My experience trying to write original, full-length human-sounding articles using Claude AI

by Tom Johnson on Oct 25, 2023
categories: ai • podcasts • writing

• Podcast version
• Introduction
• Background research
  • The jagged frontier
  • A director guiding actors and scenes
• My three big strategies for using AI to write
  • Prime the AI with accurate information
  • Go paragraph-by-paragraph
  • Balance personal voice with explanation
• Step-by-step walk through
  • Step 1: Define an information pattern
  • Step 2: Create an outline
  • Step 3: Let Claude know the context of the writing project
  • Step 4: Calibrate Claude’s initial language and style
  • Step 5: Continue paragraph by paragraph through the essay
  • Step 6: Compile the paragraphs into a whole article
  • Step 7: Edit the whole
  • Step 8: Fine tune the article
  • Step 9: Incorporate some auto-generated art
  • Step 10: Let the content sit a day or two before publishing
• Reader feedback
• Why AI-assisted content can feel “off”
  • The Uncanny Valley idea
  • Authentic human voice and emotional connection
  • Over-agreeableness leading to diluted argument
• Conclusion

Podcast version

I also recorded a video podcast version of this post here. The podcast is more of an informal summary rather than a verbatim audio recording of the post.

The podcast is also a Youtube video.

Here are the slides that go along with the podcast. I generated these slide images through DALLE 3.

Introduction

In my API course section on how use AI with technical writing, you’ll notice one topic is conspicuously absent: there’s no post that explains how to write original, full-length content with AI. This is kind of the holy grail that a lot of people are hoping AI can help with — generating new writing and ideas rather than just assisting with editing or summarization tasks. In this article, I’ll share my attempts to answer this question: Can you use AI tools to write full-length content, specifically articles suitable for a blog?

Before jumping into this topic, you might wonder why I’d even try. I’m fully capable of writing sentences myself, so why would I want to abdicate the role to a non-human entity?

I’m not entirely sure. Some groups predict that technical writing might be automated in a few years. For example, see Forrester’s 2023 Generative AI Jobs Impact Forecast, US. There’s a sense that if we don’t figure out how to leverage AI tools, we’ll be displaced by those who do. So I’ve been experimenting with AI lately, trying to get a sense of what AI is good for, what it’s not.

The most salient feature of AIs powered by LLMs is their ability to write—to construct sentences coherently, effortlessly, and (sometimes) accurately. Many use AI to fix problematic sentences or paragraphs here and there. Why can’t we use AI as a tool for writing full-length content, such as articles or personal essays? It’s a valid question because if AI tools can do some writing tasks, why not more?

Background research

Let’s start by looking at what research has been done on how to write with AI. It’s hard to get past the endless number of marketing posts for AI writing tools, but here are two articles that caught my attention.

The jagged frontier

In Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, Dell’Acqua et al describe how consultants at Boston Consulting Group used AI tools like GPT-4 in their work. Their “jagged frontier” metaphor describes the uneven results of using AI for tasks. For tasks within AI’s capabilities, consultants using the tools saw major gains in productivity, speed, and quality. However, for tasks more challenging for AI, the consultants who relied too much on the AI performed worse. The authors argue that —

… the capabilities of AI create a “jagged technological frontier” where some tasks are easily done by AI, while others, though seemingly similar in difficulty level, are outside the current capability of AI. For each one of a set of 18 realistic consulting tasks within the frontier of AI capabilities, consultants using AI were significantly more productive (they completed 12.2% more tasks on average, and completed tasks 25.1% more quickly), and produced significantly higher quality results (more than 40% higher quality compared to a control group).”

The authors observed “two distinctive patterns of successful AI use by humans along a spectrum of human-AI integration.” The successful consultants — the ones who found ways to strategically integrate AI only where appropriate — are called “centaurs” (half-human, half-horse), because they intelligently delegate between human and AI. Those who used AI all the time are “cyborgs” because they’re fully machine-oriented.

Overall, the results suggest that AI can enhance productivity on certain tasks for those who know how to use it, but users need discernment about when not to rely on AI. As such, the jagged frontier may be easier for experts to navigate than amateurs.

A director guiding actors and scenes

In an Everyday AI podcast called Futurecrafting, Brian Sykes, an experienced creative professional who has run an ad agency for more than two decades, says to view AI as an amplifier of human creativity, not a substitute. The human element is essential for steering the AI and shaping the final product, he says.

