OpenAI Begins Testing Ads in ChatGPT — A New Revenue Model to Sustain the Free Tier

On May 7, OpenAI began testing ad displays in ChatGPT. The company describes the move as "part of our effort to keep access free," saying the ads will be clearly labeled and shown independently of response content. OpenAI emphasizes privacy protection, and users will also have controls to manage their ad settings.

ChatGPT already reaches hundreds of millions of monthly active users, and the strategy of monetizing while preserving the free tier mirrors the free model pioneered by search engines. The key going forward will be striking a balance that doesn't harm the user experience. Advertising models for AI chatbots have few precedents, making this an experiment the entire industry is watching closely.

OpenAI's Realtime Voice Models Achieve GPT-5-Level Reasoning

Also from OpenAI, three new realtime voice models are now available through the API: the reasoning-oriented GPT-Realtime-2, the translation-specialized GPT-Realtime-Translate, and the transcription model GPT-Realtime-Whisper — all optimized for realtime processing.

The most notable of the three is GPT-Realtime-2, which OpenAI says possesses "GPT-5-level reasoning capability." If voice conversations can deliver reasoning on par with text-based models, the range of applications — customer support, realtime translation, voice assistants, and more — would expand dramatically. GPT-Realtime-Translate supports more than 70 languages, while GPT-Realtime-Whisper provides realtime speech transcription.

ChatGPT Gains a "Trusted Contact" Safety Feature

On the same day, OpenAI added a "Trusted Contact" feature to ChatGPT. When ChatGPT detects signs of serious self-harm concerns, it sends a notification to trusted people the user has registered in advance. The feature is opt-in and designed with privacy in mind.

AI chatbots are increasingly involved in users' mental health, and implementing safety features like this has become an industry-wide trend. Google and Meta are also strengthening similar safety measures, and questions about where platform responsibility lies are coming to the fore.

EU Delays Most of Its AI Regulation — High-Risk Classification Now Applies from 2027 Onward

The European Union (EU) has reached agreement on the "Digital Omnibus on AI," a package that simplifies its AI regulation. It pushes the compliance deadlines for high-risk AI systems to late 2027 through 2028 and eases requirements for small and medium-sized enterprises. However, labeling obligations for deepfakes and AI-generated text will still take effect as scheduled in August 2026. The package also explicitly bans "nudification" apps.

The EU's AI Act has drawn attention as the world's first comprehensive AI regulation, but the agreement highlights both industry complaints that the regulatory burden is too heavy and the reality that legislation cannot keep pace with rapid technological progress. The delay is only a temporary reprieve, and stricter rules are ultimately expected to apply.

IMF Warns of a "Systemic Shock" Risk to Finance from New AI Models

The International Monetary Fund (IMF) has warned that the latest AI models could deliver a "systemic" shock to the financial system. According to reporting by the FT, the IMF points in particular to the rapid adoption of AI in high-frequency trading, risk assessment, and credit scoring as a possible trigger for unforeseen chain reactions.

The concern is that as AI comes to drive a large share of decision-making in financial markets, herding behavior — "everyone facing the same direction" — becomes more likely. The IMF is calling for stronger international surveillance, citing the possibility that systemic risks like those of the 2008 financial crisis could re-emerge in new forms through AI.

Anthropic Auto-Interprets LLM Internal Representations with "NL Autoencoders"

A new paper from Anthropic's research team, "NL Autoencoders Produce Unsupervised Explanations of LLM Activations," published on transformer-circuits.pub, proposes a method for automatically interpreting LLM internal representations in natural language.

Interpretability research in machine learning has traditionally relied heavily on manual annotation, but this approach can reportedly attach natural-language explanations to LLM activation patterns in an unsupervised manner. Making it possible to explain why an LLM produced a given output is an important step toward improving AI reliability and safety. Anthropic has invested heavily in interpretability research and has become a leader in the field.

---