Is This AI? How to Tell if Text, Images or Video Are AI-Generated (2026)

There is no single reliable way to prove whether something was created by AI. For images and video, the most trustworthy signals are provenance data (C2PA Content Credentials and watermarks such as Google's SynthID), plus checking the source and looking for visual inconsistencies. For text, AI detectors are unreliable and produce false positives, especially for non-native English writers, so they should never be the only evidence. The best approach combines provenance, source verification, context and human judgement.

"Is this AI?" is now one of the most-asked questions online, whether about a suspicious photo, a viral video, a student essay, a product review or a voice message from a "family member" asking for money. This guide explains how to check each type of content, how AI detectors work (and why they fail), the watermark and provenance technologies that are actually becoming useful, and what to do when you can't be sure.

Why it's hard to tell

Generative AI has improved to the point where its output is often indistinguishable from human-created content at a glance:

• Text from modern language models is fluent, varied and can mimic any style.
• Images from current models rarely have the obvious flaws (extra fingers, garbled text) of early generators.
• Video generation now produces realistic motion, lighting and sound.
• Voice cloning can reproduce a person's voice from a short sample.

And humans increasingly edit AI output, while AI increasingly edits human content. Much content is now a mix, so "AI or human?" is often the wrong question. "Is this authentic, accurate and from who it claims to be?" is usually the right one.

How to check if an image is AI-generated

1. Check for Content Credentials (C2PA)

The C2PA (Coalition for Content Provenance and Authenticity) standard, backed by Adobe, Google, Microsoft, OpenAI, camera makers and media organisations, attaches tamper-evident metadata to images and videos describing how they were made and edited. Many AI image tools (Adobe Firefly, OpenAI's image generation, Google's tools) add Content Credentials, and some cameras and phones sign authentic photos at capture.

How to check: upload the image to the Content Credentials Verify tool (contentcredentials.org/verify) or look for the "CR" icon on platforms that display it.

Limitation: metadata is often stripped when images are screenshotted, re-saved or uploaded to some platforms. Missing credentials don't mean an image is AI-generated; they just mean provenance is unknown.

2. Check for invisible watermarks

SynthID, developed by Google DeepMind, embeds an imperceptible watermark in images, audio, video and text generated by Google's AI tools, designed to survive common edits like cropping and compression. Google offers a SynthID Detector, and the technology has been adopted beyond Google's own products. Other providers use their own watermarking. Limitation: each watermark is only detectable for content from tools that use it.

3. Do a reverse image search

Use Google Lens, TinEye or Bing Visual Search to find where the image first appeared. Real news photos usually trace back to reputable sources, photographers or agencies; AI images often first appear on social accounts, AI art communities or with no credible origin.

4. Look for visual clues (less reliable than before)

• Hands, teeth, ears and jewellery that look odd on close inspection
• Text and signs that are garbled or nonsensical
• Inconsistent lighting, shadows and reflections
• Background details that melt into each other or don't make physical sense
• Overly smooth, "airbrushed" skin and perfect symmetry
• Patterns (fences, tiles, fabric) that break or repeat unnaturally

Treat these as hints, not proof: modern models avoid many of them, and real photos can look strange too.

5. Consider the context

Who shared it? Is it too perfect, too dramatic or too convenient for a particular narrative? Is any reputable source reporting the same event?

How to check if a video or audio is AI-generated (deepfakes)

• Check the source and provenance (Content Credentials, original uploader, reputable coverage).
• Look for visual artifacts: unnatural blinking, lip movements slightly out of sync, flickering at the edges of faces, inconsistent lighting, odd hand and teeth details.
• Listen for audio clues: flat emotion, odd breathing, unnatural pauses, mispronounced names, background noise that doesn't match the setting.
• Verify through another channel. This is the most important step for scams. If a "family member", "CEO" or "bank" calls or sends a voice note asking for money or sensitive information, hang up and call them back on a number you know. Many families and companies now agree on a code word for emergencies.
• Search for fact-checks from established fact-checking organisations.

Deepfake fraud is a real and growing threat; there have been publicly reported cases of companies losing large sums after employees were deceived on video calls with deepfaked executives.

How to check if text is AI-generated

How AI text detectors work

AI text detectors (such as GPTZero, Turnitin's AI detection, Copyleaks, Originality.ai and others) analyse statistical properties of text:

• Perplexity: how predictable the words are to a language model. AI text tends to be more predictable.
• Burstiness: how much sentence length and structure vary. Human writing tends to be more uneven.
• Classifier models trained on examples of human and AI text.

