KI-generierte Bilder Erkennen: Wie man Deepfakes wie den Balenciaga-Papst Identifiziert

Deepfakes erkennen: Dieses KI-generierte Bild zeigt Papst Franziskus untypisch gekleidet in einer Balenciaga-Daunenjacke. Beachten Sie die surreal anmutenden Details wie den schwebenden Kreuzanhänger und die ungreifbare Kaffeetasse – subtile Hinweise auf die digitale Herkunft des Bildes. (Bildquelle: Instagram @art_is_2_inspire)
Deepfakes erkennen: Dieses KI-generierte Bild zeigt Papst Franziskus untypisch gekleidet in einer Balenciaga-Daunenjacke. Beachten Sie die surreal anmutenden Details wie den schwebenden Kreuzanhänger und die ungreifbare Kaffeetasse – subtile Hinweise auf die digitale Herkunft des Bildes. (Bildquelle: Instagram @art_is_2_inspire)

Die Bedrohung durch KI-generierte Bilder im Kontext von Fake News, besser bekannt als Deepfakes, ist seit Jahren ein Thema in der öffentlichen Diskussion. Bis vor kurzem war es jedoch relativ einfach, ein KI-generiertes Bild von einem echten Foto zu unterscheiden. Doch diese Zeiten sind vorbei. Am vergangenen Wochenende ging ein KI-generiertes Bild von Papst Franziskus in einer Balenciaga-Daunenjacke viral und führte viele Internetnutzer in die Irre.

Die rasante Entwicklung der KI-generierten Bilder

In nur wenigen Monaten haben öffentlich zugängliche KI-Tools zur Bildgenerierung einen Grad an Fotorealität erreicht, der beeindruckend ist. Obwohl das Bild des Papstes einige verräterische Anzeichen einer Fälschung aufwies, war es überzeugend genug, um viele zu täuschen. Dieses Ereignis könnte als die erste wirklich virale Desinformation in die Geschichte eingehen, die durch Deepfake-Technologie angetrieben wurde.

Die Gefahren von Deepfakes

Opfer von Deepfakes, insbesondere Frauen, die Opfer von nicht einvernehmlicher Deepfake-Pornografie geworden sind, warnen seit Jahren vor den Risiken dieser Technologie. In den letzten Monaten sind Bildgenerierungstools noch zugänglicher und leistungsfähiger geworden, was zu qualitativ hochwertigeren gefälschten Bildern geführt hat. Mit dem raschen Fortschritt der künstlichen Intelligenz wird es noch schwieriger, echte von gefälschten Bildern zu unterscheiden. Dies könnte erhebliche Auswirkungen auf die Anfälligkeit der Öffentlichkeit für ausländische Einflussnahme, gezielte Belästigung und das Vertrauen in Nachrichten haben.

Wie lassen sich KI-generierte Bilder und Deepfakes erkennen?

Es gibt einige Tipps, wie Sie KI-generierte Bilder und Deepfakes erkennen können:

