Issue 003: Who Wrote That?

Some news about AI writing and governance, and a workflow to check your sources

A new study in Nature Human Behaviour suggests that large language models may be making human writing more alike. Across three studies, seven datasets, and more than 880,000 texts, researchers found that AI-assisted rewrites generally preserved the original meaning while reducing variation in writing style and weakening some of the language patterns associated with personality, empathy, morality, age, gender, and political affiliation. In other words, an AI-edited message may still communicate what you mean, but it may reveal less about how you naturally think, feel, and express yourself. Read the study

Consider: Ok, so this isn’t really a surprise. AI is making us all sound the same. We’ve all seen AI slop and noticed the tells. But this goes deeper into our actual writing styles and use of language. The next time you ask AI to “make this sound better,” notice what kind of better you mean. Do you want clearer wording, or are you slowly relinquishing your personal writing style? AI can be useful for spotting repetition, organizing ideas, or catching an unclear sentence. But before accepting a full rewrite, compare it with your original (you wrote your original, right?). Keep the phrases, examples, humor, uncertainty, and texture that sound like you, especially when the writing is meant to represent you, not just deliver information.

BambooHR’s new workplace AI report is framed around productivity, but the governance stats jumped out at me. Workers say AI has become part of how they solve problems, transfer knowledge, and perform their jobs, but the boundaries around that use remain unclear. Among respondents who use personal AI accounts for work, 71% reported entering client data, proprietary strategy, or other sensitive company information into tools their employers could not see. That does not tell us exactly how common the behavior really is (self-reported surveys can understate, overstate, or misinterpret behavior), but it is still a serious signal. Governance is not only about restricting AI. It is about making clear which tools are approved, what information is safe to enter, whether prompts are monitored, how outputs should be checked, and when a person must make the final decision. Read the report.

Consider: Choose one task where you use AI and walk through four questions: Which account am I using? What information am I entering? Who could access the interaction? How will I verify the result? If those questions are difficult to answer, that may signal a governance and training gap and not simply an individual failure to “know better.” Clear policies establish the boundaries, but learning and development can help people apply those rules in ordinary work: recognizing sensitive information, protecting data while still getting useful results, checking AI output, and knowing when human judgment or collaboration matters more. The goal is not merely to tell employees to use AI responsibly, but to give them enough information and practice to do so. There is a lot to unpack in this report and yes, I chose to focus on governance. But read it — the various results may surprise you.

The Justice Department is supporting OpenAI’s position that training language models on large collections of online writing qualifies as fair use. OpenAI’s argument is that the model is a predictive machine. It doesn’t just store and republish each article; it processes many sources to learn patterns that help it predict and generate new text. The New York Times and other publishers answer that the AI companies copied valuable work to build commercial products that can compete with, or substitute for, the original reporting, especially when chatbots reproduce articles or answer questions without sending readers to the source. The fact that a system can transform material for a new purpose is relevant to fair use, but it does not settle the question. U.S. law weighs several factors, including the purpose of the use, the nature of the work, how much was copied, and the effect on the market for the original. Courts apply those factors case by case, and this lawsuit has not produced a final answer. Read the latest article.

Consider: When we say that AI is trained on books, articles, images, or other creative work, it can sound as if the model is learning, just like a human. However, there are two separate questions: What does the model do with the material after training? Was the initial copying lawful or fair? A transformed output could support a fair use argument, but transformation alone is not a permission slip. The unresolved governance question is also economic: if AI companies depend on human-created work, what obligations do they have to obtain permission, pay for access, preserve attribution, or avoid replacing the original source? The answer will be shaped by courts, licensing practices, and public policy. For now, creators need to keep pushing for rights to their own work. (I created a short explainer video using Gemini Notebook if you want to know more about fair use.)

Until next time,
Michelle

Flow & Function: AI in Working Order

AI Disclosure:
This newsletter uses AI agents for researching potential articles, brainstorming workflows, and creating the workflow graphic in this newsletter. The final copy, writing, perspectives, em-dashes, and workflow design are created by Michelle.