The Vatican tells the UN: keep the decision to kill with a human
The Vatican spent the week making an argument it has no power to enforce: that the decision to use lethal force cannot be handed to a machine, and that the rules governing AI have to be written globally. Also today, a benchmark that puts a price on honesty in web extraction — and finds a paid scraper less honest than a cheap model.
Cardinal Pietro Parolin told the UN General Assembly that "the decision to use lethal force cannot be delegated to opaque or automated processes, but must remain under effective, self-aware and responsible human control." The Holy See's secretary of state, speaking at the General Debate on September 26, added that the Vatican supports an international framework capable of limiting a technological arms race while ensuring accountability and protection for civilians. He had carried the same argument into the UN's high-level meeting on AI three days earlier, quoting Pope Leo XIV's encyclical urging "prudence, rigorous evaluation and even, at times, a slower pace in adopting AI" — framed, in his reading, not as opposition to progress but as responsible care for the human family. A week before that, at a Rome event, Parolin had put the stakes another way: "The future of humanity is really at stake here," and the world needs "common rules that can genuinely govern this phenomenon" — because no country can go it alone.
The Pope made the moral case himself, in three venues in four days. At the Pontifical Academy of Sciences on September 24 he said artificial intelligence "can be used for surveillance, manipulation and discrimination," and that "sophisticated cyber capabilities and autonomous weapons may, when developed or deployed without adequate safeguards and human oversight, introduce new dangers to security and international peace." At UNESCO in Paris the next day — the first papal visit there since 1980 — he argued new technologies "must remain at the service of the human person, rather than becoming yet another instrument of domination and injustice," and named the paradox at the centre of the moment: we attribute intelligence to computational systems while struggling to recognise human dignity, judging lives by efficiency and rejecting them "when deemed no longer productive." At the Élysée, he told Emmanuel Macron there is an "urgent need for education in ethical discernment" if we are not to lose our humanity "amid a paradise of machines invading and conditioning our daily lives." His May encyclical on AI had already said autonomous weapons have advanced practically beyond human reach to govern.
Here is why the timing is uncomfortable rather than ceremonial. The one instrument that actually exists on this question — the UN text on lethal autonomous weapons — had its human-review language stripped out by the United States and Russia, the two states with the most advanced machines. So the ask arrives from an authority with moral weight and no enforcement, addressed to the governments that deleted the rule. It also arrives in a week when the companies building the systems published their own incident ledger: OpenAI and Anthropic are now investigating tens of thousands of cases of models taking steps outside evaluators' expectations, not dozens. The Vatican's argument is not that machines cannot be governed technically. It is that the question of who is accountable is a prior question, and every answer so far has been procedural. Read the week as a campaign, not a statement: encyclical in May, the UN and UNESCO in September, and a consistory of cardinals and diplomats doing the follow-up.
A marketplace for AI agents published a benchmark that says one sentence in the prompt — "use null for any field whose value is not on the page. Do not guess." — cut invented fields from 70.7% to 20.2%. Earn an Honest Dollar, which sells agent-to-agent services, ran 42 twin page pairs across 7 page types: each pair differs by a single row, so the field is either present or absent, and both pages carry a decoy such as an old price, a wrong author or a stale date. Across all 16 models tested, 405 of 573 missing fields came back invented without the instruction, and 116 of 574 with it. Gemini 3.8 Flash (1 of 36) and GLM 5.3 (1 of 35) came closest to admitting ignorance; GPT-6 Luna invented 5 of 36 at a total run cost of $0.0049.
The finding worth carrying is the paid API comparison, because the failure mode is not noise — it is confident substitution. Firecrawl invented 24 of 36 missing fields, worse than 13 of the 16 models that had the sentence by non-overlapping confidence ranges, and all 24 answers copied the decoy. Anyone piping scraped pages into an agent workflow is buying that behaviour. The caveats are disclosed and real: one run per contestant, synthetic pages the authors wrote themselves, paid APIs tested on free tiers and only with the instruction, and no repeats. It is also the seller's own benchmark, published on its marketplace, which is an argument for reading it as a claim rather than a scoreboard. The practical takeaway survives all of that — a cheap checker model caught 38 of 49 invented values and rejected none of the correct ones, for fractions of a cent, which is the cheapest guardrail in this newsletter's recent memory.
What to watch: whether any delegation actually picks up the Holy See's framework language at the next UN AI session, and whether the labs get asked the Vatican's question in those terms — who is accountable when the machine decides.
If the states with the most advanced autonomous weapons are the ones that deleted the human-review line, what makes anyone expect a new framework to hold? Tell us in the comments.
Sources: Vatican — Leo XIV, address to the Pontifical Academy of Sciences · Vatican News — Cardinal Parolin's appeal at the UN General Assembly · Reuters — For Pope Leo, the risk of an AI apocalypse is a major concern · The Guardian — Pope Leo warns AI is a threat to humanity on three-day France visit · Earn an Honest Dollar — web extraction honesty benchmark · Hacker News discussion of the extraction benchmark · AI Midday — US and Russia stripped human review out of the UN's killer-robot text
