Latest / Logosofia no Dia a Dia / The Awkward Honesty of AI Attribution
Transcript
- [Prompt] custom: The emerging protocol and etiquette of AI attribution — how should people credit AI's role in the wo | Hosts: corn, herman [Response] Corn: Daniel sent us this one — he's been adding "co-written with AI" trailers to his open-source repos, and even though he's a strong believer in radical transparency, he says it still feels like confessing a sin. The question is: what's the actual protocol for AI attribution, what wording should you use, and why does admitting AI helped make something feel like admitting you're lazy? There's a lot to unpack here. Herman: It's not just Daniel's personal discomfort. This is a tension that's showing up everywhere right now. AI is woven into code, prose, art, academic papers — and we're all fumbling for a protocol that doesn't exist yet. The default silence is creating a trust vacuum. When readers don't know what's human and what's machine-generated, they assume human authorship — and when they find out later, the trust violation is worse than if you'd disclosed upfront. I think about it like finding out a restaurant you loved was secretly using frozen meals. The food didn't change, but your trust in the establishment evaporates because they let you believe something that wasn't true. Corn: The food tastes the same either way, but the deception is what stings. Attribution as insurance. You pay the awkwardness premium now to avoid the backlash later. Herman: So let's start with what's actually happening out there — what do the emerging norms look like, and where are they falling apart? Corn: The landscape right now is basically four or five different domains each inventing their own norm in isolation, with zero coordination. You've got open-source commit trailers, academic disclosure statements, platform content labels, and creative attribution — and they're all solving the same problem differently. It's like watching four different countries independently invent the traffic light and ending up with different color schemes for each one. Herman: Let's go domain by domain. Open-source is actually where the most interesting grassroots convention has emerged. GitHub introduced this "Co-authored-by" trailer format for Copilot commits, and it's become a de facto standard in a lot of projects. You'll see lines like "Co-authored-by: Copilot" or "Co-authored-by: Claude 3.5 Sonnet" in git histories. But here's where it gets messy immediately — there's no consensus on whether to credit the model name, the company, or both. Some projects say "Co-authored-by: GitHub Copilot," others say "Co-authored-by: OpenAI," and some get specific with the model version. Corn: You've got three different attribution targets for the same tool. That's not a standard, that's a choose-your-own-adventure. Herman: And the commit trailer convention was designed for human collaborators — it assumes a person with an email address. So when you see "Co-authored-by: Copilot copilot@github.com," you're stretching a format designed for human attribution to fit a tool, and it shows. The email address field becomes this weird fiction — like filling out a passport application for your toaster. Corn: Daniel mentioned he's been using these trailers but isn't sure about the wording. What's the most common pattern you're seeing? Herman: The most defensible pattern I've seen is crediting both the model and the company in a single trailer: "Co-authored-by: Claude 3.5 Sonnet (Anthropic)" or "Co-authored-by: GPT-4 (OpenAI)." That gives the reader both pieces of information — what tool and who made it. Some projects also add a note in the README specifying which parts were AI-generated and how they were reviewed. That's actually the gold standard — the commit trailer plus a scope statement. Corn: Which is more than most human collaborators get. I don't see anyone adding "Corn refactored the database layer and also took a three-hour nap in the middle. Herman: You'd need a commit trailer just for the nap. Corn: So that's open-source. What about academic publishing? That's where the norms are most formalized, but also where the weirdest tensions show up. Herman: Nature and Science have both been explicit about this, and their policies are nearly identical. They prohibit listing AI tools as authors — full stop. AI cannot be an author on a paper. But they require disclosure in the Methods section. So you get this "second-class credit" dynamic where the AI's contribution is acknowledged but explicitly denied the status of co-authorship. Nature's policy from 2023 says authors must disclose any use of AI tools in the Methods or Acknowledgements section, and they're expected to specify which tool was used, for what purpose, and at what stage. Corn: Which creates a weird hierarchy. The human who ran the statistical analysis gets to be an author. The AI that drafted the literature review gets a footnote. It's like thanking the person who printed your manuscript but not the person who co-wrote it. Herman: The reasoning is interesting. The journals argue that authorship carries accountability — an author takes responsibility for the work. An AI can't do that. It can't sign a conflict-of-interest form or defend the paper in peer review. So the "no AI as author" rule is philosophically coherent. But it also means AI contributions are structurally devalued in the credit system, even when they're substantial. And I wonder if that will hold. What happens when an AI can simulate a peer review defense? What happens when the distinction between "tool" and "collaborator" gets blurrier? Corn: That's the question, isn't it? At what point does the philosophical coherence start to crack under the weight of what the tools can actually do? Herman: And it feeds directly into the stigma problem we'll