Author: michelle.hauptmann

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

When I started my pilot study, I mainly wanted to find out whether my survey worked well. I expected to make a few small changes before starting the main study, but I didn’t expect the pilot study to change the way I think about my research. But if I am honest, I feel that it did much more than test my questionnaire. It gave me new ideas, new questions and a much clearer direction for the next step of my project.

One thing I learned very quickly is that research rarely gives you simple answers. Instead, every answer seems to lead to another question. At the beginning, I only wanted to find out whether people can recognise authentic, AI-edited and fully AI-generated images. Now I find myself asking something different: Why do people trust some images but not others? I think this has become one of the most interesting questions in my research.

The pilot study also helped me develop a few assumptions that I would like to investigate in the main study.

• I expect that highly photorealistic AI-generated images will continue to be much harder to recognise than authentic photographs.
• I believe that people’s confidence does not always match their actual performance.
• I think that our judgement is influenced less by whether an image is objectively real or fake and more by specific visual details that make an image appear trustworthy.
• I now suspect that AI-generated images may have a second effect that I had not considered before: they might slowly reduce our trust in authentic photographs as well.

This last assumption is probably the most interesting outcome of the pilot study for me. When I started this project I wanted to understand whether people can recognise AI-generated images. But after analysing the pilot study, I began wondering if the real issue is something bigger. Could AI-generated images slowly make us trust authentic photographs less? If people start doubting real images because they look “too perfect”, then AI is not only changing how images are created. It may also be changing how we decide what we believe.

The pilot study also made me rethink the design of my main study. For the pilot survey, I intentionally selected mostly very realistic AI-generated images because I wanted to create a situation that reflects how AI could actually be used in the real world. Several participants told me afterwards that the images were “too good” and that people would normally not see such difficult examples online. I understand this opinion, but I personally see it differently. If someone wanted to spread misinformation using AI-generated images, they would most likely choose the most convincing image available, not one with obvious mistakes. For this reason, I still believe that the image selection represents a realistic scenario.

However, the feedback also gave me an idea for improving the main study. Instead of using only highly realistic AI-generated images, I now want to include images with different levels of difficulty. Some images will contain obvious AI artefacts, while others will be almost impossible to distinguish from real photographs. I hope this will make it easier to compare different age groups. My current assumption is that younger participants, especially those between 25 and 34 years old, may recognise the more obvious AI-generated images more easily than older participants because they are generally more familiar with AI tools and digital media.

Looking back, I feel that the pilot study has given me much more than I expected. It helped me improve my survey, but it also changed how I think about my research question. I now have a clearer idea of what I want to investigate and which questions I find most interesting. At the same time, I know that there is still a lot to discover. I am excited to continue this project and see whether these first ideas are confirmed when I conduct the main study with a larger and more diverse group of participants.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

Once I had analysed the overall results, I became curious about the individual images. The average recognition rate already showed that identifying AI-generated images is difficult, but I wanted to understand what was behind those numbers. Which images confused people the most? Which ones were recognised more easily? Looking at each image one by one turned out to be one of the most interesting parts of the whole project.

As I mentioned in my previous blog post, authentic photographs were generally easier to recognise than AI-generated images. But once I started looking at the individual images, I realised that the averages only told part of the story.

Some authentic photographs were recognised by more than 70% of the participants, and one image was even recognised by almost 80%. At the same time, several fully AI-generated images were correctly identified by only around 20% of participants.

One result immediately caught my attention. Image 13 was an authentic photograph, but only 14.3% of participants identified it correctly. Image 20, which was fully AI-generated, reached exactly the same recognition rate. I found it fascinating that two completely different image types produced almost identical results. So, looking at each image individually revealed details that would otherwise have been easy to miss.

Another interesting part of the analysis was the confidence ratings (1 = not confident at all, 5 = very confident). Before analysing the data, I expected participants to feel much more confident when they answered correctly. Instead, the difference was surprisingly small.

On average, participants rated their confidence as 3.13 when they answered correctly and 2.99 when they answered incorrectly. Both values are very close to the middle of the scale, meaning that participants generally felt only moderately confident about their decisions. What surprised me most was how similar these two values were. Whether participants were right or wrong, their confidence hardly changed.

I also noticed a change in how participants viewed their own ability to recognise AI-generated images. Since participants rated their own ability before and after the survey, I was able to compare both answers. At the beginning of the survey, many participants rated themselves as being quite good at identifying AI-generated content. After completing the survey, several participants rated themselves as less confident than before. Although this pilot study is too small to draw firm conclusions, it seems that taking part in the survey made some people realise that distinguishing between authentic and AI-generated images is more challenging than they had expected.

