These documents generally have a number of people writing them, and go through many rounds of scrutiny and review. That's why even apart from the AI use, you can hear different voices in the writing. John Paul II and Benedict XVI were exceptions, being academics who'd written many books prior to becoming Pope, and so their encyclicals. If I had to guess, it seems likely for Magnifica Humanitas that some aide used AI for parts they were responsible for in the final drafting, which is a shame.
With respect to having Anthropic there, I agree it sucks. The silver lining is (1) most people who will hear about the encyclical will have no idea Anthropic was there, and (2) the encyclical itself was at explicitly at odds with Olah's anthropomorphizing in his remarks.
Both these are disappointments, but I read them as blunders/fumbles moreso than hypocrisy or bad intent. He has many advisors, and likely the wrong ones were more influential in these cases.
Thanks for this -- useful info about how the encyclical was put together! I do think that much of the content of the encyclical was *really good*. Overall, I'm very happy that he/his team wrote this! :-)
In all honesty, I haven't read this in its entirety and I will later but i've read enough for now. First, thank you for the info on Anthropic I had no idea on some of those things, I do think we gotta be fair to them and you might not be giving them enough credit. However that's a different conversation.
I just stumbled onto this subject by using claude for the first time last week. I got on substack to have conversations with people.
It was definitely a huge mistake to partner with Anthropic on the rollout of the encyclical. From a purely PR standpoint, it was maybe not such a bad move, but from an ethical standpoint…well, you kind of summed it up.
I’m also shocked (perhaps naively so) to find out that some of the encyclical was written by AI. Another very bad move. (And btw, I’m inclined to agree with you on the accuracy of Pangram. Just ran my last article through it and it came up 100% Human. Phew.)
All that said, the actual sentiment of the encyclical is pretty spot on, from its focus on the primacy of human dignity, to its warnings of AI’s many limitations and dangers, to its advocacy of “disarming” AI. That this important message is doubly tarnished by the PR stunt and the use of AI is absolutely…a shame.
Your frustration really radiates from this one, Émile. I empathise, only totally blind and numb people wouldn’t be frustrated in our current circumstances.
The AI “authorship” of the encyclical – I would be more sceptical and wouldn’t really believe that easily another AI-based detection tool. I think there might be many texts that can seem AI generated but actually are not. I think there is no way to tell definitely (unlike with generated images where there are some discerning patterns that are not found in “organic” images) and the accusations might be very unfair in some cases.
The whole content of the encyclical is so anti-unethical-AI-use that even if someone from Anthropic (oh the irony of the name!) was there at the event (will someone who does not follow the AI topic closely even notice or recognise the guy?), I am not sure this gives them much of an “endorsement” or that it would be that much helpful – they are still a company pushing/developing AI, and one of the Pope’s messages is, as I understand it, “we need to regulate/govern the AI development, use and adoption”, and this governance and regulation should come from other actors than the AI-promoting companies (such as Anthropic). So if people receive this message then hopefully they will not promote any AI company (whichever it is) and won’t be blind to the bad stuff Anthropic or other similar actors do. (Note: I have not read the encyclical and only have mediated information.)
I’ll also add that apparently the impact of the words of this encyclopedic is far greater bc the whole infrastructure of the Catholic church is supposed to follow this as guidance. It can pave the way for much bigger change as Jacob Ward explained regarding the previous Pope Leo and the minimum wage.
So more impactful than the visual of inviting Anthropic to participate (“partner” feels a bit strong), which I agree is problematic (see my other comment).
With that in mind, I feel like “shame on the Pope” is a but strong. You can do better is more appropriate. On balance I think he’s done far more good with this than bad, especially since the Vatican is a major world institution with a lot of influence that had come out very strong on the side of being pro-human, in comparison to other world leaders and regional institutions (like the EU) that are still swept up in the hype and ready to strip away regulations for fear of missing out (I’m looking at you EU).
