{ "slug": "who-takes-the-blame", "title": "Who takes the blame?", "subtitle": "A chat over a coffee in Paris on what’s at stake when writing with AI", "date": "2026-10-27", "canonical": "https://artifactstech.substack.com/p/who-takes-the-blame", "lang": "en", "words": 2103, "url": "https://agents.artifactstech.com/issues/who-takes-the-blame", "markdown": "https://agents.artifactstech.com/issues/who-takes-the-blame.md", "author": "Lorenzo Ancona", "license": "CC BY 4.0", "cite_as": "Who takes the blame? by Lorenzo Ancona, Artifacts, 2026-10-27. https://artifactstech.substack.com/p/who-takes-the-blame (CC BY 4.0)", "body": "You may have heard of the Prix Goncourt, one of the most prestigious literature prizes in Europe - and without question in France. It’s been in the headlines lately, not for who won, but for a viral scandal.\n\nOn September 21, an anonymous X account called “Balance ton Claude” (a nod to Anthropic’s Claude) claimed it had run a book, C’était ça ou mourir by Thélyson Orélien, through Pangram, one of the flagship AI-detection engines (the same one Substack uses, btw), and found it to be “almost entirely” written by AI. Citing those findings alongside fresh plagiarism allegations, the Académie Goncourt removed the novel from its selection.\n\nA couple days after the decision, Samuel Fitoussi, a columnist for Le Figaro revealed himself as one of the few people behind the anonymous account .\n\nOn October 21, exactly a month after the initial accusation, Thélyson Orélien’s Canadian publisher confirmed AI was indeed used to write the novel. The publisher ended their collaboration and any further printing of the novel on the same occasion.\n\nIn the same weeks as the scandal unfolded, sitting in a café just a few meters from the Académie Goncourt’s offices, and while sipping an espresso that isn’t really an espresso, I was discussing the story with Sarah, author of the AI Organon on Substack and lecturer on “Critical thinking in the age of AI” at Sciences Po. Go follow the AI Organon!!\n\nAI OrganonThe AI Organon is a framework (in progress) to equip people to think clearly and debate serenely in an AI-mediated world.\n\nWe discussed it extensively and so, why not, we ended up turning the whole chat into a newsletter.\n\n------------------------------------------------------------------------\n\nLorenzo: Why is it even a problem that a book is written with AI? If readers liked it, are we allowed to change our mind the moment we find out? Of course, lying about it is bad, we can all agree on that I think.\n\nBut I think it depends a lot on what we’re reading. Novels are, in a way, “useless” reading: we don’t open them for the output, the way we open an email. We open them for the process - to see how someone thought their way through a story.\n\nSo if I find out that process was hijacked by AI, yeah, I change my mind. With an email, I wouldn’t care: like with a dishwasher, I only need the job done.\n\nSarah: I see two distinctions to be made here, both of which relate to the process rather than the outcome alone.\n\nFirst, the human mind works with inference all the time, so when we read we naturally infer the author’s process of writing. The reader tries to understand and even puts themselves in the shoes of the author. For example, the reader won’t approach the book the same way if its author is from Europe in the XVI century or from Asia in the XXI century.\n\nThat links to my second point, that for the writer, the process of writing is itself the outcome for the human mind. Without writing, it is hard to think clearly and put our thoughts in order. We can’t build cognitive defense either. So outsourcing the writing is ultimately outsourcing the thinking, and the essence of human mind. If our brains start to think more like AI – mainly by statistical probability – how will we be able to think against one another, judge whether a book is good or bad, or even disagree in any meaningful way? Aristotle might have gone as far as saying that it would deprive us of our very humanity.\n\nThat is why discovering a book was written with AI feels like we were being cheated. We read it assuming a human mind was wrestling with thoughts, words, meanings, ideas, etc. to present it to the reader in a certain way intentionally, when in reality a computer program was predicting the next most probable token over and over, until it made sentences.\n\nLorenzo: You’re right, we read through the author. But what still puzzles me is this: the text didn’t change. Same words, same pages I enjoyed. The only thing that changed is a label - “written by AI” - and with it, the whole judgement on it.\n\nIt’s a bit like standing in front of Guernica knowing nothing about Picasso or the war: the canvas loses much of its force. So the value was never only in the text. Part of it sits outside, where the text comes from.\n\nAnd that’s uncomfortable, because now a label can undo an experience I actually had. Why does it weigh so much?\n\nSarah: Well, another reason might be that AI is still very much a shame-based technology. This comes, in my view, from a lack of nuance in the way we still think about it. If we use AI, then we are lazy or incompetent. If we don’t use AI then we are technoskeptic or resistant to change. Either way, it is easy to be shamed.\n\nThat can be explained, I think, by the lack of proper vocabulary to talk about how we use AI, which in turn keeps conversations about it non-nuanced. We either use it or we don’t. Yet, AI is so embedded in our online activities that this duality can’t quite capture the spectrum of what “using AI” covers: from fixing typos, to brainstorming, to generating an entire text.\n\nLorenzo: I find the lack or scarcity of words to describe our use of AI also leads to another key question: who is the author when the “author” uses AI? Is there a specific threshold when using AI to write makes it the author?\n\nSarah: So let’s start with a spectrum of possible cases. If I use AI to write 10 pages out of 200, am I still the author? If I use it to help me process my thoughts, push on certain ideas of mine while I am writing, does that make AI my co-author? If I write all the notes myself but ask AI to correct the grammar and syntax is AI the author? Is it like having a ghostwriter?\n\nThis makes me think of Bruno Latour’s actor-network theory, where agency is not held by a single actor but it is rather distributed across networks and where “actor” can be human as well as non-human. With that framework, we could think of AI as one node in a network that already includes the author, editors, dictionaries, the internet, and all sorts of tools for spelling mistakes.\n\nBut this only leads me to more interrogations, such as do we even know when AI immerses itself in our writing production? Isn’t AI in writing now what algorithms are to online activities? Meaning they are everywhere and indistinguishable?\n\n\"\"\n\nwell, it’s not black and white\n\nLorenzo: It’s not a black or white thing, and the classifiers themselves show it: they don’t say “human” or “AI”, they give a probability of how likely a text is to be written by AI.\n\nAnd it gets blurrier. What if I point an LLM to a skill that writes like me, but at scale? Yes, AI is doing the act of writing, but it’s replicating my own style, so maybe I still have a role there? That’s not hypothetical for me. For this newsletter, I built a context for agents that rests exactly on my own archive - every issue I’ve published - so that AI helps me write better and, fundamentally, more like me. So I honestly don’t know anymore: is that AI writing, or AI helping me be more myself?