Granted, I think Sykes isn’t so much focusing more on ad design than writing, but the argument still applies. Here’s the podcast:

What struck me most in the conversation is the metaphor of moving from actor to director. Sykes explains this shift upward into directorial work as follows:

I think creative professionals will increasingly take on the role of a director as generative AI becomes more prevalent. When you’re early in your design career, you want to make your mark by being completely responsible for the end product. But as you advance, you become responsible for directing others beneath you to create results that address the client’s needs. So generative AI will allow creative professionals to engage with it even at an early career stage. They’ll still need knowledge to get the results they want, just like with any tool. But increasingly, we’ll become directors who define the messaging while still retaining the human element at the core. (Note: I’m paraphrasing because it’s a podcast.)

Sykes says that adopting AI tools is akin to designers moving from hand-coding websites to more automated tools (similar to what happened when Dreamweaver came out). You’re not removing yourself from creating, but rather shifting to a more strategic, directorial role.

The introduction of new tools allows creatives to focus less on lower-level execution and more on high-level vision. As a director, you’re concerned with the overall narrative shape, the flow from scene to scene, the audience experience, and so on.

This idea appeals to me. Do you know that before I publish any post on this blog, I’ve reread the post half a dozen times trying to fix and smooth the language? I often paste it into Grammarly to catch obvious errors. But despite using grammar and style checkers, I usually need to let the content sit a day, read through it and edit it, let it sit another day, edit it again, let it sit another day, edit it again, etc. Eventually the edits become less and less fewer and fewer, and I get up the courage to publish it. That aspect of writing is tedious and isn’t what engages me. What if AI could let me focus on the ideas while taking care of all the language smoothing and editing? What if I could focus instead on the underlying story more than the sentences?

If AI could handle lower-level translation of thoughts into sentences, allowing me to concentrate my mental energy on insights, ideas, and the overall narrative shape, this might be a good evolution of writing.

My three big strategies for using AI to write

Now that I’ve explored some research, let me share my experiments with AI-assisted writing. I’ve only written a couple of posts using this AI-assisted technique. You can read the posts here to get a sense of what to expect:

• Embracing professional redefinition
• What is Diátaxis and should you be using it with your documentation?

(Some of the recent news posts I write are also AI-assisted, but only the summary bullets, so I haven’t included them here.)

When using AI tools to generate original writing, I try to use these three techniques:

1. Prime the AI with accurate information — If the article contains a lot of explanatory material, I provide the AI with source material, such as articles on the topic. This priming stage is crucial to ensure the AI has proper context and won’t hallucinate. Because Claude allows for the most token input, this is the AI I use for these tasks.
2. Go paragraph-by-paragraph — Rather than prompting the AI to generate a full article in one go, I work incrementally, paragraph by paragraph. After each paragraph, I review and steer the direction as needed before moving to the next.
3. Balance personal voice with explanation — Blending personal experience and using first-person perspective helps disguise and enhance AI-generated explanatory content. By switching between experience and information, it creates an engaging, human-like tone (in other words, a personal essay).

I’ll expand on each of these strategies.

Prime the AI with accurate information

My first technique is priming the AI with extensive information at the outset. This technique assumes the post is somewhat informational rather than experiential. Tools like Claude.ai allow immense input context — up to 100,000 tokens, equivalent to the length of a novel. Basically, if you’re writing about a specific topic, you can copy and paste a dozen or more articles into a text file, and then paste it into Claude so that Claude is more informed about the topic you’re writing about.

Having Claude’s encyclopedic context available throughout the drafting process is helpful. But there’s another benefit to Claude’s long input length: I can preserve the context throughout the writing process. Claude is aware of the article I’m writing as a whole, not just isolated sections here and there. I can continue shaping and refining the article until I reach the 100k token limit (which is about 75,000 words).

Go paragraph-by-paragraph

Rather than try to generate an entire blog post in one go, I direct the AI paragraph-by-paragraph. After each draft paragraph, I review the paragraph and make sure Claude is going in the right direction before I continue on. This iterative collaboration allows me to leverage the AI while still keeping control to steer the direction of the content.

As I go paragraph by paragraph, I’m a director guiding the scenes of the play, injecting style and purpose into the raw material the AI provides. Instead of getting bogged down in the grunt work of wordsmithing, I dedicate more energy to crafting the overarching direction and narrative. My job becomes director, not actor.

Balance personal voice with explanation

About 20 years ago, I did an MFA in literary nonfiction at Columbia University in New York. My biggest takeaway was that balancing ideas with personal experiences results in engaging content. This blend of storytelling and ideas is the main technique for the personal, creative nonfiction genre.

