Google is updating its AI tools, including the Gemini large language model (LLM, the underlying technology behind conversational AI like ChatGPT), to allow users to turn off visible watermarks on AI-generated images, videos, and music. This move gives creators more control over the aesthetics of their AI outputs, even as new research independently demonstrates a significant leap in AI's capacity to solve complex mathematical problems, underscoring the rapid evolution of artificial intelligence.
The change from Google, a major player in the AI space, means that the 'sparkle' icon previously appearing in the bottom-right corner of AI-generated media can now be toggled off. This applies to content created with Gemini and Google's AI video generator, Flow. Importantly, while the visible watermark can be removed, invisible benchmarks or metadata will still be embedded within the files. These hidden identifiers are crucial for tracking the origin of AI-generated content and distinguishing it from human-created work, addressing growing concerns about deepfakes and misinformation.
Concurrently, independent research published on arXiv reveals a surprising new capability: large language models are now autonomously solving open mathematical conjectures. A new benchmark called OEIS Open, built on 492 unproven mathematical statements from the Online Encyclopedia of Integer Sequences (OEIS), was used to test these LMs. Researchers found that LMs, given minimal tools, could resolve 147 of these conjectures, achieving a 30% success rate with a budget of $50 per attempt.
This mathematical breakthrough is particularly noteworthy because these conjectures are not trivial problems. While their broader mathematical significance is still being assessed, they represent genuine, unsolved research questions. The study also tested a subset of 100 conjectures, OEIS Open Lite, where the best current LM scored 44% with a higher budget of $200 per attempt. Interestingly, providing LMs access to a vast archive of 476,000 mathematical papers from arXiv did not improve performance, nor did more sophisticated 'agent loops,' suggesting the models' inherent reasoning capabilities are driving the success.
The Google watermark update reflects a common tension in AI development: balancing creative freedom with accountability. Creators often want their AI-generated content to look professional and unbranded, especially for commercial or artistic applications. Google's decision to allow watermark removal suggests a response to user demand, while the retention of invisible metadata aims to provide a safety net for content provenance. This mirrors a broader industry trend where companies grapple with how to best integrate AI into creative workflows while mitigating potential misuse.
The mathematical research, on the other hand, highlights the accelerating pace of AI's cognitive development. While solving math problems might seem niche, it demonstrates a sophisticated form of logical reasoning and pattern recognition that goes beyond merely generating text or images. This ability to autonomously tackle open research problems at a modest cost indicates that LMs are evolving from sophisticated prediction machines to tools capable of genuine scientific discovery, even in fields traditionally considered exclusive to human experts.
This combination of developments signals a nuanced phase in AI's integration into society. Google's move empowers creators and normalizes AI-generated content, potentially increasing its presence across digital platforms. Simultaneously, the mathematical breakthroughs suggest a future where AI isn't just a tool for automation or creativity, but a partner in scientific inquiry, capable of advancing knowledge in fundamental fields. The implications are profound, from accelerating research in areas like materials science and drug discovery to potentially redefining the role of human intellect in certain academic disciplines. This also raises questions about intellectual property and attribution when AI systems are the primary solvers of complex problems.
What to watch next is how other major AI developers respond to Google's watermark policy. Will other companies like OpenAI or Meta follow suit in offering similar controls over visible AI identifiers? Simultaneously, the mathematical community will be closely observing the impact of LMs on new conjectures. Will these models continue to improve their success rates, and will they eventually tackle problems of even greater mathematical significance? The interplay between user control, ethical safeguards, and expanding AI capabilities will define the next chapter of artificial intelligence.
