You are looking at something right now and something feels off. Here is what to actually check in text and images, and an honest note up front: as models improve, the tells keep getting harder to spot, and no method here is certain. HAL can look at what you are looking at and give you its honest second read, not a verdict.
Repetitive sentence structure, generic phrasing that could apply to almost any topic, oddly perfect grammar with no natural slip-ups, and a very even, list-heavy rhythm. Watch especially for confident specifics, names, dates, statistics, that sound exact but cannot actually be verified; that is a classic AI hallucination pattern.
Openers like "In today's fast-paced world," transition-heavy phrases like "It is important to note that," and tidy summarizing endings like "In conclusion" all show up far more often in AI text than in how most people actually write or talk. One of these alone means nothing; several stacked together is worth noticing.
Language models generate the most statistically plausible next words, not verified facts, so when they lack real information they can produce a specific-sounding detail, a quote, a citation, a number, that is entirely made up but reads with total confidence. This is called hallucination, and confident tone is not evidence of accuracy in AI text the way it often is in human writing.
Check hands and teeth first, extra or missing fingers, teeth that blend together oddly. Then check any text inside the image, signs, labels, and book spines often come out garbled or nonsensical. Finally look at shadows and reflections: AI images frequently get light direction inconsistent across the same scene.
Hands are complex, high-variability shapes that appear in comparatively few training images from every possible angle, so image models historically struggled to render them consistently, extra fingers, fused fingers, wrong joint counts. Newer models have gotten much better at this specific tell, which means it is becoming less reliable on its own.
Look for a shadow falling in a direction that does not match where the light source in the image actually is, or two people in the same photo with shadows pointing different ways. Also check reflections in glasses, windows, or water; AI often gets these subtly wrong even when the main subject looks convincing.
Unnatural blinking patterns, lip movements that do not quite match the audio, warping or flickering around the edges of hair and glasses, and lighting on the face that shifts oddly between frames are the classic signs. These are general technical tells, not something HAL analyzes directly, since HAL does not process video.
Yes, noticeably. Tells that reliably worked even a year ago, garbled hands, obviously robotic phrasing, often do not anymore, because model improvements specifically target the most visible failure points. This is a real arms race, and being current on the newest tells matters more than memorizing a fixed checklist.
No, not right now, for text, images, or video. The most reliable approach combines several imperfect signals: checking the tells above, checking the source and context something came from, and applying ordinary skepticism, rather than trusting one clean test.
You can share a piece of text or an image with HAL and ask for an honest read. HAL will tell you what stands out or does not, framed clearly as a second opinion from a conversation, not a certified detector score, because that certainty genuinely does not exist yet.