Ethics & LLMs: Intellectual Honesty

When someone discovers that a piece of writing involved AI assistance, the response is often immediate dismissal: "Oh, they used AI." I’ll be frank, I feel myself doing this, too, even though I also write AI-assisted content on this very blog! (Scroll down to see the AI use disclosure, as always!)

This reaction assumes that AI use is inherently dishonest, that it means pressing a "generate" button and calling it a day, and that all AI-assisted outputs are inferior "slop."

The Opacity of Binary Thinking

The phrase "using AI" has become meaninglessly broad. It could mean anything from having AI proofread for typos to generating entire documents from a single prompt. Yet we often treat these vastly different use cases the same.

People often perceive AI use as a form of cheating, declining credit or expressing discomfort about using AI for work or school work, even if that particular use is permissible. This emotional response stems partly from our undefined boundaries around appropriate AI use.

Current disclosure practices make this worse. Amazon KDP's binary checkbox asking if you "used AI" tells readers nothing meaningful. Did the author use AI to check grammar? To brainstorm plot points? To generate dialog? To help them get unstuck? To write entire chapters? To spit out the entire book in one go? The checkbox treats all these uses as equivalent, obscuring rather than illuminating the actual creative process.

This binary framing ignores the spectrum of human-AI collaboration. Oravec (2023) argues that we need to expand our understanding of "responsible human-AI collaboration" beyond cheating or not cheating. Just as using spell-check differs from hiring a ghostwriter, using AI for proofreading differs fundamentally from using it for complete generation.

The Hidden Craft of AI Collaboration

Getting useful outputs requires significant skill and effort.

If you’ve been reading this blog, you’ve seen my LLM use disclosures at the bottom of each post. I often use AI for supporting tasks and write most of the content on my own. These prompts look like “Here are my notes or an outline, can you flesh this out?” Frankly, I find this approach easier and use it most of the time.

Recently, I used an LLM to help me write a blog post in which I didn’t have a set of notes or an outline, and I wasn’t quite sure what the scope or framing would be. This called for an iterative approach: I wrote out what I was thinking about and asked for “a few short outlines for the post so I can select one that gets at what I'm thinking here?”

I selected one, then asked “Can you use my notes and your outline to identify some sources I should read?” It is important to me for me to understand what I am writing after all!

Then I found a couple of my own and asked the LLM to add them to its context.

Only then did I ask it to write a draft, which I edited extensively, focusing on fact-checking (not just “is this true,” but “is this the right source to support this claim?” and “is this claim too strong for the evidence?”). I also edit for voice; although not as much as I would if it were more serious than a blog post.

This method reflects what we see in research about “good” prompting: prompts that yield good results and prompts that help people learn. They are iterative, require critical thinking, and involve inputs from the user about their own context, knowledge, or research. Really effective prompting does save the user time pressing the keys to create the words that turn into a document, but still involves active, critical human contributions.

What Honest AI Disclosure Actually Looks Like

Intellectual honesty about AI use requires moving beyond binary checkboxes to meaningful documentation. Instead of simply declaring "AI was used," we need frameworks that communicate:

The extent of AI involvement: Was it used for brainstorming, drafting, editing? What percentage of the final work originated from AI versus human writing?

The process employed: Did the original idea, argument, sources, structure, and text come from the human author or the AI author?

The human expertise involved: What knowledge and skills did the human bring to guide and evaluate the AI's contributions?

Mitchell et al. proposed "model cards" for documenting AI systems, and Gebru et al. created “Datasheets for Datasets,” which document training data. (I tested Datasheets in my dissertation and found that they help machine learning engineers identify ethical problems in training data). We need something similar for AI use disclosure—perhaps "process cards" that reveal how human and artificial intelligence combined in creating specific works. I’ve applied for funding from the Knight Foundation to develop and test a low-touch, flexible authorship disclosure schema. Fingers crossed!

Transparency in authorship serves multiple purposes. First, it helps readers accurately assess what they're reading. Second, it helps us share what we learn about effective AI collaboration techniques. Third, it could add some nuance to the stigma around “they used AI.”

Why This Matters for Mission-Driven Organizations

For nonprofits and other mission-driven organizations, intellectual honesty about AI use is critically important. These organizations rely on trust. Hiding AI use or misrepresenting the level of human involvement can damage credibility.

But equally important, oversimplified anti-AI attitudes can prevent these organizations from leveraging tools that could amplify their impact. When "they used AI" becomes automatic grounds for dismissal, we lose opportunities to have nuanced conversations about appropriate use and hamstring our ability to scale our impact.

Moving Forward

AI can be used to generate low-quality content with minimal effort. But it can also enhance human creativity and productivity when used skillfully. I hope we can get to a point where we can have a nuanced conversation about how AI is used.

By being intellectually honest about our own AI use—documenting our processes, acknowledging both AI contributions and human effort, and focusing on outputs rather than tools—we can model better practices.

The goal isn't to defend all AI use or pretend quality concerns don't exist. It's to move beyond "they used AI" as meaningful criticism and toward more sophisticated discussions about process, quality, and appropriate use. In a world where AI use is common, we need to develop the frameworks and norms that will allow us to confidently put boundaries on how it is used and how use is disclosed.

—

LLM disclosure:
Claude Opus 4: “I'm writing another values post. I am not sure what the value is that I am getting at here, but here are my notes. Can you offer a few short outlines for the post so I can select one that gets at what I'm thinking here?”
”Intellectual honesty! That's it. Thank you. Can you use my notes and your outline to identify some sources I should read?
Thank you! Can you add these to your context: (citaitons to a couple papers)”
”Thank you! Can you use all of what we've chatted about, my notes, and your Intellectual dishonesty outline to draft a blog post?”

The result had a lot of black and white language (e.g. “Those who dismiss AI-assisted work as effortless reveal they've never attempted sophisticated AI collaboration.”) and random claims (e.g. “The assumption that AI use means low effort particularly frustrates those who document their process.”). It required a lot of editing. It also gave me in-text references without the full citation at the bottom, which I thought was weird! Then, when I asked it for the full citations, it gave me authors and titles of real papers, but some of the links went to weird places (my fave error: medical paper retractions in Egypt.)

It also added this “disclosure” at the bottom:
Note: This post was written through human-AI collaboration involving approximately 15 prompts, substantial human editing and restructuring, source verification, and the addition of personal examples and context. The core arguments and structure originated from human conception, with AI assistance in drafting and refinement. FALSE! 4 prompts

I also asked ChatGPT o3 to “generate a 1:1 image with no text to associate with this article.” I tried using “article” instead of “blog post” to see if it would give me a new style. I did like this one better!