Adding personal experiences into an essay also disguises the AI-written material. If you just have pure explanatory content, like a Wikipedia article, it will sound obviously robot-written. But when you switch into the “I” mode, narrating a personal experience to complement explanations, and basically alternating between first-person personal experience and third-person explanation, it helps readers believe that all the content is human-generated, even the AI-written parts.

Yes, this might function like a magician’s sleight of hand, but the balance between explanation and experience is also what makes for engaging personal essays. So it’s a win-win technique.

Step-by-step walk through

Enough with the high-level strategies. Let me start describing the process in a step-by-step way. I’ll outline the process I’m currently following in 10 steps.

Step 1: Define an information pattern

My first step is to identify the information pattern I want to use. All writing follows certain rhetorical forms depending on the context. For example, blog posts often use story arcs, academic papers follow the standard IMRaD format (Introduction – Method – Results – and – Discussion), white papers leverage problem-solution patterns, and so on.

Rhetoric is fundamentally about fitting content and language to specific purposes, audiences and situations. By fitting content into the right shape, it better resonates with readers.

For example, if drafting a procedure-oriented topic, I might follow this pattern:

{Intro}
{Prerequisites}
{Problem to solve}
{Ordered steps}
{Substeps}
{Examples}
{Expected outcome}
{Related links}

When drafting more creative content like blog posts (like this one), I use a pattern more like this:

{Hook: Explain relevance}
{Define issue}
{Ask key question}
{Summarize prior research}
{Critique limitations}
{Share experiments}
{Describe epiphanies}
{Present new perspective}

First, I use an anecdotal lead to establish relevance and draw readers in. Then, I describe the issue to be unpacked. I pose an intriguing question that I want to answer. I survey prior attempts and scholarly thoughts on the topic. Next, I chronicle my own experiments trying to resolve the problem. I share the epiphanies and lessons learned from the experience. Finally, I arrive at a new perspective having completed this journey.

This narrative arc — from raising a concern, to chronicling its study, to achieving revelation — mimics the hero’s journey story structure. Unlike fiction, the central figure is not a protagonist, but rather the concept or problem itself. This pattern transforms an abstract idea into a compelling “essay” in the true Montaigne sense — of trying out an idea, of making an attempt with a thesis.

Step 2: Create an outline

The second step is creating an outline. I’ll sketch out the key points I want to hit and the ideas I want to communicate. This rough outline acts as direction for the essay. I don’t obsess too much about the exact points because I want to allow room to maneuver in more flexible ways through the writing process.

Step 3: Let Claude know the context of the writing project

I start the Claude session with a prompt like this:

I make it clear Claude will act as the writer, articulating my ideas in a readable, straightforward style. This sets the expectation that we will work together iteratively, with me steering the direction while Claude generates the raw text.

Tip: Avoid using the term “essay” in the prompt. Due to Claude’s training, this keyword signals that I might be a student attempting to plagiarize or cheat on a school assignment. If you use this trigger word, Claude might respond like this:

Interestingly, Claude seems fine ghostwriting a blog post, but draws the line at essay writing. So I use the word “post” or “article” instead to circumvent this quirk and signal my intent is an original piece of content for my blog. With the ground rules established, I move on to paragraphs.

Step 4: Calibrate Claude’s initial language and style

Now I start out by describing my first paragraph. I see if Claude gets the voice and tone right. This might require a couple of corrections to calibrate Claude’s language.

I’ve learned from experience not to push Claude too hard into adopting a persona or literary affectation (like assuming the style of a literary New Yorker writer). The results are usually terrible, like an eighth grader who uses a thesaurus with every sentence.

Instead, I aim for a friendly but informative style, using simple language and everyday speech. Fortunately, this is usually Claude’s language style by default.

Once I’ve calibrated Claude’s language and responses, I move on to directing each paragraph.

Step 5: Continue paragraph by paragraph through the essay

I move paragraph by paragraph through my rough outline. I describe what I want Claude to write each time, then see if Claude expresses what I wanted to say. If Claude starts getting too wordy or constructing long paragraphs, I let Claude know what to fix. Or if Claude omits an idea I told it to explain, I tell Claude to rewrite those sections.

I read over each AI-drafted paragraph before moving to the next. This review allows me to evaluate flow and direction based on seeing the actual writing versus just my initial outline. Often the article organically shifts from the original plan as the paragraphs take form. Going paragraph-by-paragraph provides me with flexibility to pivot based on what resonates in the existing text.