The second part of my pilot study was a little different. This time, participants did not have to classify a single image. Instead, they saw two similar images and had to decide which one they thought was authentic. I was curious to find out if comparing two images directly would make the task easier.

The average recognition rate was 55.4%, which is slightly higher than in the first part of the survey. Even so, I was surprised that the results were still so close to chance. Some image pairs were much easier than others. Pair 3 was recognised correctly by 71.4% of the participants, while Pair 4 was the most difficult, with only 42.9% choosing the authentic image. Most of the other image pairs were recognised correctly by about half of the participants.

For me, this part of the study added another interesting perspective. It showed that even when participants could compare two images directly, deciding which one was authentic was often still difficult.

Analysing the results in this level of detail helped me understand much more than the overall recognition rate alone. It showed me which images were particularly successful for the study and which ones challenged participants the most. These observations will surely help me improve the image selection for my main study.

!!Attention!!

!!Please only look at the documents if you have already completed the study or if you do not intend to participate in it!!

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

Seeing the first responses come in was a really exciting moment. After spending so much time planning the survey, choosing the images and creating the questionnaire, I was finally able to look at the data. I was curious to find out whether people could actually tell the difference between authentic, AI-edited and fully AI-generated images. Even more importantly, I wanted to understand where people struggled and whether any interesting patterns would appear.

In total, 14 people completed my pilot study. They came from different age groups, although most participants were between 25 and 34 years old. Each participant looked at 24 images and decided whether each one was authentic, AI-edited or fully AI-generated. Altogether, this gave me 336 individual answers to analyse.

The first thing I looked at was the overall recognition rate. After comparing every answer with the correct solution, I found that participants classified only 42.3% of the images correctly. I honestly expected the result to be higher. At first, this seemed surprisingly low, but the more I thought about it, the more it made sense. AI-generated images have become incredibly realistic, and even manipulated images are often very difficult to recognise.

I also wanted to see whether some image categories were easier than others. Authentic photographs achieved the highest recognition rate, with 57.1% of participants identifying them correctly. AI-edited images were recognised correctly in 42.9% of the cases, while fully AI-generated images were the most difficult, with only 26.8% correct answers. This immediately caught my attention because it showed that participants struggled most with images that were created entirely by AI.

While these numbers already gave me a good overview, I quickly realised that there is more to it. Some images performed much better than others, so I decided to analyse every image individually. Looking at the results in more detail helped me understand which images confused participants and which visual characteristics might have influenced their decisions. This became one of my favourite parts of the analysis.

Besides the recognition rates, I also collected confidence ratings. After every image, participants indicated how confident they felt about their answer. I included this question because I wanted to know whether people were aware when they were uncertain. It is one thing to answer correctly, but it is another thing to know how reliable your own judgement actually is. I will analyse this in more detail in the next blog post.

One thing that became clear after analysing the pilot study is that recognising AI-generated images is more difficult than many people might expect. Before starting the study, I already assumed that people tend to overestimate their ability to recognise AI-generated images. Although this pilot study is too small to confirm that assumption, the results encourage me to explore this question further in the main study by comparing participants’ self-assessment with their actual performance.

!!Attention!!

!!Please only look at the documents if you have already completed the study or if you do not intend to participate in it!!

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

Conducting my first interview was a valuable experience. Not only because of the participant’s answers, but also because it taught me a lot about my own interviewing skills.

Before the interview, I was mainly focused on asking the right questions. I prepared an interview guide and thought about the order of the questions, but I quickly realised that asking questions is only one part of the process. Listening carefully is just as important.

One challenge I faced during the interview was taking notes while also paying full attention to the participant. Sometimes I found myself looking down to write something instead of maintaining eye contact or thinking about a good follow-up question. Although I still managed to collect all the important information, I noticed that my attention was divided.

Despite this challenge, I felt that the atmosphere during the interview was very relaxed. The participant seemed comfortable and answered all questions openly. I think creating a friendly and comfortable environment made it easier for them to share their thoughts honestly. This was also reflected in the observation sheet completed by another student, who described the interview as friendly, respectful and professional. The observer also noted that the participant appeared comfortable throughout the conversation.

Another thing I learned was how valuable follow-up questions can be. Some of the most interesting answers were not the direct responses to my prepared questions but came from asking participants to explain their thoughts in more detail. This helped me understand not only what they think but also why they think that way.

