As of 2026-05-18. Detection is a moving target — model capabilities improve every few months. The article points to durable signals (provenance, context, source) over telltale-of-the-month. For current detection tool status, see thepromptbench.com.
Why this is hard
A few years ago, AI-generated content was easy to spot: stiff text, extra fingers in images, eerie voices. Today's models produce text indistinguishable from competent human writing, images that pass casual inspection, voices that sound right, and short videos that look real.
The arms race is asymmetric. Generators improve faster than detectors can keep up. Each new model fixes the previous generation's telltales. And once a "tell" is publicly known, generators specifically train against it.
The honest message: you can't reliably detect AI content from the content alone. What you can do is look for specific patterns that still show up, use detection tools knowing their limits, and — most importantly — develop habits around source, context, and provenance.
Telltales in text
Text is the hardest case because well-edited AI output is indistinguishable from human writing. But there are still patterns:
Generic structure. AI loves "First, ... Second, ... Third, ..." and "In conclusion." Heavy use of obvious transitions ("Moreover," "Furthermore," "Additionally") suggests AI or a writer trained to imitate it.
Vague specificity. AI tends to say "many experts," "various studies," "a recent report" without naming them. Real writing tends to either name specific sources or admit the lack of citation.
Polished but generic. AI output is grammatically clean and structurally tidy but often lacks the personal detail, idiosyncratic word choices, or specific anecdotes a human writer would include. If a piece reads smoothly but you can't recall a single specific point afterwards, that's a hint.
Stock phrases. "In the rapidly evolving landscape of..." "It's important to note that..." "When it comes to..." "Delve into" became a meme in 2024 partly because ChatGPT used it so often. These phrases are not exclusive to AI but their frequency is.
Lists that don't quite match. AI sometimes lists items that overlap, or includes one item that doesn't quite fit, suggesting the list was generated by an autocomplete process rather than a deliberate selection.
Confident but unverifiable claims. "Studies show..." "Research indicates..." without specifics. Or specific-sounding numbers ("a 2019 study found 73%") that turn out to be fabricated when you check.
No personal stakes. Real writing usually has a point of view or stakes — the writer believes something, cares about something. AI writing often reads from nowhere in particular, balanced to inoffensiveness.
Hedging everywhere. "While it can be argued that... it's also worth considering that..." A hedge in every paragraph suggests AI being trained to avoid controversy.
None of these is conclusive. A human writer might use any of them. But a piece that hits 4-5 of these is more likely AI-influenced.
Detection tools for text
Several products claim to detect AI-generated text: GPTZero, Originality.ai, Copyleaks, Turnitin's AI detector, Hive AI detection, and others.
The state of these tools through 2024-2025: many major institutions restricted or cautioned against using AI-detection scores as primary evidence, though policy varies by institution. The problems:
- High false-positive rates. They flag human-written text as AI-generated regularly. Non-native English speakers, simple direct writing styles, and academic prose all get flagged. Several universities (including Vanderbilt and Michigan State) suspended use of Turnitin's AI detection after high-profile false-positive incidents, and OpenAI shut down its own AI Text Classifier in 2023 citing low accuracy.
- Easy to defeat. Light editing (a few rephrased sentences, some personal anecdotes added) dramatically lowers detection scores.
- Worse on shorter text. Most tools need 300+ words to be even slightly reliable. Tweet-length AI text is essentially undetectable.
- Mixed text confuses them. Human writing edited with AI assistance, or AI writing heavily edited by humans, gets ambiguous scores that mean little.
Reasonable use: detection scores can be one signal in a broader investigation (does this person typically write this way? is this similar to other AI patterns I've seen? does the content match what a human source could plausibly know?). They're not standalone proof.
Telltales in images
AI images have improved dramatically. Older artefacts:
- Extra or missing fingers
- Asymmetric eyes or eyeglasses
- Wrong number of teeth
- Background text that's gibberish
- Inconsistent shadow directions
- Floating earrings
- Mismatched hands
These have been largely reduced (though not eliminated) in current frontier image models — Stable Diffusion XL (released 2023), Adobe Firefly 3 (announced 2024), recent Midjourney releases, OpenAI's DALL-E/image models. Modern AI images often pass casual inspection, especially at typical web display sizes.
What still shows up:
Hands. Still hard. Especially when hands are interacting with objects (holding a phone, gripping a steering wheel) or each other.
Text within images. Words on signs, t-shirts, book covers, name tags often look almost-right but with subtle errors. Some newer models handle this better, but it's still a weak spot.
Hair behavior. Hair often looks "painted on" — strands don't behave with realistic physics, especially around the edges.
Eyes and pupils. Look closely at the eyes. Pupils should be circular and reflect a consistent light source. Some AI images have asymmetric pupils or inconsistent eye reflections.
Symmetry where there shouldn't be. Faces are slightly asymmetric in nature. AI-generated faces are sometimes too symmetric.
Backgrounds that don't make architectural sense. Buildings with impossible geometry, room corners that don't meet correctly, sky that doesn't match the lighting.
Smooth uniformity. Skin too clean, lighting too even, surfaces too pristine. Real photography has noise, dust, imperfections.
