Jean-Nicolas Gilles Hopfer

My strategy was actually quite simple. First, the setup gave a huge advantage from the start: with only two images to choose from, this provides a 50/50 chance. That got rid of the uncertainty of a single-image challenge, where you'd have to figure out if something's real without comparison, which would be much harder!

I spent more time than I probably should have each day, opening the calendar and zooming in as much as possible on my laptop. I'd look at everything that could be spotted. Often, it was straightforward things like text-rendering glitches or architectural issues. But some images were much trickier and required zooming in on tiny details while also using logic: which picture is more likely to be from a stock photo library, and which could be AI-generated?

The first day was actually the trickiest, and only a third of participants got it right. In that case, the AI-generated image was too perfect. My first gut feeling was: "It's a trap!" When I compared these two images, the AI version had the lions in almost identical poses with nearly the same facial features. In comparison, the stock image (which turned out to be the correct answer) had some imperfections.

There was some kind of out-of-focus fencing or structure in the background contaminating the bokeh, you couldn't tell if it was from a sanctuary, rescue centre, or a zoo... but crucially, the lion's fur was much more detailed and messy, if you will, compared to the AI version, where everything looked almost studio-perfect. The lion and cub even had nearly identical facial features in the AI image, which struck me as suspicious. I know from experience that AI has difficulty consistently generating multiple faces, and here it seemed like the cub was just a copy-paste of the adult lion's facial characteristics, including patches of hair colour. That level of detail convinced me.

Some other challenging days included the pedestrian crossing (day 5). For the crossing image, I couldn't read the kanji characters to verify their accuracy, and I couldn't read the out-of-focus license plates on the vehicles. Without those tell-tale signs, I had to rely on intuition, the composition of the right image seemed less interesting, less professionally framed, more like it was captured in the moment rather than perfectly posed. Luckily, that turned out to be correct.

Absolutely. I developed several patterns to watch for:

• Grain and lighting: AI-generated images, unless you adjust specific parameters, tend to be very smooth and polished with perfect lighting. Real photography has that characteristic grain that's hard to fake. The photographic texture is a dead giveaway.
• Architecture and details: I'd always check windows, doors, architectural elements, light switches... looking for issues with hinges, inconsistent placement, or structural elements that don't make sense. For example, on day 3 with the bridge image, the support planks underneath were positioned strangely in the AI version, and the connections were in odd locations. When I compared it to the real image, where the support planks were regular and aligned with the tension cables, it made much more sense.
• Composition and poses: AI often creates symmetry or poses that don't make sense. On day 12, with the snowstorm image, the person looked like they were throwing their hands up in surrender while laughing: a very unnatural, symmetrical pose. The real image, while you couldn't see much detail, felt much more natural.
• Plausibility: sometimes it's about what's more believable. On day 14 with the pastries, I asked myself: what's more plausible, cakes that are extremely perfect and identical, or ones with minor imperfections? The paper cupcake liners also had illogical patterns on the AI version as well.

I've been using AI image generation platforms for over two years. I started with Midjourney version 3 and still use it occasionally on certain projects. I find that Midjourney has the most potential in terms of style and medium, you can really push the boundaries as an art director. It offers tremendous artistic control.

DALL-E, I've always found it very flat: all the images tend to look the same, as you can probably see on LinkedIn nowadays.

Nano Banana is much better and promising, as demonstrated in this Advent calendar. But no matter which model is used, I find that, by default, the generated images often present artefacts, glitches, or other issues, and it's necessary to use Photoshop or other retouching tools afterwards to correct them. Or you must run countless generations and prompt-refining to get the right seed.

My workflow with image generation typically involves Midjourney first, then I'll go through a pass with Photoshop, potentially integrating real elements, for example: if a hand was badly generated, or if there are issues with eyes or finger positions, which still sometimes happen. But it's becoming easier and easier.

However, I think what helps me most in detecting AI-generated images isn't my knowledge of the tools themselves, but rather my knowledge of photography. I've been doing photography for over 25 years, self-taught, starting with film photography, including developing black-and-white film, then moving to digital photography. Through extensive practice in both shooting and editing RAW files, I believe I've gained a deep understanding of how light is captured through a lens, depth of field, bokeh, grain, textures, reflections, and perspective...

This photographic knowledge helps me spot images that are "off" compared to that foundation. Ironically, what works best in photography, and therefore in AI image generation, are the imperfections. When I generate an image in Midjourney and try to achieve "photorealism, I actually add extra layers and textures afterwards to make the image less pure, by adding grain and texture, making it less polished and less artificial. These are all things you find in the imperfections of real photography or camera lenses. I think that with a bit of practice, you can train your eyes to be more aware of what makes real photography look real.

Obviously, these tools are constantly changing: by the time this interview is published, there will probably be new models that change things. Initially, the obvious tells were hands and faces, especially in group shots with multiple people. There were many problems generating correct hands with five fingers.

Eyes looked strange, all the female models seemed to have freckles, faces looked too similar, lack of diversity... But that's changing now. With DALL-E, Nano Banana, and Midjourney, fingers and hands are much more advanced, making it harder to catch artefacts. Sometimes you can still spot issues in hair: maybe there's too much volume or the haircut is too perfect.

Rule #1: try to find the source. With this calendar, all the stock images could be traced back to their original source. While using reverse image search would have been "cheating" in the context of this contest (full disclosure: I only used it afterwards to verify), tools like Google Images (right-click search) or other online tools are great for checking viral images or news content. It's crucial to find the source: is this from an official photographer, a press agency, or another reliable source? For images without a direct source, you need to look more carefully at lighting, posture, grain, and what's happening in the image.

EXIF metadata: in a real photograph from a digital camera, the image contains metadata that can display the camera model, date taken, lens type, GPS position, etc. Usually in photojournalism, this metadata is available. An AI-generated image won't have this information, though it won't necessarily say "generated by AI" either. There are tools and systems being put in place to mark AI-generated content (like C2PA/Content Credentials), but these can be easily bypassed. An image captured by a camera can be saved in another format and its metadata removed. So while metadata can be useful to check, it's not perfect.

Online detection tools exist, but I tested 3-4 of them out of curiosity during this challenge, and what was interesting was that some real stock images were flagged as... AI-generated! These tools aren't 100% reliable, so you need to do manual checking.

Advanced technical analysis: there are some interesting studies on mathematical detection methods. For example, researchers have found that AI-generated images can have unusual noise patterns. In real photographs, noise from the camera sensor is random and varies across the image, but some AI models generate images with noise that's too uniform or has patterns that don't match real camera sensors. This is pretty technical stuff, and you need specialised tools to analyse it (you can get somewhere with Photoshop), but it seems like one of the best methods being developed to spot AI-generated content at the moment.