Looking back, there are also a few things I would improve next time. Instead of writing detailed notes during the interview, I would like to record the conversation first, with the participant’s permission, and transcribe it afterwards. This would allow me to focus completely on the conversation, maintain better eye contact and react more naturally to what the participant is saying. It would also reduce the risk of missing interesting details while writing.

Overall, I learned that conducting an interview is a skill that improves with practice. Preparing good questions is important, but creating a comfortable atmosphere and actively listening are just as essential. This interview was a great learning experience for me and gave me more confidence for future qualitative research. I am looking forward to using these experiences in the next interviews I conduct.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

After spending a lot of time designing my survey, I wanted to include another research method in my project: an interview. While the survey focuses on collecting data from many people, the interview gave me the opportunity to explore one person’s thoughts in much more detail.

Before the interview, I prepared a list of open questions. I wanted to understand how the participant thinks about AI-generated images, whether they trust images online and which clues they use to decide if an image is authentic.

The participant told me that they use AI quite often and spend a lot of time online. Because of this, they felt quite confident in recognising AI-generated images. At the same time, they also said that it is becoming more and more difficult to tell real and AI-generated images apart. I found this interesting because even someone with a lot of experience with AI feels uncertain sometimes.

Another topic we discussed was what people actually look for when judging an image. The participant explained that they usually pay attention to unrealistic content, strange details or common AI mistakes such as unusual hands or facial features. They also mentioned that images with very “perfect” skies or unrealistic scenes immediately make them suspicious. These answers were similar to what I had already read in some of the literature, which made me feel that my research is moving in the right direction.

One answer surprised me the most. When I asked how much they trust images online, the participant explained that they are generally careful. Not only because of AI, but also because authentic images are often taken out of context or edited to support a certain story. This reminded me that image authenticity is about much more than AI. The context in which an image is shared also plays an important role.

We also talked about the importance of the source. The participant said that trustworthy sources can increase confidence, but they still prefer to verify information instead of believing it immediately. I think this is an interesting point because it shows that people do not only judge the image itself but also where it comes from.

Overall, I really enjoyed conducting this interview. It helped me look beyond simple survey answers and better understand the thoughts behind people’s decisions. Even though it was only one interview, it gave me useful ideas that I can keep in mind when analysing my survey results later. I am curious to see whether other participants think in a similar way or whether completely different patterns will appear.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

After finishing the questionnaire, I thought the hardest part was behind me. I was wrong. The next step, choosing the images for the survey, turned out to be just as challenging.

At first, I simply started collecting images that I thought would fit my study. But after a while, I realised that this approach would not work. Every image I included could influence the results, so I needed to think much more carefully about my selection.

I also wanted to avoid focusing on only one type of subject. If all images showed people, the results might only tell me how well participants recognise AI-generated faces. To get a broader picture, I chose images from different categories, including people, animals, landscapes and everyday scenes. This will hopefully help me see whether some types of images are easier to judge than others.

One category was especially important to me: AI-edited photographs.

Many studies and online discussions focus on images that have been created completely by AI. However, I think AI-edited images are just as interesting, if not even more interesting. In many cases, these images look almost identical to a real photograph because only a small part of the image has been changed. An object might have been added, removed or replaced, while everything else remains authentic. These small changes are often much harder to notice than a completely AI-generated image.

Besides the individual images, I also added a second part to my survey. In this section, participants compare two very similar images. One image is authentic, while the other has been edited using AI. Instead of deciding between four answer options, they simply have to choose which image they believe is the real one.

I included this second task because I am curious whether people perform differently when they can directly compare two images. Maybe small AI edits become easier to recognise when there is an authentic version right next to them, or maybe they are still difficult to spot. I am looking forward to finding out.

One final challenge was the survey platform itself. Google Forms reduces the image quality slightly, so very small details are not always visible. Because of this, I avoided choosing images where the answer depends on tiny artefacts or imperfections. Instead, I selected images that can still be judged even at a lower resolution.

Looking back, choosing the images took much longer than I expected, but it also made me think more deeply about my research question. Every image is now part of my experiment and plays an important role in helping me understand how people perceive authenticity in a world where AI-generated content is becoming more and more common.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

In my last blog post, I wrote about the process of designing my survey and how a small pre-test helped me improve the instructions. After making these changes, I finally felt ready to share the questionnaire with participants.

