Grok-3 AI Model Announced as Next-Generation Contender

AI Hallucinates While Colorizing History

While artificial intelligence (AI) technologies are rapidly advancing in visual content production and editing, they are also revealing some of their limitations and interesting behaviors. Recently, a test conducted on leading image editing models, specifically trained to automatically colorize historical black-and-white photographs, yielded unexpected results. The models made additions that could be described as 'hallucinations' in the task of colorizing the famous Solvay Conference photograph from 1927, which brought together the brightest names in quantum physics.

Solvay Conference: A Test Scene

One of the most iconic photographs in the history of science, the 1927 Solvay Conference frame shows giants of the era such as Albert Einstein, Niels Bohr, Marie Curie, and Werner Heisenberg together. This historic meeting, also covered by TÜBİTAK Bilim Teknik, became a perfect testing ground for AI models. The models were asked to colorize this black-and-white photograph in a way that was consistent with historical facts. However, the results went beyond expectations.

Color Palettes and Objects Diverging from Reality

Upon examining the images produced by the models that underwent the test, a series of inconsistencies and imaginary additions were noticeable. For example, the jacket and tie worn by Albert Einstein took on bright and vivid colors that did not exist in reality. Some models added furniture or decorative elements to the background that were inconsistent with the historical context. One model even placed an object (a book or a glass) in the hand of one of the scientists in the photo, which was not present in the original frame. This phenomenon, referred to as 'seeing visions' or 'hallucination,' is interpreted as the models overgeneralizing patterns from their training datasets or attempting to fill missing information with their own creativity.

The Balance Between Creativity and Accuracy in AI

This situation highlights a fundamental dilemma of artificial intelligence: the balance between creativity and historical/social reality. As noted on Wikipedia, artificial intelligence aims to mimic cognitive functions unique to human intelligence. However, in this mimicry process, it can sometimes lack the contextual and historical awareness that humans possess. While creativity in image generation is valuable for artistic projects, accuracy and fidelity should be the most important criteria when processing a historical document.

As emphasized in the Ministry of National Education's Ethical Declaration on Artificial Intelligence Applications, technology should be used solely to support objectives and enhance quality. When processing historical content, this concept of 'quality' must include the accuracy of information and its non-manipulated state. AI's arbitrary addition of colors or objects to historical photographs carries the risk of inadvertently distorting the historical narrative and paving the way for the spread of misinformation.

The Role of Google Gemini and Other Assistants

Personal assistants mentioned in web sources, such as Google's AI assistant Gemini, promise to help users with writing, planning, and brainstorming. The developers of these tools state that they take user feedback seriously and continuously improve their tools. Similarly, image processing models also need to be continuously trained and improved with user and expert feedback to minimize such 'hallucination' errors. The tests demonstrate how vital this development process is.

Implications and Warnings for the Future

The Solvay Conference test reminds us that, alongside the exciting potential of AI technologies, they must be used carefully, especially in fields requiring representation of reality. In the future, focusing on the following points in the development of such models may be beneficial:

• Contextual and Historical Accuracy Checks: