How to Spot AI-Generated Content

No single visual "tell" will save you anymore, so start with the two things that actually scale: provenance metadata (C2PA Content Credentials and Google's SynthID watermark) and old-fashioned source-checking. Treat the finger-counting and vocabulary tricks below as tie-breakers, not verdicts, because today's models have fixed most of the obvious glitches and the rest are disappearing month by month.

Check for provenance before you squint at pixels#

The most durable signal is not inside the content, it is attached to it. Two systems are worth knowing by name.

C2PA Content Credentials are a cryptographically signed manifest recording how a file was created and edited. Adobe, Microsoft, OpenAI, Leica, and others attach them. To read one, drag the image onto verify.contentauthenticity.org, or click the small "Cr" icon that some platforms now show. A credential reading "AI generated (DALL-E 3)" is strong evidence. The absence of one proves nothing, though, because screenshots, re-saves, and most social platforms strip the metadata on upload.

SynthID is Google DeepMind's invisible watermark, embedded in the pixels themselves (and in audio, video, and Gemini's text) rather than the metadata, so it survives cropping, compression, and moderate edits. Images from Google's Imagen, plus a growing share of other generators, increasingly carry it, and Google runs a SynthID Detector portal, though public access is still limited.

Neither is a magic wand. Most content in the wild carries no watermark at all, and a missing one is not a clean bill of health. But when a credential is present, it beats every gut-feeling heuristic below.

Reading AI-generated text#

Large language models write fluent, grammatically spotless prose that is oddly averaged-out, the statistical middle of everything they were trained on. That flatness is the real tell, and it surfaces in specific ways.

The words and rhythms that recur#

• Signature vocabulary. "Delve," "tapestry," "testament to," "navigating the complexities," "it's important to note," "in today's fast-paced world," "boasts," "underscores," "landscape" used as a metaphor. One of these means nothing; a paragraph stacked with them is a pattern.
• The rule of three, everywhere. Models love triads: "clear, concise, and compelling." When every list runs to three items and every sentence balances two clauses against a third, you are reading a rhythm, not a mind.
• Both-sides hedged endings. A conclusion that "ultimately depends on your needs" and reminds you to "weigh the pros and cons" is refusing to hold an opinion, which is a very LLM move.
• Confident, unsourced authority. "Studies show" and "experts agree," with no study and no expert named. AI is superb at sounding informed and weak at being checkably right, which is why it states wrong facts in flawless sentences, the failure known as hallucination.

Human writing leaves fingerprints that AI sands off: a weirdly specific example, a real stake in the argument, an unexpected joke, a number that could turn out wrong. The absence of specific, verifiable detail is more telling than any single buzzword.

Why detection tools are not proof#

Do not paste text into GPTZero, Originality.ai, or Turnitin and call the matter settled. These tools estimate statistical "perplexity" and "burstiness," and they are wrong often, in both directions. A widely cited 2023 Stanford study found that popular detectors disproportionately flagged writing by non-native English speakers as AI, a real and unfair harm. OpenAI itself quietly shut down its own AI-text classifier in 2023 for low accuracy. Use a detector as one weak input, never as a gavel.

Inspecting images#

The 2022-era giveaways, like six-fingered hands and melting ears, are largely gone from Midjourney v6, DALL-E 3, and Adobe Firefly output. Look instead at the parts these models still fumble.

• Text inside the scene. Signs, labels, book spines, and license plates still dissolve into plausible-looking gibberish, near-letters that spell nothing. This is the single most reliable visual tell left.
• The physics of light. Reflections in eyes, mirrors, and windows that do not match the scene; shadows falling from inconsistent directions; catchlights that differ between the two eyes.
• Symmetry that should not be there. Mismatched earrings, glasses arms that change thickness, teeth that are unnaturally uniform, patterned fabric that fails to line up across a seam.
• Backgrounds and crowds. Faces that blur to mush at a distance, repeating textures, architecture that bends where it should run straight.

Then leave the image itself. Run it through Google Lens, TinEye, Yandex, or Bing Visual Search. If a "breaking news" photo turns out to have appeared two years ago in an unrelated context, you have your answer. Check the EXIF metadata too: a genuine photo usually carries a camera model and timestamp, while a bare file with none is at least worth a second look.

Video and cloned voices#

Video generators such as OpenAI's Sora, Runway Gen-3, and Google's Veo fail mostly on continuity over time. Watch for objects that morph or pop in and out, textures that shimmer frame to frame, hair and finger boundaries that flicker, and physics that quietly breaks, like a poured liquid that never settles or a walk that never quite lands. Face-swap deepfakes betray themselves at the jaw and hairline, in skin tone that does not match the neck, and in lip movements running a half-beat off the audio.

Cloned voices are the most dangerous, because they are the cheapest. Tools like ElevenLabs can imitate a voice from a minute or two of sample audio. Synthetic speech tends to skip breaths, hold unnaturally even pacing, and sit on suspiciously clean silence. Here, context is your best defense: a "relative" phoning in a panic and asking for money or gift cards is the classic cloned-voice scam. Hang up and call the person back on their known number, or ask something only they would know. Verify through a second, trusted channel before you act, every time.

The mistakes that trip people up#

• Treating a detector score as proof. It is an estimate, and a biased one, not evidence you can accuse someone with.
• The em-dash panic. Long dashes, "however," and clean grammar are not AI signatures. Plenty of careful humans write that way, and branding their work fake is its own kind of harm.
• Assuming no watermark means human. Most real and fake content is unwatermarked, so absence tells you nothing.
• One clue, instant verdict. Any single tell can be coincidence. Confidence should come from several independent signals pointing the same way, plus a source you can actually trace.

FAQ#

Are AI content detectors accurate?#

Not reliably. Text detectors like GPTZero and Turnitin produce both false positives and false negatives, and research has shown they unfairly flag non-native English writers. Treat any score as a weak hint, never as evidence strong enough to accuse someone or drive a high-stakes decision.

What is the single most reliable tell?#

For images, it is garbled text on signs and labels. For writing, it is the absence of specific, checkable detail behind confident claims. But the most durable signal is provenance, meaning a C2PA Content Credential or a SynthID watermark, combined with tracing the item back to a trustworthy original source.

Can I remove a SynthID or C2PA watermark?#

C2PA metadata is easy to strip; a simple screenshot removes it, which is exactly why its absence proves nothing. SynthID is embedded in the pixels or audio and is designed to survive cropping and compression, although heavy editing can degrade it. Neither is foolproof, and neither is present on the majority of content you will encounter.

How do I protect myself from voice-cloning scams?#

Agree on a family "safe word" for genuine emergencies. If a call or voice message demands money or gift cards urgently, hang up and call back on a number you already have saved. Verify through a second channel, and remember that urgency plus an unusual payment method is the scam's core pattern, not a coincidence.