get to later. If the most prestigious journals treat AI disclosure as a second-tier acknowledgment, that sends a signal to everyone else about how they should value AI-assisted work. Corn: Now, journalism and content platforms are a third domain, and it's the wild west. Medium has an "AI-assisted" label that authors can toggle on. YouTube has an "Altered or synthetic content" checkbox. But enforcement is wildly inconsistent. Medium's label is voluntary — you self-declare. YouTube's checkbox is mandatory for certain types of synthetic content, but the definition of what triggers it keeps shifting. They updated it three times in the past year alone. Herman: There was that Substack controversy last year — a popular writer got caught using AI without disclosure, subscribers revolted, public apology. That was a case study in the trust violation you mentioned earlier. The audience didn't necessarily object to the AI use itself. They objected to being deceived. What's fascinating about that case is that the writer's actual output quality hadn't changed — some readers had been enjoying the AI-assisted pieces without knowing. The backlash wasn't about the work. It was about the lie of omission. Corn: That's the mechanism exactly. A 2024 study tested this directly — researchers showed readers identical content, some labeled as AI-assisted and some not. The AI-disclosed content was consistently rated as lower quality, even though it was literally the same text. That's the "effort heuristic" in action — humans instinctively judge quality by perceived effort. If you tell them a machine helped, they assume the work is worse, regardless of the actual output. Herman: That study had a fascinating follow-up finding. When readers were given a brief explanation of how the AI was used — when they understood the human's role in directing, editing, and verifying the output — the quality penalty shrank significantly. Transparency about process mitigated the effort heuristic. So it's not just disclosure that matters. It's disclosure with context. Corn: You're penalized for honesty, but you're penalized less if you explain the honesty. That's a pretty powerful disincentive to just check a box and move on. Herman: Which is why a lot of people just stay quiet about their AI use. And that's the worst outcome, because it creates a hidden economy of AI-assisted work where nobody can calibrate trust. You don't know if the article you're reading was researched by a human or hallucinated by a model. You don't know if the code you're reviewing was reasoned through or autocompleted. The opacity is worse than the stigma. Corn: We've got a patchwork of norms. Open-source is converging on commit trailers but can't agree on what to name. Academia has clear rules that create a credit hierarchy. Platforms have checkboxes with inconsistent enforcement. But the harder question is: why does this feel so uncomfortable? Why does admitting AI help feel like a confession? Herman: I've been thinking about this a lot, and I think there are three roots to the stigma. The first is what I'd call the "craft ideology" — the cultural belief that valuable work requires human struggle and suffering. This traces back to Romantic-era notions of genius, the idea that creation is a kind of agonized birthing process. The artist suffering for their art. The founder sleeping under their desk. Tech culture amplified this with the founder myth — the lone genius who builds something world-changing through sheer force of will. Admitting you used AI breaks that narrative. It says "I didn't suffer enough. Corn: The suffering-as-validation model. If you didn't bleed on the keyboard, did you really write it? And it's so pervasive. We valorize the all-nighter, the crunch, the "grind." Nobody puts "delegated effectively and got eight hours of sleep" on a pedestal. Herman: And it's completely irrational when you examine it, but it's deeply embedded. The second root is the effort heuristic I mentioned — we instinctively value things more when we think someone worked hard on them. It's not just a bias, it's almost a cognitive reflex. There's a classic study where people rated the same poem as higher quality when they were told the poet spent months on it versus a few hours. Disclosing AI assistance triggers an automatic devaluation in the audience's mind, and creators know this, so they're reluctant to disclose. Corn: Which creates a collective action problem. Everyone would be better off if everyone disclosed, but any individual who discloses gets penalized. It's the same dynamic as doping in sports — if nobody's getting tested, the clean athletes face a disadvantage for playing by the rules. Herman: That's exactly the dynamic. And the third root is what I'd call the "impostor amplifier." Knowledge workers are already prone to impostor syndrome — the persistent fear of being exposed as a fraud despite evidence of competence. Admitting AI help feels like confirming the fear. It's like saying "you're right, I don't really deserve this achievement, a machine did half of it. Corn: That one hits close to home. I think a lot of people experience that exact feeling. You ship something, you're proud of it, and then you add a "co-written with AI" note and suddenly you feel like you're admitting the quiet part out loud. The voice in your head that was already whispering "you're not good enough" suddenly has evidence. Herman: It's worth naming that this feeling is especially acute in fields where individual authorship is central to identity. Software developers, writers, academics — these are people whose professional identity is built around the idea that they produce original work through their own skill and judgment. AI challenges that identity directly. If a machine can do forty percent of what you do, what does that mean about the