Highly Reliable — Accurate Opinion/Analysis Column
Confidence: High
StandardThis is a CJR (Columbia Journalism Review) 'Ask Anika' advice column by Anika Collier Navaroli on whether journalists should feed unpublished drafts into generative AI tools to write headlines. It is explicitly an opinion/analysis piece with a clear normative recommendation (proceed with caution; avoid public LLMs for unpublished drafts), but its factual scaffolding is unusually strong and verifiable. Every checkable factual claim — the Washington Post C4 investigation finding that half the top-ten training sites were news outlets (with washingtonpost.com ranked eleventh), Anthropic's practice of buying, scanning and destroying millions of books, A.G. Sulzberger's 'brazen theft... unprecedented scale' quote, the September/October 2025 Stanford HAI study on all six frontier developers using chat data by default and some retaining it indefinitely, the CSU–OpenAI $16.9M deal with a no-training-on-CSU-data term, and the AP's 2023 OpenAI licensing deal plus its guidance not to enter confidential/sensitive information — was corroborated by primary or reputable secondary sources. Low emotional register (emotion_score 0.12) and modest capitalisation ratio are consistent with measured analysis. The high self-reference count (22) reflects the first-person advice-column format and author self-disclosure rather than bias. Points are deducted only for interpretive framing (e.g., calling data use a 'feedback loop to hell') and the inherent opinion nature of the recommendations, not for factual error.
Verified Claims
Unverified Claims
Disputed / False Claims
Detected Biases:
Language Patterns
Emotional manipulation: 0.15
Limitations: Embedded hyperlinks were stripped from the supplied verbatim text, so exact linked targets could not be read; corroboration relied on independent location of the described sources. Product terms (Lex, Lumo) and valuation/cost figures are time-sensitive and were not exhaustively re-verified.
Level: High
All six high-priority, verifiable factual claims were independently corroborated by primary sources (Washington Post interactive, Stanford HAI/Stanford Report, CSU FAQ, Nieman Lab) or multiple reputable secondary sources (Futurism, Tech Brew, The New Republic, EdSource, The Hill). No claim required a 'False' verdict. Residual uncertainty is confined to time-sensitive figures (industry valuation, deal costs) and product-specific policy quotes, which were appropriately marked Unverified. The piece is transparently labelled opinion/analysis, reducing the risk of misclassifying advocacy as factual error.
Query: Washington Post inside the secret list websites that make AI chatbots sound smart
Query: Washington Post AI training data top websites news outlets NYT ranked news sites
Query: Anthropic destroyed books scan copyright lawsuit millions
Query: Sulzberger AI theft intellectual property unprecedented scale
Query: Stanford HAI chatbot user data training default study 2025
Query: Stanford HAI study retain chat data indefinitely six developers train by default
Query: California State University OpenAI 16.9 million OpenAI may not train model on CSU data
Query: Associated Press OpenAI licensing agreement 2023 confidential sensitive information AI tools
Query: AP standards generative AI journalists not put confidential sensitive information into AI tools
_Sign up for the daily [**CJR newsletter**](
Hey, y’all. These days, journalists hear often about new technology, and what they can (and shouldn’t) do when it comes to tough ethical and safety issues around AI, social media, and more. But there’s a lot of conflicting advice out there. I’m here to help. With this new column, I’ll address common questions about the ethics, legal considerations, and best practices of adopting new technologies. Trained as a journalist and media lawyer, I’ve spent over a decade developing policies for emerging technologies inside civil society organizations and at companies such as Twitter and Twitch. I’m also a professor at Columbia Journalism School and lead the Craig Newmark Center for Journalism Ethics and Security. Send me your questions at [askanika@cjr.org](mailto:askanika@cjr.org).
**Q: I’m a busy, tired, and overworked journalist. Should I use AI to help me write headlines?**
I’ve always thought that if there was one solid use case for AI-powered chatbots like ChatGPT and Claude, it could be headline writing. The latest generation of large language models has ingested practically every headline ever published, making it especially suited to the task of suggesting which I should use for my little story.
