How to Spot AI-Generated Content (Images, Text, and Voices)

AI detectors look for statistical fingerprints of generated text — and those fingerprints fade with every model generation, while human writing is too varied to score cleanly. The result: detectors that miss modern AI text and flag real students and non-native speakers. Courts, universities, and platforms have learned the same lesson.

The reliable approach is content-level verification: the tells below work because they check whether the world described actually coheres — not whether the words pattern like a machine. Pair them with reverse search and source-checking, and you'll catch the overwhelming majority of fakes.

Generate-ready models improved enormously, but composites still leak in the details. Scan in this order:

• Hands and feet: finger count, joints, jewelry that merges into skin. Still the classic tell in mid-tier generations.
• Text in the image: signs, labels, book spines, phone screens — garbled glyphs, near-words, mirror letters.
• Backgrounds: people who melt into objects, glasses that hold no liquid, shadows pointing different directions, architecture that doesn't quite join up.
• Accessories: eyeglass frames that double, earrings at different heights, logos that are almost brands.
• Skin and teeth: porcelain smoothness at close crop, tooth counts that drift between frames in a 'video'.

Then the metadata check: reverse-image search first (Google Lens, TinEye). Real photos have histories; generated ones appear from nowhere. C2PA content credentials (newer cameras and Photoshop embed them) are the emerging provenance standard — check for them where present.

Forget 'sounds robotic' — modern AI writes well. The reliable tells are structural:

• Perfect-but-empty: flawless grammar, no typos, no distinct voice, and somehow no information density. Human experts make small, revealing errors and hold opinions.
• Hedging walls: every claim softened ('it's important to note', 'can vary', 'many factors') — a style trained to be inoffensive.
• Fabricated precision: statistics, citations, or quotes that don't resolve when you check them. Click every citation; invented references remain the hard tell.
• Listicle gravity: everything becomes a tidy numbered structure regardless of whether the content wants one.
• Time-defying details: references to events slightly off-date, products that don't exist, or a 'current' claim that misses what changed last month.

Voice cloning is now caller-ID-quality; video deepfakes are improving yearly. The defenses that hold:

• The interaction test: live-video scam attempts fail when you ask the person to turn their head, wave a hand near their face, or answer something only-the-real-them knows. Rendering lags interaction.
• Audio tells in clones: flat emotional range under pressure, breathing/room tone that never varies, odd pauses where emphasis should be, and word-level artifacts in names and numbers.
• Channel verification: the family 'emergency' call gets returned to the known number; the boss's wire request gets a call to the number on file. The safe-word for families remains undefeated.
• Context: urgency + secrecy + money is the scam shape regardless of how real the face looks — our phishing guide's two-question check applies to video and voice identically.

One habit catches most fakes, whichever the medium: triangulate. One source is zero sources.

1. Reverse-search images and key claims (Google Lens, TinEye, search the exact quote).
2. Find a second independent source for events — not a repost chain.
3. Check provenance: account age, posting history, C2PA credentials when present.
4. Apply the pressure test: fakes push action (share, pay, panic); truth tolerates a pause. Urgency is the tell beneath all tells.

These skills pair naturally with our AI verification habits — same discipline whether the content arrives by search, email, or your aunt's group chat.