters more.
Check the source, not the prose
A far more reliable test than analyzing style is asking who published this and on what basis.
- Does the author have a real name and a verifiable background?
- Does the site disclose who operates it, with a contact route?
- Do claims link to checkable sources?
- Is there a record of corrections? Outlets that publish corrections are the ones worth trusting.
Signals in AI-generated images
- Hands and lettering: finger counts, how objects are gripped, and text on background signage remain frequent failure points.
- Physical inconsistency: shadows falling in different directions, eyeglass arms passing through an ear, a railing that stops mid-span.
- Skin and lighting: poreless, porcelain-smooth skin under lighting that exists nowhere in particular.
- Background crowds: faces that dissolve and body proportions that drift the further back you look.
Caveat. Generation quality improves every year and these tells shrink accordingly. Their absence proves nothing.
Three habits that beat tell-spotting
Reverse image search
Upload the picture to an image search tool and look for its origin. A genuine news photo traces back to an outlet; a generated one has no upstream source or surfaces in generation communities. This takes ten seconds and is the single most effective check available.
Cross-reference
The more shocking the claim, the more you should confirm other outlets carry it. Major events are rarely reported by exactly one source. If a search returns many pages repeating identical sentences, that is one source copied widely, not independent confirmation.
Check your own reaction
Content engineered to spread runs on outrage and fear. Inserting one search between feeling angry and pressing share filters out most manipulation on its own.
Where labeling standards stand
Major AI companies are embedding invisible watermarks and content credentials into generated media, and several jurisdictions are moving toward disclosure requirements.
These help, but watermarks survive poorly through screenshots, editing and recompression, and disclosure rules only bind those who follow them. The last line of defense is still the reader.
The short version
AI content is not inherently a problem. Content designed to deceive is. Check the source, reverse-search the image, confirm elsewhere — and when you are not sure, do not share it.
작성자 소개