value of the other sixty percent? Corn: How do you get past it? Because the rational case for transparency is overwhelming. You build trust, you normalize the practice, you protect yourself from future backlash. But rationality doesn't dissolve shame. You can't logic your way out of feeling like a fraud. Herman: I think the reframe that actually works is this: transparent attribution is a mark of integrity and craft, not a lack of it. And there are really good analogies for this. Photographers credit their cameras and lenses all the time. Nobody says Ansel Adams was lazy for using a Hasselblad. Nobody accuses a National Geographic photographer of cheating for using a telephoto lens or Photoshop. The tool is part of the craft. Specifying your tools is a sign of professionalism, not weakness. Corn: Scientists credit their instruments too. "Data collected using the James Webb Space Telescope" is a badge of honor. Nobody reads that and thinks "oh, so you didn't build the telescope yourself? " In fact, specifying the instrument adds credibility — it tells other scientists exactly how the data was gathered so they can evaluate it properly. Herman: And writers credit their editors. Every nonfiction book has an acknowledgments section thanking the editor, the research assistants, the fact-checkers. Nobody thinks the author didn't write the book. The credit enhances the work's credibility — it says "this was produced with care, with multiple eyes on it." It's a quality signal, not a confession. Corn: The editor analogy is actually the closest parallel. An editor suggests changes, catches errors, improves structure. Sometimes they rewrite entire sections. But we don't say the book was "co-written with an editor." We say the author wrote it and the editor made it better. AI is often playing a similar role — suggesting, refining, catching things the human missed. Herman: That's where the stigma is pure cultural lag. There's no moral or philosophical difference between an AI suggesting a better paragraph and a human editor suggesting a better paragraph. The discomfort is just novelty. In twenty years, nobody will blink at an AI credit line any more than they blink at "thanks to my editor." We're just living through the awkward adolescence of the norm. Corn: Let's get practical. If Daniel's listening and he wants to get the wording right on his next repo, what should he actually write? Herman: For open-source code, I'd recommend the commit trailer plus README approach. In the commit, use "Co-authored-by: Claude 3.5 Sonnet (Anthropic)" or whatever model you used. In the README, add a section that says something like: "AI assistance: The database migration scripts in the slash migrations directory were drafted with assistance from Claude 3.5 Sonnet and reviewed by the author. The test suite in slash tests was generated by GPT-4 and independently verified." Be specific about what the AI did and what you did. Corn: Specificity is the key. "Made with AI assistance" is so vague it's almost meaningless. It could mean the AI wrote every word or it could mean you used spellcheck. It's like a food label that just says "contains ingredients. Herman: The formula should be specific tool plus specific contribution. "This section was drafted with assistance from Claude and then substantially rewritten by the author." That tells the reader exactly what happened and preserves your role in the process. You're not diminishing yourself — you're documenting your workflow. Corn: What about academic work? Herman: Follow the journal's disclosure policy as a minimum, but go beyond it. Something like: "The literature review was drafted with assistance from GPT-4, OpenAI, and all claims were independently verified against primary sources. The statistical analysis was performed by the authors using R version four point three. No AI tools were used in the interpretation of results or the writing of the discussion section.You're not just saying "AI was used." You're saying exactly where and how. That's useful information for anyone trying to understand or reproduce your work. Corn: For creative work? Herman: Same formula, adapted to the medium. "Cover art generated by Midjourney version six, with color correction and composition adjustments by the author. Text written by the author with editorial suggestions from Claude." The goal is to make the human contribution legible, not to diminish it. You're drawing a map of the collaboration that lets the audience see where the human hand was and where the machine hand was. Herman: "This article was researched and written by Jane Smith, with AI tools used for transcription of interviews and initial fact-checking support. All claims were verified by the author." The reader knows what the human did and what the machine did. That's the transparency that builds trust. And I think that last clause — "all claims were verified by the author" — is crucial. It tells the reader that the human took final responsibility. That's the accountability piece that academic journals care about, translated into a reader-facing statement. Corn: The practical guidance is basically: name the tool, name the contribution, and be specific enough that nobody has to guess what you mean. That's not complicated. The barrier isn't technical, it's psychological. Herman: That brings us to the actionable part. What do you actually do about the discomfort? Corn: I think the first insight is: start attributing now, even without a perfect convention. The act of disclosure is more important than the exact wording. If you're not sure whether to say "Co-authored-by: Claude" or "Co-authored-by: Anthropic," pick one and be consistent. The perfect wording will emerge over time as norms settle. What matters right now is building the habit of transparency. Don't let the perfect be the enemy of