
Tech • AI • Robotics • Game
AI-generated and AI-edited images are increasingly hard to spot by eye, but forensic methods still offer useful ways to detect manipulation, while text detection remains far less reliable unless outputs are explicitly watermarked.
A rapid real-or-fake test using photos, edited images and fully generated visuals showed how unreliable instinct has become. Some authentic photos looked suspicious because of motion blur, compression, lighting or smartphone effects, while several AI creations appeared convincing at first glance. Small edits such as adding a pasta packet or changing shorts into trousers on a real photo were especially difficult to notice.
Realistic synthetic images no longer require cutting-edge proprietary models. Fine-tuned open-source systems can imitate the look of old cameras, film stocks or earlier smartphone photos, making fabricated images feel more credible. That “vintage” aesthetic helps hide errors because viewers often associate imperfect quality with authenticity.
Image forensics is already used against insurance scams, false parcel-return claims and food-delivery fraud. A manipulated picture showing damage, spoiled food, insects or hair can trigger refunds or payouts. In criminal cases, the stakes are higher: suspects may now argue that authentic surveillance or photographic evidence was AI-generated, a tactic known in the United States as the liar’s dividend.
One basic check is the metadata attached to files, which can reveal device type, location or processing history. The European Union is moving toward rules requiring generative AI providers to attach provenance metadata to online outputs, backed by cryptographic safeguards. But metadata remains fragile because it can be stripped, altered or lost through re-exporting, so it is only one element in a wider investigation.
Invisible watermarking embeds a hidden signal inside an image, audio file or text so detectors can later identify the origin. The approach is already familiar in copyright protection, TV audience measurement and film anti-piracy systems. Major AI companies have adopted versions of it, including Google’s SynthID, while camera and media companies are also using provenance marks to certify authentic content.
Tests on generated images showed that watermark detection can be inconsistent. In some cases, only a minority of outputs were correctly flagged even before users altered them. Attackers also look for ways to weaken or bypass watermarks, though robust systems are designed to survive cropping and routine edits better than simple metadata.
When no watermark is present, analysts search for statistical artifacts left by generation models. Unlike a camera, an AI system has no physical sensor, optics or natural sensor noise. That difference can leave unusual frequency patterns or noise residues in images and video, allowing detectors to identify signals tied to specific model families, at least before heavy recompression or repeated transformations erase them.
Modern systems often modify only a small visible element while preserving the rest of a real image, making visual inspection unreliable. In some cases, though, the editing model regenerates more of the image than users realize, spreading detectable traces beyond the changed object. Forensic tools can also compare statistical differences between one region and the surrounding background to spot composites such as face swaps.
Some of the strongest long-term defenses may come from proving authenticity at capture. Apple has promoted provenance features for iPhone imagery, and firms such as Imatag have worked with camera makers and news agencies including AFP and Nikon to embed secure marks at the start of the imaging chain. The goal is to confirm that a photo traveled from camera to publication without alteration.
Reliable text detection is much weaker than image detection because there is no camera-like physical process to analyze. Most detectors can only guess from writing style, which creates a high risk of false positives, especially as people absorb common AI phrasing or lightly edit generated drafts. The more promising route is text watermarking, in which a model subtly biases word choice across long passages so origin can later be verified with lower error rates.
The contest between AI generation and AI detection is becoming an arms race, but images still leave more forensic clues than text. In the long run, trusted provenance systems may prove more durable than trying to judge authenticity by eye alone.
Ask a question