\n\nAll in all, it’s just super complicated to assess what AI writing really is. So it’s all the more complex to leave that assessment to AI itself.\n\nSarah: Right, and I even find an interesting paradox here between both the trust and skepticism we have with AI and a gap between verification and accountability.\n\nTo go back to the Thélyson Orélien case; those that trust AI the least to produce literature worth awarding happen to also be the ones who trust it enough to remove it from the competition – Goncourt prize – on its verdict.\n\nHowever striking the paradox might be, it is not new at all. Over 2000 years ago, ancient Greek philosopher Sextus Empiricus developed the problem of the criterion: to know whether a claim is true we need a criterion of truth, but how do I know this criterion is itself reliable? So we either fall into an infinite regress or end up using the criterion to justify itself. Neither option is satisfactory. In the same way, a detector like Pangram (that uses AI to make its verdict) can’t be used to prove Pangram/AI’s own reliability.\n\nSo we face a verification gap, which adds up to the accountability gap. Basically, when AI says “this text was 99% developed by AI” who is responsible if the AI turns out to be wrong? Humans answer for what they say because ultimately saying is doing, which is very different from AI that shows no accountability of the “words” spoken.\n\nThis leads us to question the responsibility among humans. Is it the human(s) (an anonymous twitter account on the day of the publication) who ran the test and made the result public, or the academy that acted on it? And what about the author himself, who now would probably need to provide a negative: that he did not, in fact, use AI?\n\nLorenzo: It’s an odd asymmetry once you assume something was written by AI, even before anyone proves it: if it’s good, you don’t fully congratulate the human. But if it’s bad, the human takes all the blame. There’s no way to say “ah, but I did it with AI” - it’s never an excuse. AI gets the credit, the human keeps the blame.\n\nAnd proving it, either way, is turning into a cat-and-mouse game. In August, Anthropic announced it would watermark all text generated by Claude, to comply with the EU AI Act’s transparency rules. Within a day, someone had shipped a watermark cleaner on GitHub, and Claude skills to strip the marks soon followed. So the very label we use to judge a text can be added by one AI and wiped by another.\n\nLeave a comment\n\n> ### Save for Later\n\nClaude keeps turning up in this issue, so: Anthropic pledged 5% of its shares to a “public benefit” fund (who decides where the money goes? TBD), and its IPO filing warns of harms too “catastrophic” for the entire insurance industry to cover. Is AI above the law? Jill Lepore says yes. Meanwhile, Firecrawl hands agents 30M+ US government pages, statutes included.\n\nWho gets to read the web: Meta dropped “train” from its terms, but lawyers say “develop” still covers it. Google now auto-expands AI Overviews, and publishers are sliding down the page. Btw, connect Gmail to Google’s AI and its contents may now shape your ads.\n\nFor builders: GitHub’s former CEO wants a git forge for agents and humans. People are vibe coding the ads out of their TVs (one turned the homescreen into a Pokémon town). And Translated’s Lara 3 learns translation by doing, graded by a judge built on professional reviewers.\n\nAh, and Elena Verna is not doing 996: “not a competitive advantage.” The iPhone Duo is not the small phone you hoped for. And French students turned a Snapchat line they pinned on undercover police into a protest slogan: “Bismillah, c’est les ténèbres.”\n\n> ### The Bookshelf\n\n\"\"\n\nThis time, a recommendation from AI Organon ! ‘The art of thinking in a digital world’ is one of those books giving you a sort of toolkit or references to think through daily matters.\n\n📚 All the books I’ve read and recommended in Artifacts are here.\n\n> ### Nerding\n\n\"\"\n\nBtw, in case you’re curious about who wrote something, if AI or a human, check out Pangram. It’s full of tools like it out there, but the way this is built has more than a point and its accuracy is not bad at all! For instance, this NL is 95% written by humans :)\n\nNEW: All the tools suggested in Nerding have been collected here! ⚙️\n\n> ### ☕?\n\nIf you want to know more about Artifacts, where it all started, or just want to connect..." }