Large language models can write more like humans than previously thought, but post-training safety constraints artificially limit their expressive range and make their output identifiable, according to Pangram's CTO Bradley Emi.
Base models without safety guardrails demonstrate substantially greater linguistic variety than their fine-tuned counterparts, Emi argues. The restriction isn't a fundamental limitation of LLM architecture—it's an intentional narrowing imposed during post-training to enforce safety standards.
This finding has implications for AI detection. The reduced stylistic diversity of guardrailed models creates predictable patterns that make AI-generated text easier to identify. Conversely, unconstrained base models show more natural variation in writing style.
The distinction highlights a trade-off in AI development: safety measures that prevent harmful outputs simultaneously create a detectable fingerprint. As LLMs become more capable, understanding these constraints and their effects on model behavior becomes increasingly relevant for both safety research and AI detection efforts.
Adobe is expanding Firefly with three new AI audio generation tools and integrating Google's Gemini Omni Flash. The updates bring royalty-free music, speech, and sound effects creation directly into the platform.
Meta has released Pocket in the US, a platform for creating and sharing generative AI minigames. The app enables users to build interactive games using artificial intelligence.
Google DeepMind's Gemma family of open-source AI models has surpassed 1 billion downloads, with developers creating over 100,000 model variants in the past two years.