  1. Trügerische Details: Betrachtet man das Bild des Balenciaga-Papstes aus der Nähe, lassen sich einige verräterische Hinweise auf seine Herkunft aus der KI erkennen. Das Kreuz, das scheinbar ohne Kette in der Luft hängt, oder die Kaffeetasse, die ohne erkennbaren Henkel in seiner Hand steht, sind solche Hinweise.
  2. Unnatürliche Physik: KI-Generatoren verstehen oft nicht, wie Objekte in der realen Welt interagieren. Unlogische Elemente wie schwebende Objekte oder ungewöhnlich geformte Körperteile können ein Indikator für KI-Generierung sein.
  3. Detailgenauigkeit: KI-Bildgeneratoren sind im Wesentlichen Musterreplikatoren. Sie haben gelernt, wie der Papst aussieht und wie eine Daunenjacke von Balenciaga aussehen könnte, und können beides erstaunlich gut kombinieren. Aber die Gesetze der Physik verstehen sie (noch) nicht. So wird der scheinbar schwebende Kreuzanhänger oder die unlogische Verschmelzung von Brillengläsern und deren Schatten zum verräterischen Detail.
  4. Fehlende Logik: In den Randbereichen eines Bildes können Menschen intuitiv Unstimmigkeiten erkennen, die KI nicht versteht. Diese Unstimmigkeiten können ein Hinweis auf eine KI-Generierung sein.
  5. Technische Grenzen: KI-Generatoren haben Schwierigkeiten, komplexe und detaillierte Szenen fehlerfrei zu reproduzieren. Achten Sie auf Anomalien in Texturen oder ungewöhnliche Muster.
  6. Inkonsistente Beleuchtung: Beleuchtung und Schatten sind für KI-Generatoren oft schwierig korrekt darzustellen. Achten Sie auf inkonsistente Lichtquellen, unstimmige Reflexionen in den Pupillen oder unnatürlich wirkende Schatten.
  7. Unnatürliche Proportionen: KI-Generatoren können Schwierigkeiten haben, die richtigen Proportionen von Gesichtern oder Körpern zu reproduzieren. Achten Sie auf ungewöhnliche oder verzerrte Proportionen als Hinweis auf eine KI-Erstellung.

Wie wir uns in Zukunft nicht durch Deepfakes täuschen lassen

Im Moment sind Medienkompetenztechniken vielleicht Ihre beste Möglichkeit, um mit KI-generierten Bildern Schritt zu halten. Fragen Sie sich: Woher kommt dieses Bild? Wer teilt es und warum? Wird es durch andere verlässliche Informationen widerlegt?

Suchmaschinen wie Google bieten ein Tool für die umgekehrte Bildsuche an, mit dem Sie überprüfen können, wo ein Bild bereits im Internet geteilt wurde und was darüber gesagt wird. Dieses Tool kann Ihnen dabei helfen herauszufinden, ob Experten oder vertrauenswürdige Publikationen ein Bild als Fälschung eingestuft haben.


Dieser Artikel ist ein Auszug aus dem Buch KI für Content Creation von Alexander Loth. Alle Infos zum Buch und eine kostenlose Leseprobe findet ihr bei Amazon.

Diskutieren Sie mit auf LinkedIn, wie man Deepfakes erkennt und was das für unsere digitale Zukunft bedeutet:

„KI-generierte Bilder Erkennen: Wie man Deepfakes wie den Balenciaga-Papst Identifiziert“ weiterlesen

GPT-4 Launches Today: The Rise of Generative AI from Neural Networks to DeepMind and OpenAI

OpenAI GPT-4 launch illustrated with Stable Diffusion (CC BY-SA 4.0)
OpenAI GPT-4 launch illustrated with Stable Diffusion (CC BY-SA 4.0)

With today’s launch of OpenAI’s GPT-4, the next generation of its Large Language Model (LLM), generative AI has entered a new era. This latest model is more advanced and multimodal, meaning GPT-4 can understand and generate responses based on image input as well as traditional text input (see GPT-4 launch livestream).

Generative AI has rapidly gained popularity and awareness in the last few months, making it crucial for businesses to evaluate and implement strategies across a wide range of industries, including e-commerce and healthcare. By automating tasks and creating personalized experiences for users, companies can increase efficiency and productivity in various areas of value creation. Despite being in development for decades, it’s high time for businesses to apply generative AI to their workflows and reap its benefits.

Before you dive into OpenAI GPT-4, let’s take a quick look back at the evolution of generative AI…

The history of generative AI begins in the late 1970s and early 1980s when researchers began developing neural networks that mimicked the structure of the human brain. The idea behind this technology was to assemble a set of neurons that could pass information from one to another with some basic logic, and together the network of neurons could perform complicated tasks. While minimal advances were made in the field, it remained largely dormant until 2010, when Google pioneered deep neural networks that added more data, hardware, and computing resources.