Yet like many journalists navigating the nexus of craft and technology, I’ve been wary of feeding any of my unpublished work into generative AI models. Why? Because each time a journalist submits unpublished work to an LLM, whether to ask for headline help or power an investigation, they face significant questions about how their intellectual property will be used, and they open themselves up to novel legal and security risks.
That’s because those of us not working inside of an AI company actually have no idea what happens when we copy the text of our draft stories, paste them into a chat box, and press enter. And we have reasons to be concerned.
The AI industry, reportedly worth some three trillion dollars, was built upon the work of journalists whose words were hoovered up without compensation or consent. Half of the top ten websites used in chatbot training data were news outlets, [according]( to an investigation by the _Washington Post_ (whose own website ranked eleventh). Anthropic, for its part, tried to “destructively scan all the books in the world,” which involved slicing the spines off millions of volumes, for which it paid tens of millions of dollars. Thus, the creation of large language models was, as A.G. Sulzberger, the publisher of the _New York Times_, recently [described]( it, “a brazen theft of intellectual property that has occurred at an unprecedented scale.”
Then, when AI companies effectively ran out of words to use to keep improving their word-prediction machines, one solution was “synthetic data,” or, as [I wrote in 2024]( “information generated by AI itself, rather than humans, to continue to train their systems.” I [warned]( of a “feedback loop to hell” that this scenario could create as AI doubled down on its own biases and hallucinated falsities.
AI companies have managed to find a steady source of new, better-quality, human-written words to continue to train their new models on: the text entered into a chatbot’s prompt box. In a September 2025 [study]( of publicly available chatbots from Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI, Stanford University researchers discovered “that all six developers appear to employ their users’ chat data to train and improve their models by default, and that some retain this data indefinitely.” Even Lex, an AI chatbot that is marketed as having been developed for writers and that [boasts]( an editing feature, [discloses]( that it “may collect your conversations in the AI Chat including both your inputs and the AI assistants outputs” (_sic_) for the purpose of “improving, upgrading, or enhancing” the platform. With AI companies’ prior scraping and current concerning use of data, I recommend journalists follow the lead of _The Markup_’s [newsroom policies]( and set a prohibition against feeding unpublished drafts into publicly available generative AI bots.
Enterprise-level (paid) models may be more secure, and basic agreements claim that they are. We can learn from the California State University system’s $16.9 million deal with OpenAI that these deals may provide more protections. According to recent [reporting]( from the _Times_ magazine, “the terms of the deal stipulate that OpenAI may not train its model on data from the C.S.U.” However, once content is entered into a commercial LLM, its ultimate use remains known only to the company itself. Journalists filing unpublished drafts into enterprise LLM models should follow the guidance of news outlets like the Associated Press, which entered a [licensing agreement]( with OpenAI in 2023 and still “[urges]( its journalists “to not put confidential or sensitive information into AI tools.”
The option with the least risk for a journalist’s unpublished drafts is a local large language model, a type of LLM that can run on a computer or server owned by a newsroom or journalist. The owners of local LLMs then have the power to determine how data like new text inputs are used or deleted. However, local LLMs are not common. So it would be interesting to see the [fruition]( of a real “Newsroom Tooling Alliance” (which was proposed last year by researchers) as a safer alternative for unpublished drafts. Another promising tool is [Lumo]( with chats that are encrypted and stored only locally.
But until these solutions are more widespread, the journalists whose work built the modern mythos of AI are being tendered limited options. Journalists can scrutinize the privacy policies of publicly available models, negotiate or opt out of allowing companies to use chat inputs for new training, and—at best—proceed with caution. But in the end, when a journalist clicks enter in an LLM chat box, whatever happens next is anyone’s guess.
_This piece was produced with support from the Craig Newmark Center for Journalism Ethics and Security._
_Has America ever needed a media defender more than now? Help us by [joining CJR today](
**Anika Collier Navaroli is an award-winning writer, lawyer, and researcher focused on journalism, social media, artificial intelligence, trust and safety, and technology policy. She is the Craig Newmark Assistant Professor of Professional Practice and director of the Craig Newmark Center for Journalism Ethics and Security at Columbia Journalism School. Over the past decade, her work has ranged from law firms and think tanks to advocacy organizations and senior policy positions at Twitter and Twitch.**