Reverse image search. Search the image on Google Images, Bing Image Search, TinEye. If it's stock art or stolen content, you may find the original. If nothing similar exists anywhere on the web, that's slightly suspicious.
Metadata. Many AI image generators tag their outputs in metadata (with the C2PA standard) or via specific identifiers in EXIF data. Tools like Adobe Verify can read these tags.
Telltales in video and audio
Video and audio "deepfakes" are improving but still have telltales:
Video — face: Mouth movements that don't quite match speech (especially for non-English languages or unusual sounds). Subtle facial muscle twitches that look unnatural. Blinking patterns that are too regular or too rare. Skin texture that smoothes oddly when the face moves.
Video — body and environment: Lighting on the face that doesn't match the environment lighting. Edges of the face/hair that "shimmer" against the background. Camera artefacts that don't match what a real camera would produce.
Audio — voice: Subtle background hiss that's too uniform (real recordings have variable background noise). Breathing patterns that are absent or too regular. Mouth sounds (lip smacks, swallows) absent. Tonal range that's slightly too flat. Reverberation that doesn't match the supposed environment.
Audio — calls/scams: A phone call from a "loved one" asking for money urgently is the most common deepfake-audio scam. Set a family code word — a phrase that only your family knows — to confirm identity on stressful calls. Always verify by calling back on a known number.
The actual best practice: care about provenance
Trying to detect AI from the content alone is increasingly futile. The more productive question is: where did this come from?
Source. Was the content published by a known organization with editorial standards (a newspaper, a research institution, a verified social media account) or by an anonymous source? Source quality matters more than detection score.
Chain of custody. Was the original captured by a known camera or device? Has it been published in multiple places that would have caught a fake? Has anyone close to the supposed subject verified or denied it?
Plausibility. Does the content fit other facts you know? Is the supposed speaker doing or saying something they'd plausibly do or say? Extraordinary claims need extraordinary evidence; an out-of-character statement is more likely fabricated.
Cross-check. If this is a major news event, are other independent outlets reporting it? Major events leak across many sources; a story that only appears in one place is suspect.
C2PA / Content Credentials. The C2PA standard (backed by Adobe, Microsoft, Google, Sony, BBC, and others) is an open framework for cryptographically signed provenance metadata. As adoption grows, you can check whether an image, video, or document carries a credible signed history of its creation and edits. Some camera models support C2PA-signed captures — the Leica M11-P (2023) was the first to ship with Content Credentials, and Sony added C2PA support to the Alpha 9 III via firmware. Adoption is growing but not yet on by default in most consumer cameras. OpenAI and Adobe tag AI-generated images from their generators.
The standard is at https://c2pa.org/. Look for "Content Credentials" UI in image viewers and social platforms — Instagram, LinkedIn, and several news sites have begun displaying them when present. Not universal yet, but heading in a good direction.
Common scams and how to spot them
A few specific patterns to be aware of:
The fake-celebrity endorsement. A deepfake video of a celebrity endorsing a crypto scheme or "miracle" supplement. Always traceable to a paid ad network or a sketchy domain. Real endorsements would be on the celebrity's official accounts and major media.
The urgent voice call. "I'm in trouble, send money." Use a family code word. Hang up and call them back at their normal number.
The AI-written news article. Especially on niche topics, low-traffic news sites publish AI-generated articles to capture search traffic. Check the site's history, the author bylines, the citations. Quality news writing has specific sources and traceable claims.
The AI-generated review. Product reviews increasingly include AI-generated content. Look for reviews with specific details (model number, room size, install date) over generic praise. Check reviewer profile history.
The AI-impersonation social media account. Accounts mimicking a real person to scam followers. Verify against the real person's verified accounts. New accounts with stolen profile photos are suspicious.
The fake research paper or "study". AI-generated papers exist on preprint servers and even some predatory journals. Real research has named authors with institutional affiliations, citations that exist, and methodological detail.
Personal practices
A few habits that help:
Slow down on emotional content. Outrage, fear, and shock are emotions adversaries use to bypass scepticism. If something makes you immediately want to share it, that's the time to verify.
Develop sources you trust. Build a list of reliable outlets (newspapers, scientists, experts, journalists) whose content you trust because they've been right over time. Default-trust them more than random social media.
Reverse-search images. Easy and fast. Right-click → "Search image" in most browsers. Often shows the original or earlier uses.
Ask the supposed speaker. If a friend "says" something shocking via text or audio, message them through a different channel.
Be sceptical of perfection. Real photos have imperfections. Real writing has personal voice. AI content tends toward smoothness; that smoothness itself is a hint.
If you'd like a guided 5-minute course on AI media literacy, NerdSip can generate one.
The takeaway
AI-generated content is increasingly hard to detect from content alone. Text telltales (generic structure, vague specificity, stock phrases) and image telltales (hands, text-in-images, hair, symmetry) still exist but get fixed with each new model generation. Detection tools (GPTZero and similar) are unreliable enough that institutions have stopped trusting them as proof. The durable answer is provenance: care about where content came from, the source's reputation, cross-checks, and increasingly the C2PA content-credentials standard. Skip the "spot the AI" game when you can — verify the source instead.