In 2011, Apple launched Siri, the first mass-market speech recognition application. In 2012, Google used the technology to identify cats in YouTube videos, finally reviving the field of neural networks and AI. Both Google and NVIDIA invested heavily in specialized hardware to support neural networks. In 2014, Google acquired DeepMind, which built neural networks for gaming. DeepMind built AlphaGo, which went on to defeat all the top Go players, a pivotal moment because it was one of the first industrial applications of generative AI, which uses computers to generate human-like candidate moves.

OpenAI was founded to democratize AI as a non-profit organization

In 2015, OpenAI was founded to democratize AI and was established as a non-profit organization. In 2019, OpenAI released GPT-2, a large-scale language model capable of producing human-like text. However, GPT-2 sparked controversy because it could produce fake news and disinformation, raising concerns about the ethics of generative AI.

In 2021, OpenAI launched DALL-E, a neural network that can create original, realistic images and art from textual description. It can combine concepts, attributes, and styles in novel ways. A year later, Midjourney was launched by the independent research lab Midjourney. Also in 2022, Stable Diffusion, an open-source machine learning model developed by LMU Munich, was released that can generate images from text, modify images based on text, or fill in details in low-resolution or low-detail images.

OpenAI launched ChatGPT in November 2022 as a fine-tuned version of the GPT-3.5 model. It was developed with a focus on enhancing the model’s ability to process natural language queries and generate relevant responses. The result is an AI-powered chatbot that can engage in meaningful conversations with users, providing information and assistance in real-time. One of the key advantages of ChatGPT is its ability to handle complex queries and provide accurate responses. The model has been trained on a vast corpus of data, allowing it to understand the nuances of natural language and provide contextually relevant responses.

Today’s launch of OpenAI GPT-4 marks a significant milestone in the evolution of generative AI!

This latest model, GPT-4, is capable of answering user queries via text and image input. The multimodal model demonstrates remarkable human-level performance on various professional and academic benchmarks, indicating the potential for widespread adoption and use. One of the most significant features of OpenAI GPT-4 is its ability to understand and process image inputs, providing users with a more interactive and engaging experience.

Users can now receive responses in the form of text output based on image inputs, which is a massive step forward in the evolution of AI. Depending on the model used, a request can use up to 32,768 tokens shared between prompt and completion, which is the equivalent of about 49 pages. If your prompt is 30,000 tokens, your completion can be a maximum of 2,768 tokens.

Token limitations of GPT models with real-world scenarios
Token limitations of GPT models with real-world scenarios

Bing has already integrated GPT-4 and offers both, chat and compose modes for users to interact with the model. With the integration of GPT-4, Bing has significantly enhanced its capabilities to provide users with more accurate and personalized search results, making it easier for them to find what they are looking for.

The disruptive potential of generative AI is enormous, particularly in the retail industry. The technology can create personalized product recommendations and content, and even generate leads, saving sales teams time and increasing productivity. However, the ethical implications of generative AI cannot be ignored, particularly in the creation of disinformation and fake news.

To sum up, generative AI is here to stay, and companies must evaluate and implement strategies swiftly. As generative AI technology advances, so do the ethical concerns surrounding its use. Therefore, it is critical for companies to proceed with caution and consider the potential consequences of implementing generative AI into their operations.

Are you already using generative AI for a more productive workflow?

What improvement do you expect from OpenAI GPT-4 in this regard? I look forward to reading your ideas in the comments to this LinkedIn post:

„GPT-4 Launches Today: The Rise of Generative AI from Neural Networks to DeepMind and OpenAI“ weiterlesen

GPT-3: A Leap in Language Generation But Not True AGI– Insights from Decisively Digital

Explore the intricate relationship between AGI and GPT models like OpenAI's GPT-3, as revealed in the much-awaited book "Decisively Digital."
Explore the intricate relationship between AGI and GPT models like OpenAI’s GPT-3, as revealed in the much-awaited book „Decisively Digital.“

Artificial Intelligence (AI) has been making significant strides in recent years, particularly in the realm of generative AI. Among these advancements, OpenAI’s GPT-3 (Generative Pre-trained Transformer 3) has emerged as a groundbreaker. While its language generation capabilities are astonishing, the question remains: Are we any closer to achieving Artificial General Intelligence (AGI)? In this article, we’ll explore the complex world of GPT-3, its potential, and its limitations, as discussed in my forthcoming book Decisively Digital.