The survey starts with a short introduction. I explain the purpose of my research and introduce the three image categories that participants will see during the study: authentic photographs, AI-edited photographs and fully AI-generated images. Since the pre-test showed that the term AI-edited could be misunderstood, I also explain that normal photo editing, such as changing the brightness or colours of an image, is not considered AI editing in my study. Only changes to the actual content of an image count as AI editing.

Before participants start looking at the images, I ask them three short questions.

The first question asks for their age group. Since I want people of different ages to take part in my survey, this information will help me compare the results later. I am curious to see whether younger participants, who may have grown up with AI tools and social media, perform differently from older participants.

The second question asks how familiar participants are with AI-generated images. Today, many people have already seen AI-generated content online, but others may have very little experience with it. I think this could have an influence on how people judge the images in my study.

The third question is one of my favourites because it asks participants to rate themselves before the experiment even begins. I ask them how well they think they can recognise AI-generated or AI-edited images. Later, I can compare their self-assessment with their actual results. It will be interesting to see if people who feel confident are really better at recognising AI images, or if confidence and performance are not always connected.

After these introductory questions, the main part of the survey begins.

Participants see one image at a time and answer the same two questions for every image.

First, they decide how they think the image was created. They can choose between an authentic photograph, an AI-edited photograph, a fully AI-generated image or Not sure.

After that, they rate how confident they are in their decision on a scale from one to five.

I decided to include the Not sure option because I did not want to force participants to guess if they were genuinely unsure. I also think that the confidence rating adds another interesting layer to the results. Someone might choose the correct answer but still feel very uncertain, while someone else could confidently choose the wrong category. Looking at both the answers and the confidence ratings might reveal interesting patterns later.

While the survey itself is quite simple, every question has a purpose. My goal was to create a questionnaire that is easy to complete but still provides enough information to answer my research question.

In my next blog post, I will explain how I selected the images for the survey and why this turned out to be one of the most difficult parts of the whole project.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

Over the past few weeks, I have been working on the survey for my master’s project. At first, I thought creating the questionnaire would be one of the easier parts of my research. I already knew what I wanted to investigate, so I assumed I only had to write a few questions and choose some images. However, I quickly realised that designing a good survey is much more complicated than it seems.

My research is about whether people can tell the difference between authentic photographs, AI-edited photographs and fully AI-generated images. Because of this, I wanted the survey to be easy to understand while still collecting useful data. Every participant should understand the questions in the same way, otherwise the results could become misleading.

I decided to create an online survey because it allows many different people to participate. Everyone sees the same images and answers the same questions, which makes it easier to compare the results later. I also wanted the survey to feel similar to how people normally look at images online. Usually, we don’t spend several minutes analysing a picture, we scroll through social media, see an image for a few seconds and immediately decide whether we think it looks real or not. That is why I ask participants to trust their first impression instead of thinking too long about every image.

One challenge appeared while I was writing the introduction. I had to explain the difference between authentic and AI-edited images. At first, I thought my explanation was clear enough, but before sharing the survey publicly, I asked three people to test it for me.

The purpose of this small pre-test was not to collect data. Instead, I wanted to find out if there were any questions that were confusing or difficult to understand. This turned out to be a very good decision.

The three participants gave me valuable feedback. One thing became clear very quickly: the term AI-edited was not as obvious as I had expected. Some people thought that every edited photograph should be classified as AI-edited, even if it had only been adjusted using brightness, contrast or colour correction.

Because of this feedback, I changed the introduction of my survey. I now explain more clearly that normal photo editing does not count as AI editing. Small adjustments like brightness, contrast, saturation, sharpness or colour grading are very common in photography and do not change the content of an image. In my study, an image is only considered AI-edited when artificial intelligence has been used to add, remove, replace or change objects or other elements within the picture.

This small change made me realise how important clear instructions are. Even if a survey is carefully designed, people may still understand questions differently than intended. Testing the questionnaire before starting the actual study helped me identify these problems early and improve the survey.

Looking back, creating the survey was much more than writing questions. It became an important part of my research process and taught me that small details can have a big impact on the quality of the results. In my next blog post, I will introduce the structure of the questionnaire and explain why I chose these specific questions.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

In my previous blog post, I introduced the topic I will be exploring this semester and discussed the growing difficulty of distinguishing between authentic and AI-generated images. One question kept coming up during my research: are people really able to tell the difference?

Many people seem confident that they can spot an AI-generated image immediately. Common clues that are often mentioned include unrealistic hands, strange facial features, or unusual details in the background. However, image generation technology is improving rapidly, and many of these obvious signs are becoming less common. This made me wonder whether people are actually as good at identifying AI-generated images as they think they are.