At first I was disappointed and troubled to see Anthropomorphic up there, but @Jacob Ward’s take on Barathunde Thurston’s recent Life with the Machines provides an interesting perspective. He saw it as a diplomatic manoeuvre - a keep-your-enemies-closer type move (bear hug) - rather than a partnership or endorsement. I still think for a less-informed public that thinks Anthropic is a “good AI company” the optics worked in Anthropic’s favour, but the Pope was actually attacking their entire business model with the Tower of Babel comparison etc and I think putting the lie to their lofty self-image. It had to make him squirm inside. All that to say I’m not sure where I fully land on this but I agree more balance by bringing in many groups Timnit Gebru listed would hv been powerful, fair and effective —- really force Anthropic to confront the harms they’ve contributed to….
From a Pope PR perspective - he wants to be a global unifier, and will likely do his diplomatic best not to antagonise even when deeply critical. The tech broligarchy is ofc powerful, so what better way to give them a public dressing down with them right by your side?
Hi Emile. Appreciate the link and the righteous anger. 3 quick caveats:
#1 I don’t think the pope himself was directly responsible for AI use in the encyclical. These projects often tend to be multi-people affairs with dozens of people contributing, and the different paragraphs with very different rates of AI usage indicates that some people (not all, and probably? not Leo himself) used AI.
(I think I wrote that caveat fairly carefully in the article. I do not take responsibility for this).
#2 I think while some paragraphs register as 40-100% AI, most did not. It’s hard to say exactly how many because a) it varies from language to language and b) Pangram has more false negatives than false positives. Snopes below covers what I think is sort of a lower bound of maybe 6% based on Pangram high-certainty hits in English, in other languages it’s higher.
(I take some responsibility for not being more precise about it in the article, but I also wasn’t sure at the time).
#3 Note that “100% AI-generated” by Pangram just means that the words in that paragraph are arranged in a way that looks like they’re with very high probability AI origin. It *does not* mean that no human effort was put into the doc, or even that the human effort was minimal.
For example if a human carefully wrote the outline and iterated many times, but the writing itself is AI-generated, it’d still pick up as 100% AI by Pangram, Pangram’s not smart enough to figure out whether the ideas are human or synthetic. It’s also just a hard problem!
On the other hand, 100% AI generated also can’t mean that the writing was only light AI assistance (eg ask AIs to fix typoes, or even translate from Italian to English). The AI stylistic features will be much milder if that were the case.
this is why I used the word “substantially” in the subtitle, to distinguish AI assistance from very light proofreading to more substantial writing.
I did not explain this distinction at all in the article. I think this was in-hindsight obvious that many readers would be confused by this distinction. I’ll be more careful going forwards to make the distinction clear.
I take full responsibility for this specific oversight in my article.
I agree with everything you say here -- I didn't mean to imply that Leo wrote the entire encyclical (it's a team effort, as you note!), nor do I think that "100% AI-generated" means they copy-pasted what ChatGPT spit out. Thanks for these clarifications, though -- I've changed some wording to be more accurate! Really appreciate it.
If you want to read more, I think the Snopes coverage is fairly good here. With the advantage of time, they came to similar overall conclusions to me, but is also shorter and has many details that I didn’t cover.
Yeah, without more of a story around the invite, I can imagine a lot of reasons why an invitation is not endorsement. Maybe Olah is Catholic? In which case, commanding him to attend could have its own meaning.
I think you seriously misrepresent what Linch claimed and what the evidence suggests. For example, Linch never claimed that Leo himself used a chatbot in writing the encyclical. Snopes offers a good summary of the issues here, I think:
"Yet the Pope crafted an official contribution to Church doctrine using AI chatbots. Imagine a religious leader saying one thing yet doing another! :-)"
But Linch was careful to say that he didn't think Leo himself used AI, but that a subordinate may have when crafting parts of the encyclical. That may seem to be a subtle distinction, but considering how critical you seem to be of Leo himself here, I think it matters.
