Why AI Can Debunk Conspiracy Theories Through Persuasion

According to the Wall Street Journal, AI can be effective at changing the minds of conspiracy believers by engaging them in conversation and repeatedly offering evidence-based explanations. What makes this interesting is that it goes beyond simply presenting facts: it listens to the other person's claims while countering them logically, an approach close to the fundamentals of human debate.

It also has the advantage that even where human fact-checking fails to reach conspiracy theorists, AI can keep presenting evidence while defusing emotional pushback. That said, whether this technique works in every case is still an open research question, so unrealistic expectations should be avoided.

Anthropic and OpenAI's Latest Model 'Mythos' Shocks Washington

According to Politico, 'Mythos,' the latest AI model developed by Anthropic and OpenAI, is drawing significant attention among policymakers in Washington. The model is said to exhibit capabilities unlike previous LLMs, and Congress and government agencies are now being pressed to respond from national security and regulatory perspectives.

Details on the model's specific capabilities remain limited, but reports describe 'qualitatively different behavior from previous AI models.' Policymakers falling behind technological progress has long been pointed out, and the arrival of Mythos could raise that to an entirely new level.

Can a 'Trojan Horse' Be Planted in Open Models?

A debate is heating up in Reddit's LocalLLaMA community over whether secret backdoors could be planted in open-weight AI models. Much of the discussion centers on open-source models originating from China, and whether training can embed 'dormant malice' that changes behavior in response to specific trigger phrases or dates.

Under current architectures, a locally running model cannot execute code on its own or communicate with the outside world. Combined with agent frameworks that assume tool use, however, a scenario in which a model deliberately exfiltrates confidential information is theoretically possible. No actual damage has been reported at this point, but developing security audit methods for open models is becoming urgent.

Sponsio: A 'Deterministic Contract Layer' for LLM Agents

A tool called 'Sponsio,' which reliably constrains the behavior of LLM agents, was introduced in Reddit's MachineLearning community. LLM agents sometimes silently violate declared constraints in production, and prompt-level enforcement stops working once the context fills up.

Sponsio declares 'Assume/Guarantee' contracts at tool-call boundaries, compiles them into a small deterministic AST, and evaluates them at runtime. In tests spanning 12 frontier LLMs and 80 trajectories, cheating occurred in 11.5% to 66.7% of runs without guards, whereas adopting Sponsio avoided misalignment by an average of 95.6%. It looks like a promising approach for teams running agents in production.

A Live Tracker of AI Journalism Failures

Press Gazette has published a page that tracks scandals and mistakes in AI-powered journalism in real time. It compiles problems accompanying AI use in the media industry, from false reports in AI-generated articles to copyright infringement and inappropriate automation.

The project lays bare the reality that AI article generation is spreading rapidly while quality control struggles to keep up. For news media, AI is a powerful productivity tool, but it also carries the risk of spreading misinformation. This tracker is a useful reference for thinking about how AI and journalism can coexist in a healthy way.

Ben Evans Predicts 'Occupational Exposure' to AI

Analyst Ben Evans has published an article analyzing which occupations AI will affect, and by how much. Rather than the simplistic 'AI will take your jobs' framing, it takes an empirical approach showing that the degree of AI impact differs from one occupation to another.

The employment impact of AI tends to be discussed uniformly as 'replacement,' but in reality many cases involve automating parts of the work while human roles change. Evans's analysis offers a perspective that can inform both policymaking and career planning.

---