Spotdle·Sep 26, 2026·Updated Sep 28, 2026·6 min read
The line is blurrier than the puzzle pretends.
GeneralGE
A press photographer dodges and burns a print to bring out a face in shadow. A studio composites a dragon onto a real Belfast hillside. Your phone "enhances" a night shot until the moon looks sharper than your eyes ever saw it. A model invents a person who never existed from a sentence of text. Four images. Ask a room full of smart people which ones are "real" and you will not get a clean answer. You will get an argument.
That argument is the whole point. Spotdle asks you to sort five items into real and AI each day, and the binary is honest enough for a game. But the world it samples from is not a binary at all but a gradient, and the gradient has been getting longer for a century.
Photography has been edited since the darkroom. Retouching, dodging, burning, cropping, color grading: none of it was ever controversial, because everyone understood a photograph as an interpretation, not a deposition. Cinema has been compositing fiction onto real plates since before computers. CGI did not break the line between real and fake; it just made the fakery cheaper and more convincing.
What generative AI changed is the default, not the existence of manipulation. The cost of producing a plausible image from nothing dropped to roughly the cost of typing a sentence. The skill floor collapsed. And the result is a flood in which the genuinely synthetic, the lightly assisted, and the heavily staged-but-photographed all swim together.
So when we say "is this real," we are usually smuggling in a question we have not defined. Real as in photographed? Real as in unedited? Real as in not deceptive? Those are three different tests, and they disagree constantly.
A more honest frame is: where did this come from, and what happened to it on the way here? That is a question about provenance (the chain of custody) rather than about some imaginary state of untouched truth. A retouched magazine cover has clean provenance and heavy editing. A "raw" screenshot from an anonymous account has murky provenance and could be anything. Purity is a fantasy. Provenance is something you can actually chase.
This is also why the entire policy world has stopped trying to police what gets made and started trying to police what gets disclosed.
In the EU, the AI Act's transparency rules (Article 50) have applied since 2 August 2026. They are not a ban on synthetic media. They are a disclosure regime, and worth understanding precisely because the details matter.
Two obligations carry the weight. Providers of systems that generate synthetic image, audio, video, or text must mark the output in a machine-readable format, detectable as artificially generated or manipulated. Separately, anyone deploying a deepfake (content made to resemble real people, places, or events) must disclose that it has been artificially generated, and AI-generated text published to inform the public on matters of public interest must be disclosed too.
Then come the carve-outs, which are where the messiness reasserts itself:
Artistic and creative work gets a softer standard: disclosure that the work exists in manipulated form, done in a way that "does not hamper the display or enjoyment of the work."
Standard editing that performs an assistive function and does not substantially alter the input is exempt from the marking rule. Your phone sharpening a photo is, broadly, not a "deepfake."
Editorial control matters: AI-generated text that a human has reviewed under editorial control can be exempt from the text-disclosure rule.
There's also a softer landing for older systems. The EU's Digital Omnibus on AI, provisionally agreed by Parliament and Council in May 2026 and formally adopted in July as Regulation (EU) 2026/1744, gives generative systems already on the market before 2 August 2026 until 2 December 2026 to meet the machine-readable marking requirement.
The takeaway is not "the law solves it." The takeaway is that even the most serious regulatory attempt to draw the line had to draw it with disclosure, exemptions, and a phase-in, because a bright line does not exist.
The platforms moved faster and looser. TikTok integrated C2PA Content Credentials to auto-detect and label content carrying provenance metadata (the first video platform to do so) and requires creators to disclose realistic AI content. Meta and YouTube run their own disclose-and-label regimes, with YouTube's hardest requirements aimed at political and election content. The approaches differ in who is on the hook and how strictly it is enforced, but the direction is shared: label it, or risk removal.
The plumbing underneath much of this is Content Credentials, the "nutrition label" standard from the C2PA coalition: Adobe, Microsoft, the BBC, and now hundreds of companies including OpenAI, Google, and camera makers. Some hardware now signs images at capture: the Leica M11-P was the first camera with Content Credentials built in, and Google's Pixel 10 signs photos with hardware-backed keys.
It is a real step. It is also brittle. Screenshotting an image discards the credential entirely. Most social uploads recompress the file and strip the metadata in transit. As RAND has noted, the scheme depends on end-to-end cooperation across an open ecosystem, which is exactly what an open ecosystem does not reliably provide. So a missing label tells you nothing, and a present one only vouches for the part of the journey that survived. Useful signal. Not a verdict.
We could hide the seams. Plenty of "spot the fake" content does: it treats the answer as a magic trick and keeps the method to itself. We think that is the opposite of useful.
So for every synthetic item in a Spotdle edition, the reveal shows you the prompt that summoned it. No mystery box. You get to see the exact sentence that produced the thing you may have just been fooled by, which is a far better teacher than being told you were wrong.
Our stance, plainly: disclosure is the product. We are not selling the thrill of a gotcha. We are trying to recalibrate your eye, and you cannot calibrate against a black box. The honest move (for a regulator, a platform, or a daily puzzle) is to say where the thing came from and let you judge.
No. Article 50 is a transparency rule, not a ban. Since 2 August 2026 it has required providers to mark synthetic output in a machine-readable format, and deployers to disclose deepfakes and certain AI-generated text. It does not prohibit making the content.
If an image has a Content Credentials label, is it trustworthy?
The label tells you what tool touched the file: useful, but fragile. Screenshotting an image, or uploading it to most social platforms, strips the metadata. A missing label proves nothing, and a present one only covers the chain that survived. Treat it as one signal, not a verdict.