Obviously one could still critique Leo for not ensuring that such AI content entered the workflow--if, indeed, it did--but I think this would result in a much milder take than the one you present here.
There's also the question, of course, on whether the AI-detection tool is as ironclad as you think it is. I'm not saying that it isn't--I don't know--but I don't know that it warrants quite the confidence you place in it here. Greater epistemic humility is, I think, warranted.
Again, this doesn't mean I am sure that AI wasn't used. I just think the evidence is more ambiguous. And I just don't think the internet needs more hot takes, in general. Cool takes often present more of the truth, in my opinion.
Let me also add that, in general, I agree with and value your writing on substack. I'm not suggesting your general approach is significantly flawed or that your perspective on AI is wrong (since, broadly, I share it!) But I do think you misstep in this particular piece.
Has someone already created a network analysis of EA-figures, companies and thinktanks? They seem to increase and grow like crazy. It's hard to keep track.
Also has someone looked closer into Rutger Bregman? He seemed to be a leftwing progressive for a long time, but is closely connected to Peter Singer, 80000 hours and other EA-proponents
These documents generally have a number of people writing them, and go through many rounds of scrutiny and review. That's why even apart from the AI use, you can hear different voices in the writing. John Paul II and Benedict XVI were exceptions, being academics who'd written many books prior to becoming Pope, and so their encyclicals. If I had to guess, it seems likely for Magnifica Humanitas that some aide used AI for parts they were responsible for in the final drafting, which is a shame.
With respect to having Anthropic there, I agree it sucks. The silver lining is (1) most people who will hear about the encyclical will have no idea Anthropic was there, and (2) the encyclical itself was at explicitly at odds with Olah's anthropomorphizing in his remarks.
Both these are disappointments, but I read them as blunders/fumbles moreso than hypocrisy or bad intent. He has many advisors, and likely the wrong ones were more influential in these cases.
Thanks for this -- useful info about how the encyclical was put together! I do think that much of the content of the encyclical was *really good*. Overall, I'm very happy that he/his team wrote this! :-)
In all honesty, I haven't read this in its entirety and I will later but i've read enough for now. First, thank you for the info on Anthropic I had no idea on some of those things, I do think we gotta be fair to them and you might not be giving them enough credit. However that's a different conversation.
I just stumbled onto this subject by using claude for the first time last week. I got on substack to have conversations with people.
Are you one of them?
https://bwoodstock.substack.com/p/four-days-in-i-found-what-anthropic?utm_source=share&utm_medium=android&r=7ymdms
It was definitely a huge mistake to partner with Anthropic on the rollout of the encyclical. From a purely PR standpoint, it was maybe not such a bad move, but from an ethical standpoint…well, you kind of summed it up.
I’m also shocked (perhaps naively so) to find out that some of the encyclical was written by AI. Another very bad move. (And btw, I’m inclined to agree with you on the accuracy of Pangram. Just ran my last article through it and it came up 100% Human. Phew.)
All that said, the actual sentiment of the encyclical is pretty spot on, from its focus on the primacy of human dignity, to its warnings of AI’s many limitations and dangers, to its advocacy of “disarming” AI. That this important message is doubly tarnished by the PR stunt and the use of AI is absolutely…a shame.
Your frustration really radiates from this one, Émile. I empathise, only totally blind and numb people wouldn’t be frustrated in our current circumstances.
The AI “authorship” of the encyclical – I would be more sceptical and wouldn’t really believe that easily another AI-based detection tool. I think there might be many texts that can seem AI generated but actually are not. I think there is no way to tell definitely (unlike with generated images where there are some discerning patterns that are not found in “organic” images) and the accusations might be very unfair in some cases.