The Evolution of Generative AI and GPT-3’s Arrival

Generative AI has seen considerable growth in recent years. OpenAI first introduced GPT-3 in a research paper published in May and subsequently initiated a private beta phase. Selected developers have been granted access to further explore GPT-3’s capabilities. OpenAI has plans to turn this tool into a commercial product later this year, offering businesses a paid subscription to the AI via the cloud.

The Capabilities of GPT-3

The evolution of large language models like GPT-3 is worth examining in the context of Natural Language Processing (NLP) applications. From answering questions to generating Python code, GPT-3’s use cases are expanding by the day. Generative AI has been escalating at an unprecedented rate. OpenAI’s recent launch of GPT-3 has created a buzz in both the tech community and beyond.

The software has moved into its private beta phase, with OpenAI planning to offer a cloud-based commercial subscription later this year. This move marks a significant stride toward integrating GPT models into business applications, bringing us one step closer to the AGI GPT reality.

The Marvel of GPT-3: A Milestone in AGI Evolution?

GPT-3 is a machine-learning model with an impressive 175 billion parameters, making it capable of generating astonishingly human-like text. It’s been applied to numerous tasks, from generating short stories to even coding HTML. These capabilities have been turning heads and inciting discussions around AGI GPT models. But is it all it’s cracked up to be?

GPT-3’s predecessor, GPT-2, laid the foundation for the current model. While the underlying technology hasn’t changed much, what distinguishes GPT-3 is its sheer size—175 billion parameters compared to other language models like T5, which has 11 billion parameters. This scale is a result of extensive training on data largely sourced from the internet, enabling GPT-3 to reach or even surpass current State-Of-The-Art benchmarks in various tasks.

The Limitations and Weaknesses

Despite its staggering capabilities, the GPT-3 model is not without its flaws. Despite its human-like text generation capabilities, GPT-3 is still prone to generating hateful, sexist, and racist language. It’s a powerful tool but lacks the genuine smarts and depth that come with human cognition. In essence, while the output may look human-like, it often reads more like a well-crafted collage of internet snippets than original thought.

Most people tend to share positive examples that fit their bias towards the machine’s language „understanding.“ However, the negative implications, such as the generation of offensive or harmful content, need to be considered seriously. For example, GPT-3 has been found to generate racist stories when prompted with specific inputs, which raises concerns about the technology potentially doing more harm than good.

Not Quite AGI

Many have been quick to label GPT-3 as a stepping stone towards AGI. However, this might be an overestimation. GPT-3 can make glaring errors that reveal a lack of common sense, a key element in genuine intelligence. As OpenAI co-founder Sam Altman notes:

„AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.“

Sam Altman, CEO, OpenAI

Decisively Digital: The AGI GPT Discourse

My upcoming book Decisively Digital devotes an entire chapter to the role of GPT-3 in business and its potential to serve as a stepping stone toward AGI. From automating customer service to generating insightful reports, GPT-3 offers a wealth of opportunities for enterprises. However, the book also delves into the ethical considerations and potential pitfalls of adopting this powerful technology.

Concluding Thoughts: AGI GPT—A Long Road Ahead

While GPT-3 serves as an intriguing glimpse into the future of AGI, it is just that—a glimpse. We have a long road ahead in the quest for AGI GPT models that can mimic true human intelligence. As we navigate this fascinating journey, a balanced perspective is crucial.

To stay updated on these critical topics and much more, connect with me on Twitter and LinkedIn, and be on the lookout for the release of Decisively Digital.