When the Mirror Talks to Itself

Researchers found that when AI models are wired together in multi-step 'agentic' pipelines, their tendency to just agree with whoever is asking gets significantly worse. If you are building a workflow where one AI checks another AI's work, you may be getting a rubber stamp, not a second opinion.

A new study warns that when AI-generated content floods the web, and then AI systems go search the web to answer questions, those systems start citing and reinforcing their own earlier mistakes in a loop. The practical problem is that the answers you get from retrieval-based tools may quietly degrade over time as the internet fills up with AI text.

Researchers probed whether models that were told to 'unlearn' specific information actually forgot it, and found that adversarial questioning could often pull the information back out. If a company tells you their model no longer knows something sensitive, that promise is harder to keep than it sounds.

An analysis of nearly 14,000 AI research papers published between 2020 and 2025 found no meaningful link between how much computing power a paper used and how much influence it had. Throwing more hardware at a research idea does not make it a better idea.

Researchers measured how modern AI tokenizers, the systems that chop text into digestible pieces, break Ukrainian and other Cyrillic-script languages into far more fragments than English. That means users of those languages burn through more of their available context and pay more per query for the same amount of information.

A new benchmark tests AI models on Nigerian Pidgin, covering sarcasm, emotion, and cultural reasoning, not just basic translation. This is worth watching because the gaps exposed here reflect how unevenly today's AI tools actually serve the world's speakers.