The whole content of the encyclical is so anti-unethical-AI-use that even if someone from Anthropic (oh the irony of the name!) was there at the event (will someone who does not follow the AI topic closely even notice or recognise the guy?), I am not sure this gives them much of an “endorsement” or that it would be that much helpful – they are still a company pushing/developing AI, and one of the Pope’s messages is, as I understand it, “we need to regulate/govern the AI development, use and adoption”, and this governance and regulation should come from other actors than the AI-promoting companies (such as Anthropic). So if people receive this message then hopefully they will not promote any AI company (whichever it is) and won’t be blind to the bad stuff Anthropic or other similar actors do. (Note: I have not read the encyclical and only have mediated information.)
You, Gebru and Becker are doing God’s work, literally. Down with these hubristic institutions!
I’ll also add that apparently the impact of the words of this encyclopedic is far greater bc the whole infrastructure of the Catholic church is supposed to follow this as guidance. It can pave the way for much bigger change as Jacob Ward explained regarding the previous Pope Leo and the minimum wage.
So more impactful than the visual of inviting Anthropic to participate (“partner” feels a bit strong), which I agree is problematic (see my other comment).
With that in mind, I feel like “shame on the Pope” is a but strong. You can do better is more appropriate. On balance I think he’s done far more good with this than bad, especially since the Vatican is a major world institution with a lot of influence that had come out very strong on the side of being pro-human, in comparison to other world leaders and regional institutions (like the EU) that are still swept up in the hype and ready to strip away regulations for fear of missing out (I’m looking at you EU).
At first I was disappointed and troubled to see Anthropomorphic up there, but @Jacob Ward’s take on Barathunde Thurston’s recent Life with the Machines provides an interesting perspective. He saw it as a diplomatic manoeuvre - a keep-your-enemies-closer type move (bear hug) - rather than a partnership or endorsement. I still think for a less-informed public that thinks Anthropic is a “good AI company” the optics worked in Anthropic’s favour, but the Pope was actually attacking their entire business model with the Tower of Babel comparison etc and I think putting the lie to their lofty self-image. It had to make him squirm inside. All that to say I’m not sure where I fully land on this but I agree more balance by bringing in many groups Timnit Gebru listed would hv been powerful, fair and effective —- really force Anthropic to confront the harms they’ve contributed to….
From a Pope PR perspective - he wants to be a global unifier, and will likely do his diplomatic best not to antagonise even when deeply critical. The tech broligarchy is ofc powerful, so what better way to give them a public dressing down with them right by your side?
Pangram kicks back 7:8(acts) Strong’s #1160.
That’s a cypher.
Hi Emile. Appreciate the link and the righteous anger. 3 quick caveats:
#1 I don’t think the pope himself was directly responsible for AI use in the encyclical. These projects often tend to be multi-people affairs with dozens of people contributing, and the different paragraphs with very different rates of AI usage indicates that some people (not all, and probably? not Leo himself) used AI.
(I think I wrote that caveat fairly carefully in the article. I do not take responsibility for this).
#2 I think while some paragraphs register as 40-100% AI, most did not. It’s hard to say exactly how many because a) it varies from language to language and b) Pangram has more false negatives than false positives. Snopes below covers what I think is sort of a lower bound of maybe 6% based on Pangram high-certainty hits in English, in other languages it’s higher.
(I take some responsibility for not being more precise about it in the article, but I also wasn’t sure at the time).
#3 Note that “100% AI-generated” by Pangram just means that the words in that paragraph are arranged in a way that looks like they’re with very high probability AI origin. It *does not* mean that no human effort was put into the doc, or even that the human effort was minimal.
For example if a human carefully wrote the outline and iterated many times, but the writing itself is AI-generated, it’d still pick up as 100% AI by Pangram, Pangram’s not smart enough to figure out whether the ideas are human or synthetic. It’s also just a hard problem!
On the other hand, 100% AI generated also can’t mean that the writing was only light AI assistance (eg ask AIs to fix typoes, or even translate from Italian to English). The AI stylistic features will be much milder if that were the case.
this is why I used the word “substantially” in the subtitle, to distinguish AI assistance from very light proofreading to more substantial writing.