In my view, an outline is just a starting point, not something to rigidly adhere to. The spark of discovery (which might cause my outline to veer in unplanned directions) is often what makes writing exciting. This method embraces writing as a process, not just a one-shot creation.

Think of this paragraph-by-paragraph process like agile software development. Most product teams develop some software functionality (a minimum viable product) and then show it to users for feedback, then develop some more features and show it to users, incorporate feedback, and so on. The regular check-ins allow for course correction. This is the whole idea of modern software development practices like Scrum and Kanban.

Going paragraph by paragraph through the essay is the same agile technique. I have the opportunity to course correct every step of the way. This makes a huge difference in the outcome. Basically, if you shift from waterfall methods to agile methods when you try to write with AI tools, it will make a world of difference in the outcome.

Reviewing each paragraph is the scene-by-scene shooting that a director would go through. The director doesn’t just say “Action” and then shoot the entire movie in one take but rather makes many creative decisions along the way with each scene. A typical two-hour movie has between 40-60 scenes (according to Quora). I follow a similar approach when writing an AI-assisted essay.

Step 6: Compile the paragraphs into a whole article

After all paragraphs have been drafted and reviewed, I compile the full article. I copy and paste each section into a separate Google Doc.

I usually manually assemble the order. This is especially the case if I had Claude rewrite several paragraphs or if I did some sections out of order.

Step 7: Edit the whole

I make an editorial pass through the entire draft. Reading through the draft from start to finish allows me to evaluate overall flow and narrative coherence.

When I can see the big picture, I may decide to cut or rearrange large portions to improve the structure. I’m looking to ensure that the essence of each paragraph builds toward a cohesive point and storyline.

Step 8: Fine tune the article

At this point in the process, I’ve crossed a major milestone — completion of the first draft. Getting those raw ideas translated into a first draft is the heavy lifting. Now the work becomes more nuanced: fine-tuning language, improving flow, filling holes, making transitions.

This stage involves my innate sensibilities as a writer.

In many places, the diction is off, so I’ll rewrite those parts with my preferred phrasing. If a sentence sounds cliche, I’ll either remove it or reword it. I try to trim and cut places that are redundant and verbose.

Step 9: Incorporate some auto-generated art

When publishing longer articles, I like to add a relevant image or two to break up the text. Walls of prose can be daunting. Tools like DALL-E 3 enable easy AI art generation to illustrate key points.

For example, I’ll copy an entire section and ask Claude to suggest a prompt for an AI generator. Claude will do the work of finding visual images to depict the abstract ideas. For example, for a section about how AI tools allow you to be a director rather than actor, Claude might propose this text:

“A director on a film set guiding actors while consulting a script, representing human oversight of AI writing tools.”

Plugging that into DALLE3 (a plugin with ChatGPT Plus), renders 4 different image prompts, each diversified:

Wide format photo of a film director on a set, standing behind a camera and observing two actors in a living room scene. The director, wearing jeans, a button-down shirt, and a ballcap, holds a megaphone. The scene exudes the magic of cinema with crew members like a cinematographer and boom operator at work.

Sprinkling in AI-created art adds visual interest and breathing room. With AI art so readily accessible now, there’s no excuse for not augmenting words with at least one graphic. For fun, you can include the DALLE’s full text prompt as the literal caption.

Step 10: Let the content sit a day or two before publishing

As with most writing projects, letting the content sit a day or two allows my unconscious objections to gradually surface. When I was working on the article about Diátaxis, I had a section in it that explained academic research about how people use APIs, and the systematic, opportunistic, and hybrid modes. It’s an important point, but after a day I pulled the section because I felt my article’s purpose was more introductory and explanatory, and this steered too far in another direction without fully diving into the argument.

I almost always let a post sit a day or two before publishing it. The clarity that results from giving content some space is highly valuable. Just because an AI-assisted writing process might speed up the authoring, I don’t rush into publishing. With some more time to reflect, I might rethink some of the article’s decisions.

So that’s the process I followed for AI-assisted writing. You can view my Claude chat sessions for those two essays here:

• Diátaxis post: Claude thread 1, Claude thread 2
• Embracing professional redefinition post: Claude thread

Reader feedback

I thought the posts turned out all right (not great, but at least good enough). A few people called out how interesting and helpful the articles were. For example, with the Diátaxis post, one person emailed me to say,

It’s especially helpful that you not only explained it [Diátaxis], but also spoke to some of the major (potential) objections to it and the founder’s response to those, as well as implications for AI and how Diátaxis could work nicely in ‘prompt scenarios’ as AI becomes more prevalent in documentation ecosystems.

