What New Research Says About Scams, Deepfakes and Photos

Three studies test different parts of online safety

New research on online deception and privacy looks at three different questions: whether scam warnings can spread through a community, whether people can learn to spot fake face images, and whether AI can infer where a photo was taken. The studies suggest possible ways to reduce risk, but they do not establish how well those approaches work in everyday life.

Each paper is a preprint. That means it has been shared publicly but has not yet been checked by other scientists through peer review. The scam-warning and photo-location papers report computer-based tests; the deepfake paper reports an online experiment and a later follow-up. These are different kinds of evidence, and none alone settles a broad question about online safety. Findings from a model or a particular set of images may not carry over to other people, places, tools or situations.

Warnings helped in a simulation, but did not spread on their own

The researchers behind Scam Town wanted to know whether people who learn about scams can pass useful warnings to others, and whether that shared knowledge might protect a community. Earlier scam-prevention research, they note, often looks at one person at a time.

They built a simulated town in which a large language model played each resident. Scam centres called the simulated residents, who could then warn people they knew. Across 80 simulated towns, everyday conversations reduced the number of people scammed by roughly a fifth. But each warning led to fewer than one additional warning on average, so word of mouth did not grow on its own. The largest simulated benefit came when an official warning reached most residents and people also discussed it: that combination reduced scams by about half in the model. ¹

This is a computer simulation, not a trial with real communities. Its residents were language-model stand-ins, and the authors say their predictions need testing with people. The combined official-warning and conversation result was also a simulated combination, not something tested live. The work suggests that warnings may be more useful when they reach people likely to be targeted and are easy to pass along. It does not tell us how much any particular real-world campaign would reduce scams.

Training improved deepfake detection in most tested conditions

This study asked whether people’s ability to identify fake face images changes depending on how the images were made, and whether training can help. The researchers examined five image-generating models: three based on GANs, a type of image-generation system, one diffusion model, and one multimodal large language model, GPT-5.

In an online experiment with 524 participants, detection accuracy differed substantially by model. More recent and advanced models were not necessarily harder for participants to detect. Training improved performance across most conditions, and training focused on a particular model generally produced larger numerical gains than general-purpose training. In a follow-up test five months later, involving 342 people, the researchers found no significant decline in detection accuracy and no significant increase in perceived stress. ²

The reported results apply to the images, training materials and participants tested; they do not show how people will perform with every deepfake they encounter online. The paper is a preprint, and its findings still need independent checking and replication. The authors report an online experiment, but the source summary does not describe enough detail to assess how participants were recruited or how closely the test matched everyday image viewing. For ordinary readers, the study suggests that detection skills can improve under tested conditions, while also showing that accuracy depends on the image source.

Small image clues helped AI estimate photo locations

The photo-location study explored whether reasoning-focused AI systems could infer where ordinary pictures were taken, even without famous landmarks or direct location indicators. The researchers tested a method called GeoTrace on 525 images from natural, urban, residential and human-centered settings.

Using Gemini-3-Flash, the method placed 25.71% of images within one kilometre of their location; with GPT-5-mini, it did so for 16.95%. The researchers found that broad scene features helped with rough location estimates, while small, distinctive details contributed more to precise ones. They then tested GeoShift, an image-editing defense designed to alter high-impact location clues while keeping the picture visually coherent. In the reported test, street-level localization fell from 16.95% to 2.48% for GPT-5-mini and from 25.71% to 3.81% for Gemini-3-Flash. ³

These results come from a benchmark of 525 images and two named AI systems, not from every photo-sharing situation or location tool. The defense was evaluated in the researchers’ experiments, including against adaptive attacks, but that does not establish how it performs with other systems or in routine use. The paper is also a preprint. Its findings highlight that removing obvious location data may not remove every clue in an image, but they do not show that every ordinary photo can be reliably traced.

The studies point to useful possibilities, not settled protection

Together, the papers suggest that online safety can involve more than one layer: people passing on scam warnings, training to recognize certain fake images, and changes to visual details that reveal a photo’s location. Each approach showed promise in its own test, but the evidence is preliminary and the methods are not interchangeable.

The most important limit is that a result in a simulation, a controlled online image task or a specific benchmark is not a guarantee of protection in everyday use. The scam simulation needs testing with real communities; the deepfake findings need checking beyond the tested images and models; and the photo defense needs evaluation across more tools and settings. Until such evidence is available, these studies are best read as starting points for further research, rather than as proof that people or current tools can reliably prevent online deception and privacy risks.