How to spot AI text, images and video in 2026 — a field guide that checks sources, not just pixels

The uncomfortable truth first: your eyes are no longer a reliable detector. Modern AI models produce faces, voices and paragraphs that fool most people most of the time. The tools meant to catch them are also far weaker than their marketing suggests. So in 2026, “spotting AI” is less about superhuman vision and more about knowing what to check, in what order, and what the law now requires.

Why “I can feel it” stopped working

AI text detectors have a serious credibility problem. OpenAI quietly retired its own official AI text classifier in 2023 because its accuracy was simply too low. Around the same time, Stanford researchers found that popular detectors consistently flagged writing by non-native English speakers as AI-generated. That finding matters deeply in the EU, where English is often a second or third language — for students, job applicants and employees.

Image and video detectors fare no better. A detector score is not evidence. It is, at best, a reason to look closer. The moment we treat it as proof, we get false accusations, damaged reputations and legal trouble under GDPR.

Text: check the process, not just the prose

AI-written text in 2026 rarely looks robotic. The old clichés — constant “delve”, “landscape”, “it is important to note” — are now the first thing both writers and models avoid. What still stands out is a certain artificial smoothness: every paragraph flows neatly into the next, every argument is calmly balanced, and nothing ever seems messy, surprising or slightly off-topic.

Signs worth a closer look:

• Many paragraphs are almost the same length and each makes exactly one point.
• The text is generic at the exact moments where a real person would add local, personal or numerical detail.
• Citations look impeccable — but some lead nowhere or do not exist. Hallucinated references remain one of the most reliable red flags.
• There are no inconsistencies, no typos, no hesitations, no humour that actually lands.

None of these signs prove AI involvement. A nervous student or a careful colleague can produce very clean text too. That is why the strongest verification is about process:

1. Ask for an earlier draft, an outline, or the version history in Google Docs or Word.
2. Ask a follow-up question that requires knowledge of the specific lesson, meeting or project — details that were never written down anywhere public.
3. Verify the sources yourself. Do the quoted studies, reports and laws actually exist and say what the text claims?
4. If you use a detector, treat a high score as an invitation to talk, never as a verdict.

Under GDPR, an important decision about a person — a mark, a grade, a job, a dismissal — should not rest solely on an automated score. Give the person a real chance to explain their process and show their work.

Images: labels and provenance first, pixels second

For years, guides told us to count fingers and look at teeth. By 2026, top-tier image models have largely fixed those problems. Pixel-level guessing is now the weakest method in your toolbox. Something else works much better: provenance.

Many professional generators now embed machine-readable information into their outputs. Adobe, Google, Microsoft and camera makers support Content Credentials, built on the C2PA standard. Google’s SynthID adds invisible watermarks designed to survive cropping, compression and resizing.

Practical image checks in 2026:

• Look for visible or hidden labels such as “AI-generated” or the Content Credentials icon, then open the file’s metadata if the platform lets you.
• Reverse-search the image with Google Lens or TinEye. If it supposedly shows a real event, where else does it appear? On trustworthy news sites, or only on anonymous accounts?
• Ask for the original file. A screenshot sent through WhatsApp has usually lost its metadata — that loss is itself a warning.
• Examine fine details only as a secondary clue: garbled text on shop signs, impossibly repetitive background patterns, or light that behaves differently on every object.

One decisive fact: if metadata has been stripped during sharing, no visual skill can tell you the image’s true origin. In that case, the question shifts from “is this AI?” to “can I verify where this file actually came from?”

Video and voice: the expensive fakes

Video is where AI deception hurts most. The Hong Kong case showed that a live video call can be faked convincingly enough to move tens of millions of dollars. Similar voice-cloning frauds have been reported across Europe, often targeting elderly people with a panicked call from a “relative”.

If you are on a live call and something feels off:

• Ask the person to turn their head to the side, move a hand slowly across their face, or change the camera angle. Real-time face-swap systems still struggle with deliberate, awkward movements.
• End the call and dial back on a number you know is correct — from the company website, the bank card, or your phone’s saved contacts. Do not use a number the caller gives you.
• Agree on a family or team code word in advance for really sensitive conversations. It sounds old-fashioned, but it is one of the few checks that still works perfectly.

For recorded videos, ask the basic questions of journalism: Who published this first? Do independent, reliable outlets confirm it? Have fact-checkers already examined it? A video that only exists as reposts without a verifiable original source is already suspicious, regardless of how real it looks.

What the EU now requires — in plain language

This is the year the transparency rules of the EU AI Act fully take hold. The Act has been phasing in since February 2025, with most remaining obligations applying from August 2026. In practice, this means:

• Providers of systems that create deepfakes must make their output detectable as artificial in a machine-readable way.
• Companies, media and influencers who publicly share realistic AI-generated or manipulated content must clearly disclose that it is artificial.
• Chatbots must tell you that you are talking to an AI — unless it is obvious from the context.

There are sensible exceptions for clearly artistic, satirical or fictional work. A fantasy poster does not need a warning label; a realistic video of a politician or a fake news report absolutely does.

Important European detail: the AI Act applies to any provider targeting the EU market, even if the company is based in the US or Asia. So the label you see on a major platform is not voluntary kindness — it is law. It is also not a perfect system: open-source tools, offline models and generators outside the EU may not label anything, and social networks often strip metadata when you download and re-upload.

Quick field table for everyday use

Build a ten-second verification habit

Most of us will never face a 25-million-dollar deepfake. We will face smaller, daily decisions: whether to believe a screenshot, share a video of a politician, accept a student’s explanation, or forward a suspicious invoice. A useful habit takes ten seconds:

1. Who is the original author or publisher?
2. Is the original file available, or only a repost and screenshot?
3. Do independent sources confirm it?
4. If it is a person asking for money or sensitive data, verify through a completely different channel.

And when you are the creator, not just the reader: label your AI content clearly. It is legally required in many European contexts, and it builds the trust that the next decade will desperately need.