I did not explain this distinction at all in the article. I think this was in-hindsight obvious that many readers would be confused by this distinction. I’ll be more careful going forwards to make the distinction clear.
I take full responsibility for this specific oversight in my article.
I agree with everything you say here -- I didn't mean to imply that Leo wrote the entire encyclical (it's a team effort, as you note!), nor do I think that "100% AI-generated" means they copy-pasted what ChatGPT spit out. Thanks for these clarifications, though -- I've changed some wording to be more accurate! Really appreciate it.
What happens when machine translation (e.g. French-English) is used, then the English is fed to Pangram? Would that affect the false-positive rate?
I missed the memo where these encyclicals aren't composed in Latin anymore. But I can't believe they're done as one big English doc.
My guess is that it affects the false-positive rate a little but I did try to backtest this nontrivially, and gave multiple pieces of evidence for why I think it's not a translation artifact. https://linch.substack.com/p/claude-author-of-the-humanitas?open=false#%C2%A7probably-not-a-translation-artifact
Since publishing the article, I also tested against translation from Spanish, and lighter tests in other languages.
Awesome, you answered that question.
If you want to read more, I think the Snopes coverage is fairly good here. With the advantage of time, they came to similar overall conclusions to me, but is also shorter and has many details that I didn’t cover.
https://www.snopes.com/news/2026/06/03/pope-leo-ai-encyclical
That is a lot of words to essentially say inviting a representative from Anthropic reflects poorly on the pope.
Given that the Holy See invites a devil’s advocate to critical deliberations suggest they are comfortable with handling alternatives.
The Pope didn’t write the encyclical, smart people who know AI did, that is why it is so good. He shaped what the position was.
Yeah, without more of a story around the invite, I can imagine a lot of reasons why an invitation is not endorsement. Maybe Olah is Catholic? In which case, commanding him to attend could have its own meaning.
Welp, I was brought up Lutheran and we were always told the Catholics are in league with the devil…
I think you seriously misrepresent what Linch claimed and what the evidence suggests. For example, Linch never claimed that Leo himself used a chatbot in writing the encyclical. Snopes offers a good summary of the issues here, I think:
https://www.snopes.com/news/2026/06/03/pope-leo-ai-encyclical/
What have I misrepresented?
Most notably you say this:
"Yet the Pope crafted an official contribution to Church doctrine using AI chatbots. Imagine a religious leader saying one thing yet doing another! :-)"
But Linch was careful to say that he didn't think Leo himself used AI, but that a subordinate may have when crafting parts of the encyclical. That may seem to be a subtle distinction, but considering how critical you seem to be of Leo himself here, I think it matters.
Obviously one could still critique Leo for not ensuring that such AI content entered the workflow--if, indeed, it did--but I think this would result in a much milder take than the one you present here.
There's also the question, of course, on whether the AI-detection tool is as ironclad as you think it is. I'm not saying that it isn't--I don't know--but I don't know that it warrants quite the confidence you place in it here. Greater epistemic humility is, I think, warranted.
Again, this doesn't mean I am sure that AI wasn't used. I just think the evidence is more ambiguous. And I just don't think the internet needs more hot takes, in general. Cool takes often present more of the truth, in my opinion.
Let me also add that, in general, I agree with and value your writing on substack. I'm not suggesting your general approach is significantly flawed or that your perspective on AI is wrong (since, broadly, I share it!) But I do think you misstep in this particular piece.
Has someone already created a network analysis of EA-figures, companies and thinktanks? They seem to increase and grow like crazy. It's hard to keep track.
Also has someone looked closer into Rutger Bregman? He seemed to be a leftwing progressive for a long time, but is closely connected to Peter Singer, 80000 hours and other EA-proponents
I think the likelihood that the translation to English was the AI use is high in this instance.