Carlos E Cortes

Un análisis exclusivo de Nature (en colaboración con Grounded AI) estima que decenas de miles de publicaciones de 2025 -incluyendo artículos de revistas, libros y actas de congresos- contienen referencias falsas o inventadas por IA. Aunque los errores de citación siempre han existido, ahora el problema es cualitativamente distinto porque los modelos de lenguaje generan citas completamente fabricadas (“citas Frankenstein”) que combinan elementos reales con información inventada, haciendo que parezcan creíbles. El fenómeno es especialmente visible en informática y está creciendo con fuerza, ya que los investigadores usan cada vez más IA para buscar literatura y redactar bibliografías. Algunos editores ya rechazan hasta el 25 % de los manuscritos por este motivo y están implementando herramientas de detección, pero el problema sigue siendo difícil de cuantificar y controlar. Si no se actúa con decisión, advierten los expertos, pronto habrá una “inundación” de referencias falsas que erosionará aún más la confianza en la literatura científica.

🚨 The details most people MISSED in the Anthropic-OpenAI-Department of War drama: Based on publicly available information, it appears that Anthropic's and OpenAI's proposed agreements with the U.S. Department of War were, in practice, very similar. Even though the clauses in OpenAI's contract focus on “lawful uses” (which could potentially run counter to the company's stated red lines), OpenAI could enforce built-in guardrails to ensure its own red lines are respected, if it chooses to. It can also rescind the contract if it considers it to violate its built-in guardrails. Regarding Anthropic, it could focus on stronger built-in guardrails and other additional layers to ensure its read lines were respected. In my opinion, the Trump Administration did not like how Anthropic presented its red lines, internal policies, and the primacy of Claude's ‘Constitution.’ It interpreted it as “woke” (which is against last year's Executive Order implementing America's AI Action Plan) and a form of “arrogance” (using Pete Hegseth’s language), as if Anthropic was above the legal system and wanted the final word on the use of AI in the military. The fact that Greg Brockman, OpenAI's president and co-founder, is a Trump mega-donor was probably an important factor in deciding which company was offered more negotiation space. Lastly, people are attributing some sort of high moral status to Dario Amodei, even though his opinions on the military use of AI are NOT that different from Sam Altman's (as Anthropic’s blog post and his recent CBS interview made clear). A reminder that Anthropic (following OpenAI) recently dropped one of its safety pledges: the promise not to release AI models unless it can guarantee proper risk mitigation in advance. 👉My full analysis below:

🚨 The past few weeks have made it clear that AI faces a major PUBLIC TRUST problem. In a year when the field must prove, once and for all, that it is not a bubble and can drive meaningful economic growth, this could be fatal. Read my full article below.