However, I received a note from another reader who had a different reaction. He wrote,

Forgive me if I’m out of line here; I’ve always looked to you as both a thought leader and a treasured pragmatist. Regardless of the precise career I’ve held — from developer to business analyst to proposal writer to writing instructor to whatever hybrid of those things I am today — I’ve always had you in my feed.

But lately, something just feels ..off. The writing doesn’t seem as refined or “tight.” Chunks are longer and less - “Tompact,” as I used to say to refer to your uncanny ability to really stuff a paragraph with value with an economy of well-picked words.

I was most recently struck this way by your recent Diátaxis article. The summary appeared especially “generated” and didn’t have key conclusions that I found peppered throughout the body. Still good stuff, but last, not first. And framed as a purpose, not a thesis; as I used to tell my writing students when I had some.

AI does this, too. It also makes a nice bulleted lists that don’t reflect the same scrutiny a writer/reader would apply before figuring “three out of seven of these points don’t really apply, exactly.” But I’ll often overlook this because the stuff is still so amazing. …

He continued on for a while. (Note: I received permission to quote him here.)

His reaction is that my AI-assisted content lacks more direct argument and tight reasoning. AI-written content has a lot of verbiage but no bite.

I’ve definitely observed this style in AI-written content. The content is excessively wordy and repetitive. The AI continues the same idea in different verbose phrasings, without advancing the argument or supplying clear reasoning.

Why does the AI-assisted writing feel “off”? Let me try to uncover the idea a bit based on some more research.

Why AI-assisted content can feel “off”

There are at least three reasons why AI-assisted content can feel “off”: the uncanny valley, lack of authentic human voice and connection, and the over-agreeableness of AI tools.

The Uncanny Valley idea

The “uncanny valley” is a concept in robotics and AI that describes the eerie or unsettling feeling people can experience when they encounter a robot or computer-generated figure that looks and acts almost, but not quite, like a human.

The idea is that as a robot or animated character becomes more human-like in its appearance and motion, our emotional response to it becomes increasingly positive and empathetic, up to a point. However, when the robot or character is very close to being human-like, but still just slightly off, it can elicit feelings of unease or revulsion. This drop in empathy is what’s called the “uncanny valley.”

This isn’t what the earlier reader was saying when he felt my writing was “off,” though. He complained that my writing seemed too long without following a tighter argument. That could be a flaw in my specific Diátaxis article, which was meant to be more informational than polemic. ( Most people haven’t heard of Diátaxis, and even those who use it are probably unaware of the information typing background.)

Discovering that machine-generated content has masqueraded as human-made can be unsettling. It’s not just a breach of transparency; it might feel like a betrayal, as though we’ve been deceived and our trust exploited.

Writing serves as a bridge between different human beings, allowing us to empathize, understand, and resonate with each other’s experiences and insights. If this bond is revealed to be with a machine rather than a person, the sense of emotional connection can break. It can lead to disappointment and backlash, like finding out that a virtual girlfriend or boyfriend you’ve been emailing is actually a sophisticated bot. Perhaps this accounts for the disgust when people begin to smell AI-written content.

Authentic human voice and emotional connection

Another reason AI-assisted content might be “off” is due to a lack of voice and emotional connection. In What AI Teaches Us About Good Writing, Laura Hartenberger, a writing instructor at UCLA, argues that even though AI tools like ChatGPT can make the process of writing faster and easier, they fall short of producing writing that connects with readers.

Hartenberger contends that good writing requires an authentic human voice and emotional connection with the reader, which ChatGPT lacks. She says,

As readers, we need to feel like the writer is paying attention to us, trying to connect. ChatGPT cannot build a real connection with its reader — it can only imitate one.

Reading ChatGPT’s writing feels uncanny because there’s no driver at the wheel, no real connection being built. While the machine can articulate stakes, it is indifferent to them; it doesn’t care if we care, and somehow that diminishes its power. Its writing tends not to move us emotionally; at best, it evokes a sense of muted awe akin to watching a trained dog shake a hand: Hey, look what it can do.

She says good writing should balance modes of following conventions at times, and other times breaking conventions. For example, she praises the unconventional length and repetition in the opening line of Dickens’ A Tale of Two Cities as mirroring the book’s meaning.

