PubTech Radar 第 44 期：期刊 AI 写作政策、Claude 文本水印与 Papers AI 发布

Hello, and thanks for still being here! As it’s been a while since I last sent out a newsletter, here's a quick reintroduction. I’m Helen, and my background is in academic publishing and technology. Ex-Nature Publishing Group, ex-Sage, ex-BMJ, plus a stint on the Crossref board and a couple of STM committees. You can find me at Maverick or Innovation Ideas. I’m currently writing from North Wales, where I’m taking some time out to care for someone and where the latest AI and technology developments feel a million miles away.

I thought I would focus this issue on articles about AI writing that have caught my attention, partly because I’ve recently had the pleasure of being on a couple of panels that were largely about this, including TORQUE in Bruges, on AI in scientific publishing, and the ISMTE conference, on AI safeguards, and partly because my views on AI writing are changing as I notice what it is doing to my writing and thinking.

AI disclosures and writing policies:

I think it’s important that a journal’s AI policies reflect the opinions of the communities who read and write for them and that we don’t have a one-size-fits-all approach.

• Philosophy & Public Affairs [tweet thread] explains their policy and why they want no AI-authored writing. “Academic journals serve at least two functions: the promulgation of new knowledge, and identification and credentialing of talented researchers... Submitting AI-authored essays makes the task of identifying talented researchers harder.”

• At the other end of the spectrum, Elsevier’s new Generative AI policies for journals are very permissive and allow AI use for literature reviews, identifying research gaps, generating ideas, structuring manuscripts, improving writing, checking manuscripts, and creating some explanatory images. See their FAQ for details.

• Florent Daudens on Stanley Druckenmiller's admission that his WSJ op-ed was written with AI: Who gets to use AI to write? What a billionaire, a Harvard professor, ghostwriters, and a few research papers tell us about the increasingly messy rules of authorship. "For the WSJ's opinion section, an AI-assisted op-ed can still be authentic if the underlying argument belongs to the person whose name appears on it. For the FT, the process of producing the words is also part of the editorial contract."
  
  A year ago I would have been in the WSJ camp, but now I'm veering towards the FT's position, especially for editorial content, for two reasons. The first is that I think AI writing patterns are becoming stronger and are very hard to get rid of, so I find myself dismissing writing as AI slop even if the ideas are good. Secondly, unless you are really clear about your view (which I think Stanley Druckenmiller would be), the AI anchors what you write. Even if you're just using it to take voice notes, it tends to give these a spin, which can subtly change your thinking. [I'm assuming AI is being used collaboratively and that the user is giving it their own thoughts and iterating on them, rather than asking it to write the whole thing.]

• I like Clay’s (a go-to-market automation platform) AI writing policy. It’s a simple approach, but requires a lot of user input:

  • You must stand behind every idea and sentence

  • Writing is thinking [I’m less certain this is true for all of us]

  • More time should be spent writing a document than consuming it

  • Longer is not better

• Scholarly Kitchen asks if Research Needs a Shared Standard for AI Disclosure? Whilst a neat and tidy AI disclosure taxonomy appeals to me, I don’t think this would be as stable or straightforward as CRediT contributor roles. I also suspect it might date badly and need regular updating.

• Claude will watermark text to comply with the EU AI Act. In this case, watermarking is done via word selection - see this accessible explainer about the technicalities of how this works. [I think that if you read enough unedited Claude-written text, you probably don’t need a watermark to identify Claude writing.] The Claude team note: “Can’t someone just edit the text to get around the watermarking? To some extent, yes. Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will.” I am uncertain whether this kind of watermarking has a lot of real-world value in scholarly publishing. I would have thought anyone creating papermill-type content isn’t going to be using Claude or can easily work around this.

• If you’re interested in testing how easy it is to fool Pangram with academic writing, here’s a simple exercise that takes around 15 minutes and follows the same passage as it is generated by AI, processed by a humaniser, and then copy-edited by another AI tool. Pangram 4 is harder to fool than earlier versions, but not that hard to fool:

🚀 Launches:

• Digital Science (Overleaf) has launched Papers AI, which says it “brings your writing, data, references and code into the same flow, so you can follow the thought without rebuilding the context.” Mark Hahnel’s write-up here.

