Fakes, labels and rights
Lesson 12 of 13 in the course “AI Images and Slides”. Free, with animations and a quiz.
The Pope in a puffer
In March 2023 a photo of the Pope in a stylish white puffer jacket went around social media. Millions believed it, journalists included. It was made in Midjourney by Pablo Xavier, a 31-year-old construction worker from Chicago, just as a joke. Since then models have become far more realistic. The takeaway: “it looks like a photo” no longer proves anything. And if you make images yourself, you’re also responsible for how people read them.
How to check a picture
Looking for six fingers is almost useless now: 2026 models draw hands well. It’s more reliable to check the origin. Who posted it first, and is it in serious media? A reverse image search in Google Lens or Yandex Images often finds the original or a debunk. If you suspect Google AI, upload the picture to Gemini and ask: it will check for the invisible SynthID mark. Odd lettering and backgrounds are only hints, not proof. Not sure? Don’t repost.
AI marks
Services mark their images in three ways. A visible caption or badge: honest, but easy to crop. C2PA metadata, a kind of passport for the file with its creation history. ChatGPT, for example, adds it, but it’s lost with a screenshot or re-saving. And an invisible watermark in the pixels themselves, like Google’s SynthID: harder to remove. No mark is perfect, so checking the source is still needed.
Labelling and the law
In the EU, Article 50 of the AI Act has applied since 2 August 2026. Whoever publishes a realistic deepfake, meaning an AI image that’s easy to mistake for real, must disclose it. That covers not just people but objects and places too. Services already on the market got until 2 December 2026 to build in machine-readable marking. YouTube, TikTok and Instagram also ask you to flag realistic AI content. Other countries have their own rules, so the habit of honest labelling saves stress everywhere.
Copyright
Can you copyright a picture drawn by AI? In January 2025 the US Copyright Office wrote in a report that a prompt alone isn’t enough to make you the author. In March 2025 a US appeals court confirmed: a work with no human author gets no copyright. And in June 2025 Disney and Universal sued Midjourney over images of their characters. Hence two rules: don’t sell other people’s characters and brands, and put your own work in: editing, retouching, text, photos.
An honest product photo
For Liya’s shop the main risk isn’t a court, it’s customer trust. If an AI photo shows the candle bigger, brighter and smoother than in real life, returns and angry reviews follow, and marketplaces can take such a listing down. The honest version: the real candle, with the background, light and mood from AI. Size and colour true to life, at least one real photo in the listing, and a short “background made with AI” note. Beauty sells once, honesty sells every time.
Check yourself
- What’s the most reliable way to check whether a viral photo is real?
Correct answer: Find the original source and do a reverse image search. Checking the origin is more reliable than hunting for errors 2026 models barely make. - Why doesn’t C2PA metadata guarantee an AI picture will be recognised?
Correct answer: They’re lost with a screenshot or re-saving. It’s data in the file, not in the pixels, so simply re-saving wipes it. - Liya wants to sell candles with Disney characters drawn by AI. What’s wrong?
Correct answer: The characters belong to Disney; AI doesn’t change that. Others’ characters and brands are protected, and Disney and Universal are already suing Midjourney.