Why AI text detectors are unreliable

• False positives: detectors wrongly flag human writing as AI. A widely cited 2023 Stanford-led study published in Patterns found that several popular detectors misclassified more than half of essays written by non-native English speakers as AI-generated, while correctly classifying essays by native-speaking US students.
• Bias: clear, simple, formal or formulaic writing (common among non-native speakers, technical writers and students taught to follow templates) looks "predictable" and gets flagged.
• Easy to evade: light editing, paraphrasing or prompting the AI to vary its style often defeats detectors.
• Mixed content: most real writing now involves some AI assistance (grammar fixes, rephrasing, outlining), which detectors can't meaningfully classify.
• No ground truth: detectors output probabilities, not proof. OpenAI withdrew its own AI text classifier in 2023, citing its low accuracy.

Bottom line: an AI detector score alone should never be used to accuse someone of cheating or to make consequential decisions.

Text watermarking

Some providers embed statistical watermarks in generated text; Google's SynthID Text is used in Gemini and has been open-sourced. Watermarks can be detected for text from participating models, but they can be weakened by heavy editing or paraphrasing, and many models don't watermark text at all.

Signs that text may be AI-written (hints, not proof)

• Generic, polished prose with little personal experience, specific detail or original insight
• Overused phrases and structures (lists of three, "it's important to note", formulaic conclusions)
• Confident claims with fabricated or incorrect citations, statistics or quotes
• Content that doesn't match the writer's usual voice, knowledge level or previous work
• Answers that address the prompt generally but miss its specifics

Better approaches for educators and editors

• Ask about the process: drafts, notes, version history (Google Docs and Word show editing history).
• Have a conversation: ask the writer to explain their argument or choices.
• Design assessments that require personal experience, in-class work, oral components or iterative drafts.
• Set clear AI-use policies so students and contributors know what's allowed.
• Check facts and sources: fabricated references are a strong red flag regardless of who wrote the text.

How to check AI-generated reviews, profiles and messages

• Reviews: look for clusters of similar wording, posting at the same time, generic praise without specifics, and reviewer profiles with little history.
• Social media profiles: AI-generated profile photos, recently created accounts, generic bios and high posting volume.
• Emails and messages: AI has made phishing more polished, so don't rely on spelling mistakes to spot scams. Check sender addresses, links and urgency, and verify unusual requests through another channel.

What's changing: regulation and standards

• EU AI Act: transparency obligations require that people are told when they're interacting with an AI system and that certain AI-generated or manipulated content (including deepfakes) is labelled; these obligations apply from August 2026.
• Platform labelling: major social platforms label some AI-generated content using C2PA data, watermarks and creator disclosures.
• Industry adoption of C2PA continues across camera makers, news organisations and AI providers.

Provenance (proving where authentic content came from) is increasingly seen as more scalable than detection (proving something is fake).

A practical checklist: "Is this AI?"

1. Who created and shared it? Is the source credible and verifiable?
2. Does it have Content Credentials or a detectable watermark?
3. Can you trace it back with a reverse image search or by finding the original?
4. Do reputable sources corroborate it?
5. Are there visual, audio or factual inconsistencies?
6. Is it asking you to act urgently (send money, share data, click a link)? If so, verify through another channel.
7. For text, can the author explain it? Don't rely on detector scores alone.

Frequently asked questions

How can I tell if something is AI-generated?

Check its source and provenance (Content Credentials, watermarks such as SynthID, reverse image search), look for visual, audio or factual inconsistencies, verify with reputable sources, and, for urgent requests, confirm through a separate channel. No single test is fully reliable.

Are AI detectors accurate?

AI text detectors are not reliably accurate. They produce false positives, especially for non-native English writers and formal writing, and can be evaded with light editing. They should never be the only evidence for accusations or decisions.

How do I check if text is AI-generated?

Look for generic content lacking specific detail, formulaic phrasing and fabricated facts or citations; compare against the author's previous work; review drafts and version history; and talk to the author about their process. Treat detector scores as weak signals only.

How can I tell if an image is AI-generated?

Check for C2PA Content Credentials, use SynthID or other watermark checks where available, run a reverse image search, examine details like hands, text, lighting and backgrounds, and consider who shared it and whether reputable sources confirm it.

What is SynthID?

SynthID is a technology from Google DeepMind that embeds imperceptible watermarks in AI-generated images, audio, video and text so they can later be identified as AI-generated, even after common edits.

What are Content Credentials (C2PA)?

Content Credentials are tamper-evident metadata, based on the C2PA open standard, that record how a piece of media was created and edited, including whether AI was used. They can be checked with verification tools, though metadata can be removed when files are re-saved or screenshotted.