• OpenAI is sunsetting its LaTeX document creation tool Prism, and OpenAI for Science, which was led by Kevin Weil, is being ‘decentralised’ [H/T: Ian Mulvaney]. Slightly off topic, but Kevin Weil is now seeking $150 million in funding for a new company to gather scientific data for AI models.

📊 Data:

• Caleb Robinson and Isaac Corley put together a Q&A from the slop trenches, including because it contains real data about what people are seeing: “We reviewed 22 ML conference submissions this summer. Fifteen had fabricated citations, hallucinated authors, or clear LLM slop, so we complain about that a bit, and release the paper references audit we now run.”

• Mahmud Omar asks: Are we losing our maybes? “Hedges fell 6.1% per 1,000 words. Boosters rose 31.4%. The hedge-to-booster ratio fell 28.6% (as you can see in the figure below).”

Reflections on AI writing:

• Dr Benn Gooch’s nuanced and interesting comments on abandoning the use of AI scribes after 18 months are worth a read. “I felt myself becoming a passive observer in encounters where I had previously been an active architect. I felt my clinical memory, my narrative identity, and my sense of connection to my patients beginning to erode at the edges.”

• Scott Bouchard at Silverchair posted What I got wrong (about AI) this year. “The enterprise and institutional deployments, where the access controls and privacy protections live, are not where most of the work is happening.” The thing I got most wrong is that I thought AI writing would be a lot better and largely undetectable by now. I’ve tried various AI-supported writing systems such as the “feed it your writing and build a voice profile” approach, various skills such as Mike Caulfield’s Jamesian skill, and an AI-tells removal skill similar to Dustin Smith’s Clean Prose skill, but on the whole, none of them seem to work very well, and AI writing is becoming increasingly formulaic.

• Scott Cunningham on “Memory is the residue of thought” and fixing that when working with AI agents about the problems he is encountering around memory when working with agents and his visual approaches to fixing this.

Longer reads on AI writing that I found interesting/thought-provoking:

• Superslop, and How It Will Kill Your Book Deal is an interesting experiment from Michael O. Church. He says, “I figured it would take about sixty hours and $2,500 worth of Claude tokens to build an AI-generated novel that could break a sophisticated autograder.” However, “The process took five hours and used $209.55 of Claude tokens; its result was a novel that, if presented to traditional publishing with a photogenic, class-fluent white woman as its author, by a literary agent able to get manuscripts read, would secure a book deal with significant support.”

• “Generative AI floods and dilutes the market for books,” by Tuhin Chakrabarty et al., unsurprisingly finds that AI-detected text is increasingly common across the Amazon self-published fiction genres they studied. Books with >25% detected AI text are 20% of titles but account for only 12.1% of sales and 11.3% of revenue, but some sell extremely well, and their share of sales is rising. Meanwhile, the number of titles competing for sales has grown much faster than revenue, and revenue per selling title has fallen. See also: Thad McIlroy’s 6,000-word response.

• Hollis Robbins on The AI campus gender wars ahead: “The most volatile demographic and technological clashes on campus this fall will not be on the streets but in the classroom. Incoming young men, increasingly conservative and highly optimistic about generative AI, will be placed in classrooms with non-tenure-track, overwhelmingly female, politically progressive first-year writing faculty fiercely hostile to AI.”

• Henriette Thiesen, Publishing Director at Gyldendal, on AI and the future of the textbook: “A textbook tells a student: ‘This is what you need to know’. Setting those boundaries is itself a form of expert knowledge, and it is something only people with genuine subject mastery can provide.”

• Dr Sam Illingworth on You Will Be Accused of Using AI. Here Is How to Prove You Wrote It.

Upcoming events:

• Building capacity with AI tools you already have: practical applications for scholarly publishing looks interesting. Sept 16 at 10:30 AM ET.

And finally…

Interesting that the people building language models to automate everyone else’s work still want a human editor for their own content:

End Notes:

• Newsletters might be even more sporadic than usual for a while.

• If you found this useful and feel so inclined, you can always buy me a coffee.

• If you need consulting help navigating any of this, find me at Maverick.

注意转载授权 · 阅读原文