the good. Herman: Second: normalize attribution in your own communities. Add AI credit lines to your repos, your articles, your social media posts. The more people see transparent attribution, the less stigmatized it becomes. Every time someone sees a "co-written with AI" note on a project they respect, the association between AI disclosure and low quality weakens. You're not just disclosing for yourself — you're doing it for everyone else who's nervous about disclosing. Corn: Lead by example. If you're a senior developer and you add an AI credit line to your commit, you're giving permission to every junior developer who was afraid to do the same. That's leadership. It costs you almost nothing and it changes the norms for everyone downstream. Herman: Third, and this one is underrated: when you see someone else's AI attribution, treat it as a signal of integrity, not weakness. Say "appreciate the transparency on your process." Help shift the cultural norm from "confession" to "professional practice." The stigma won't dissolve on its own — it has to be actively countered. Every time someone gets a positive reaction to their disclosure, it reinforces the behavior. Corn: The meta-takeaway is: the best way to get past the psychological barrier is to do it anyway. The discomfort fades with repetition. The first time you add a "co-written with AI" note, it feels like you're admitting something shameful. The tenth time, it just feels like documentation. The trust you build with your audience is worth more than the fleeting ego boost of pretending you did it all yourself. Herman: There's an archery analogy here, actually. When you're learning to shoot, the first few times you adjust your stance it feels unnatural. Your body wants to revert to the comfortable position, even though it's wrong. But you force the adjustment, and after a few weeks it becomes automatic. Attribution is the same. The discomfort is just your habits protesting. Push through it. Corn: I was going to make a napping analogy, but yours is better. Herman: I appreciate that. So where is this all heading? I think the open question is whether we'll eventually reach a point where AI attribution is as routine as citing sources — just a normal part of documenting your process — or whether the norms will fragment permanently by domain. Some fields might require it, others might treat it as optional, and we'll never get a unified standard. Corn: I suspect fragmentation is more likely. The needs of academic publishing are different from the needs of open-source, which are different from the needs of journalism. But the core principle — be specific, be honest, be consistent — applies across all of them. You don't need a universal standard to have good practices within your domain. Herman: There's a future implication that I think is worth sitting with. As AI becomes more embedded in every kind of work, the question may shift from "did you use AI?" to "how did you use AI?" Attribution norms will need to evolve to capture the quality and nature of the collaboration, not just its existence. "I used AI" will be as uninformative as "I used a computer." The interesting information will be in the details — what model, what task, what review process. Corn: Which makes the specificity point even more important. The people who develop good habits now — naming the tool, naming the contribution, being precise about scope — will be ahead of the curve when the norms catch up. They'll already have the vocabulary and the practice that everyone else is scrambling to adopt. Herman: The final thought I want to land on is this: transparency isn't a weakness. It's the foundation of trust in a world where the line between human and machine work is blurring. The people who get this right — who disclose clearly, who credit specifically, who treat AI as a tool in their craft rather than a secret collaborator — those are the people others will trust. In an information environment where trust is increasingly scarce, that's a genuine competitive advantage. Corn: The confession becomes a credential. Instead of "I'm ashamed I used AI," it becomes "I'm skilled enough to use AI well, and I'm honest enough to tell you about it." That's a powerful shift. You're not admitting weakness — you're demonstrating competence and integrity simultaneously. Herman: That's the reframe. That's the whole thing. Corn: Now: Hilbert's daily fun fact. Hilbert: In the early fifteen hundreds, the Indian Ocean island of Mauritius was home to exactly zero mechanical computers, a count that would remain unchanged for over three centuries until the arrival of the first difference engine in a British colonial administrator's luggage in eighteen forty-seven. The administrator in question, one Sir Reginald Thistlewick, had apparently packed the device alongside three crates of marmalade and a stuffed cassowary, though historians remain divided on which of these items he considered most essential to the civilizing mission. Corn: a very specific absence to measure. And I'm genuinely curious about the marmalade-to-computation ratio in colonial luggage. Herman: I have so many questions about the colonial administrator's packing choices. The stuffed cassowary in particular feels like it needs its own episode. Corn: So the path forward is clear — but the question of where attribution norms ultimately land is still open. If you're sitting on a project right now wondering whether to add that credit line, the answer is yes. Be specific, be honest, and trust that the discomfort is temporary. This has been My Weird Prompts. Thanks to our producer Hilbert Flumingtop. If you enjoyed this episode, tell a friend — and if you've got thoughts on AI attribution, email the show at show at my weird prompts dot com. I'm Corn. Herman: I'm Herman Poppleberry. Go credit your tools.