To explore this question, I am planning a small experiment.

The experiment will consist of two different parts. The first part focuses on fully AI-generated images and authentic photographs. Participants will be shown a collection of images from different contexts, including everyday situations, animals, scientific topics, and news-related content. For each image, they will be asked to decide whether they believe it is authentic or generated by artificial intelligence.

I deliberately want to include different types of content because context may influence how people judge an image. A portrait of a person might be evaluated differently than an image of a rare animal or a news event. By using a variety of subjects, I hope to gain a broader understanding of how people make these decisions.

The second part of the experiment is the one I find particularly interesting. Instead of showing completely different images, participants will be presented with two almost identical versions of the same image. One will be the original photograph, while the other will contain a modification created with AI. This modification could involve adding an object, removing a person, or changing certain elements within the scene.

Participants will then be asked a simple question: Which image is the authentic one?

This part of the experiment is designed to investigate whether people find it easier to identify AI when they can directly compare an original image with a manipulated version. While fully generated images receive a lot of attention, AI is increasingly being used to alter existing photographs rather than create entirely new ones. Because of this, understanding how people perceive manipulated images may be just as important as understanding how they perceive generated ones.

Another aspect I would like to explore is the role of age. In addition to their answers, participants will be asked to indicate their age group. This will allow me to compare the results of different generations and examine whether younger participants are better at recognizing AI-generated or AI-manipulated content.

A common assumption is that younger people may perform better because they are more familiar with digital technologies and encounter AI-generated content more frequently. However, it is also possible that the differences between age groups are smaller than expected. The experiment may reveal whether this assumption is actually true.

Of course, this will only be a small-scale experiment and cannot provide definitive answers. Nevertheless, I hope it will offer an interesting insight into how people currently interact with AI-generated imagery and whether our confidence in recognizing artificial content matches reality.

The next step will be selecting and preparing the images that will be used in the survey. Only then will it become clear how difficult this challenge really is.

To the survey: https://docs.google.com/forms/d/e/1FAIpQLSfujaTdkiyrwagGO4Cv_JjGQ_IIWt9v4tO54aVBSiF7J6a_pw/viewform?usp=header

This semester, I’ll be focusing on a different topic than I did last semester. At the time, I wasn’t completely sure where my research interests would lead me, so we were given the opportunity to adjust our topics later on if needed. Over the past few months, however, I found myself becoming increasingly interested in the way we perceive images and why we trust them. That curiosity eventually led me to change my focus and explore this topic in more depth.

Every day, we are surrounded by images. Whether we are scrolling through social media, reading the news, or simply browsing the internet, images are everywhere. Most of the time, we accept them without giving them much thought. We rarely stop to ask whether an image actually shows what it claims to show.

At the same time, the tools used to create images are becoming more powerful. The rapid development of artificial intelligence has made it possible to generate images that look surprisingly realistic. In many cases, it has become difficult to tell whether an image is a real photograph or something that was created entirely by AI.

Over the past few months, I have come across more and more examples of AI-generated images appearing outside of technology-related discussions. They show up on social media, in advertisements, and sometimes even alongside news stories. Seeing this made me wonder how much we can really trust what we see online.

For a long time, photographs were seen as evidence. Even though image manipulation has existed for decades, photographs still carried a certain sense of authenticity. A photo was often considered proof that something had actually happened. Today, that assumption feels less certain. With only a few prompts, AI can create convincing images of people, places, and events that never existed.

What I find particularly interesting is that many people believe they can easily spot AI-generated images. Common signs that are often mentioned include strange-looking hands, unusual facial features, or unrealistic lighting. However, image generation tools are improving at a remarkable pace, and many of these obvious clues are becoming less common.

During my initial research, I found several examples where people confidently identified AI-generated images as real photographs. At the same time, genuine photographs were sometimes accused of being fake. This suggests that distinguishing between real and artificial images may be much more difficult than we think.

What fascinates me most is not only whether people can correctly identify an image, but also why they trust it. Does the context matter more than the image itself? Are we influenced by familiar faces, personal experiences, or our own expectations? And what actually makes an image feel believable?

To explore these questions further, I plan to conduct a small experiment in one of my upcoming blog posts. I want to find out whether people are really as good at recognizing AI-generated images as they often claim to be. Before that, however, the next post will focus on explaining the methodology behind the experiment and how it will be carried out.