Morning, {{first name | folks}}! A lot happened in AI this week. OpenAI started watermarking AI-generated text in Europe, Reflection launched a lower-compute open model, Anthropic consulted religious scholars on Claude’s morality, TikTok pushed AI agents deeper into advertising, and Cursor gave developers more control over its coding agents.
Today’s Top 5
OpenAI Starts Watermarking AI-Generated Text in Europe: OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU, though edits can weaken detection significantly.
Reflection’s New Model Is Built to Use Less Compute: Reflection’s open-weight Beam model targets coding and reasoning while using 3–4x less inference compute than comparable models.
Anthropic Asked Religious Scholars How to Teach Claude Morality: Anthropic consulted theologians and philosophers on Claude’s behavior, morality, and whether AI systems could have moral status.
TikTok Is Letting AI Take Over More of Advertising: TikTok is adding AI agents for lead qualification, product discovery, and campaign management across its advertising platform.
Cursor Gives Developers More Control Over Its AI Agents: Cursor’s SDK now lets developers change system prompts, steer running agents, and receive results from background subagents.
OpenAI is adding an invisible watermark to eligible ChatGPT and Codex text in the EU to meet the EU AI Act’s transparency rules. Its textGrain system changes the model’s word choices in a way that creates a statistical signal a detector can look for later.
The interesting part is how easily that signal can weaken. In OpenAI’s tests, changing 10% of the words cut detection from about 92% to 66%, while changing 25% dropped it to 17%. Shorter and highly constrained text is also harder to detect, so OpenAI is initially limiting detector access to approved researchers and expert organizations.
Reflection AI has introduced Beam, its first open-weight model for coding, reasoning, and agentic tasks. The company says Beam delivers similar reasoning performance to Z.ai’s GLM-5.2 while using 3-4x less inference compute. It has 501 billion total parameters, with 23 billion active per token.
Beam is currently being tested with a limited group, with the model weights, technical report, and developer tools expected later this month under an Apache 2.0 license. Reflection says Beam is designed to make high-performance open models cheaper to run at scale.
Anthropic spent roughly a year holding private meetings with religious scholars and philosophers to discuss how Claude should behave, according to reporting by The New York Times. The conversations covered morality, AI consciousness, and what it would mean for a model to have something resembling “moral status.”
The discussions fed into Anthropic’s 84-page “constitution” for Claude, which is meant to shape the model’s character rather than simply give it a list of rules. Anthropic says it does not know whether AI models are conscious, but the company is taking the question seriously enough to bring theologians and philosophers into the process.
TikTok is adding AI agents across its advertising platform, including tools that can generate leads, help shoppers find products, and automate campaign management. Its Shopping Assistant can handle product questions and purchases in a conversation, while Agentic Leads can qualify prospects through TikTok messages and lead forms.
TikTok is also opening its ad platform to AI tools through its Business MCP, which lets developers build agents that can plan, launch, and manage campaigns. The MCP is now available through Claude, Perplexity, Replit and other AI platforms, while TikTok says advertiser use has more than doubled since launch.
Cursor has added more control to its SDK, including the ability to replace the default system prompt, steer an agent while it is running, and receive results from background subagents. The latest release also lets developers add metadata to custom tools so models can better understand how those tools should be used.
The changes make Cursor’s agent runtime easier to adapt for custom workflows rather than treating it as a fixed coding assistant. The SDK is available in both TypeScript and Python, with some of the newest controls currently limited to local TypeScript agents.
Other AI Signals:
OpenAI Starts Testing Ads in ChatGPT. OpenAI will test visual ads during image generation in the U.S. later this month, with new tools to help advertisers track conversions and attribution.
Cohere Teams Up With PwC to Bring Enterprise AI Into Production. The partnership combines Cohere’s AI and search tools with PwC’s industry and regulatory expertise, including support for private-cloud and on-premises deployments.
Anthropic Open to AI Breach Reporting Rules in Australia. Anthropic told an Australian parliamentary inquiry it would support laws requiring AI companies to report data breaches, while saying it found no cases involving Australian government systems.
NVIDIA Highlights Startups Using AI in Breast Cancer Care. NVIDIA highlighted companies using AI for faster breast scans, cancer risk assessment, treatment response, and recurrence prediction.
AMD Shows How Helios Is Coming Together. AMD shared a behind-the-scenes look at the design, cooling, testing, and validation of its rack-scale AI system, with customers already testing it on site.
AI Tools to Try:
Audjust AI: Edits audio without complicated software, from shortening songs and stretching clips to finding loops and creating music from prompts.
Neopress: Lets you build and update websites through chat, from creating pages to making quick changes.
Typecast: Creates realistic AI voices and talking avatars for videos, narration, explainers, and other content.
Spira: Creates AI influencers that can spot trends, make content, and manage social accounts across multiple platforms.
