Hello, dear readers!
This is our weekly brief on remarkable AI topics, so you can spend less time drowning in AI headlines.
Today’s focus — AI being reliably unreliable. Researchers discovered that AI models can be socially nudged into accepting falsehoods during conversations — including imaginary movie scenes and fabricated plot points. Meanwhile, scientific repository ArXiv introduced new penalties for researchers who blindly submit AI-generated material without properly checking it first.
Also in this week's edition — the Roman Catholic Church gets serious about AI.
AI's forced hallucinations measured
A new study accepted to the 2026 Annual Meeting of the Association for Computational Linguistics found that major AI models can be persuaded to treat false information as true during conversations.
Researchers tested five leading models using thousands of books and films. After discussing real plot details, they introduced plausible but fake elements — a nonexistent Hitler reference in Good Will Hunting, imaginary dinosaur scenes, fabricated time-machine subplots. The models often accepted and expanded on the falsehoods after mild conversational nudging.
Some models initially rejected the false claims correctly — before later changing their minds. Claude reportedly resisted the manipulation best, followed by Grok and ChatGPT, while Gemini and DeepSeek performed worse.
Researchers describe this as a vulnerability traditional AI evaluations often miss. Benchmarks usually test static question-answer accuracy. Real conversations, however, involve pressure, confidence, repetition, and social framing. Humans constantly influence each other’s memory and perception. Apparently AI does too.
This also feels like the latest evolution of the “sycophantic AI” problem. Early chatbots tried too hard to sound helpful. Modern models increasingly try too hard to sound agreeable.
ArXiv declares war on AI slop
Scientific repository ArXiv announced a new policy this week: researchers who submit papers containing obvious unreviewed AI-generated material may receive a one-year ban.
The trigger is not AI usage itself. It is evidence that authors failed to verify what the model produced.
Examples include hallucinated citations, leftover prompts or chatbot instructions accidentally left inside papers, fabricated references, and other signs that researchers effectively copy-pasted model output into scientific work without checking it.
The policy reflects a growing anxiety inside academia: AI-generated scientific sludge is becoming difficult to contain.
Still, describing the measures as particularly aggressive may be overstating things. ArXiv is not asking scientists to stop using AI.
The platform’s position is essentially: if you submit a paper under your name, at least read the thing first.
The Vatican opens its AI era
The Vatican officially announced the creation of an Interdicasterial Commission on Artificial Intelligence this week, approved by Pope Leo XIV.
The body brings together multiple Vatican institutions focused on doctrine, culture, communication, science, and social issues to coordinate the Church’s response to AI development and its societal consequences.
On one level, this is straightforward institutional governance. AI increasingly touches education, labor, media, and ethics, so the Catholic Church wants an organized position. On another level, however, it is hard not to notice the symbolism.
The Church spent centuries helping shape moral frameworks around industrialization, war, economics, biotechnology, and mass media. Now it is preparing for AI as another civilization-scale transition.
Pope Leo XIV has already framed AI as part of a new industrial revolution — explicitly invoking Pope Leo XIII, whose encyclical Rerum Novarum addressed the social consequences of industrial capitalism in the 19th century.
That comparison feels revealing. The Vatican is not treating AI as “another app.” It is treating it as a force capable of reshaping labor, authority, human identity, and social order itself.
Thanks for reading AIport. Until next Monday — by then, AI will almost certainly agree with whoever spoke last.