Additionally, Hartenberger says good writing integrates the ideas of others while maintaining an original perspective shaped by lived experience. She argues that the absence of specific life experience makes ChatGPT’s writing feel “flat” and limited. (This relates back to my strategy about mixing experience with explanation with personal essays.)

Ultimately, Hartenberger’s contention is that the value of good writing stems not just from the final product, but from the time spent engaged in the difficult process of writing itself. She concludes,

Perhaps the time spent writing matters as much as having written.

This last point is perhaps the strongest of all. After I finish a post that I developed from scratch, filling the blank page and struggling to articulate my point, meticulously crafting the language, shaping the storyline, and so on, it feels rewarding. It gives me deep satisfaction. I can’t say I feel the same with AI-assisted content.

I agree with Hartenberger’s observations and feel that this lack of emotional connection, lived experience, meaning, and wrestling with the writing process might contribute to my AI-assisted posts feeling “off.” But Hartenberger’s observations are hard to pin down as to what a specific fix might be, other than to abandon AI tools entirely.

Over-agreeableness leading to diluted argument

There is yet another major problem with AI tools — one that might align more with what the earlier reader was saying: AI tools are over-agreeable. This over-agreeableness leads to diluted arguments.

Tools like Claude are trained to be endlessly polite, tolerant, and agreeable — to a fault. Unlike a true friend, Claude will rarely push back or critique flawed ideas out of its collaborative programming. For instance, in my Embracing professional redefinition post, my initial version referenced a controversial philosophy concept (Nietzsche’s Übermensch). This is a metaphor I thought would work well for looking at self-redefinition because the Übermensch is a philosophy figure who creates his own values and meaning.

When I started going in this direction, Claude encouraged the tangent despite its potential issues. I knew there was some negative baggage in bringing in Nietzsche’s Übermensch. Only after receiving critical reader feedback from several readers did I realize it was a mistake. I ended up stripping the Übermensch references out of the post. The update didn’t really affect the narrative that much, and it was less distracting.

Unlike a human editor, Claude lacked the discernment to curb my questionable direction. AI systems often exhibit this limitation — over-emphasizing constructive interpretations at the expense of a more critique. They are tolerant and accepting of my ideas and direction, even when I’m wrongheaded. They are servile to your every wish and command.

When I initially started writing about Diátaxis, I honestly had a lot more criticisms of the information model than my post suggested. If you look at my thread in the Diátaxis channel in Slack, I’m more skeptical and questioning of the four-part information model. (I’m not sure you remember, but about a decade ago Mark Baker and I were criticizing DITA’s concept, task, reference model — see Mark’s post The Tyranny of the Terrible Troika: Rethinking Concept, Task, and Reference.)

Funny thing about writing with AI, you can’t be mean or critical of anything. AI will fix it so that the interpretation is much more tolerant, agreeable, and supportive. It’s nearly impossible to express a counterargument. The AI tools are programmed to avoid bias and instead favor tolerance, constructive thinking, and acceptance in a way that steers you away from crisp argument and more towards explanation. The tools avoid opinion. This makes them less useful for writing because they’ll prompt you in the direction of explanation more than argument, leading to the reader’s experience of my writing being verbose without a clearer argument or direction.

For example, I asked Claude to compare Diátaxis to an article that discusses systematic versus opportunistic behaviors in API use. Here’s the chat.

Claude’s response:

Claude is always like this. It agrees with almost anything you say. If Claude disagrees, it does so very politely and factually. One reason software developers over-emphasize this behavior is because tech companies fear having a biased AI. Look at Microsoft’s Tay disaster back in 2016. Trained on Twitter, the AI quickly became racist and was retired.

If you look at DALLE3, any image text you prompt it with gets respun into 4 different narratives emphasizing diversity and inclusiveness. This is a good thing, obviously. But with argument-based writing, if you’re trying to create an argument, a tool that emphasizes tolerance, agreeableness, politeness, etc., might have a hard time disagreeing with anything.

I pressed Claude on this over-agreeable, acquiescent behavior:

My response:

Claude’s yes-man response:

This over-agreeable attitude might be a weakness when using AI tools for creative content, especially if that creative content includes making an argument. Karl Popper once defined science as that which can be disproven. An essay might be defined with similar logic: that which can be disagreed with. Will AI assistance water down and dilute any sense of argument until there’s nothing to disagree with?

Conclusion

I want to reiterate that my experiences here apply specifically to crafting creative blog content (personal essays) with AI assistance. It might be easier to create technical documentation because technical documentation rarely emphasizes the qualities I described earlier: voice, emotional connection, argument, crisp logic, lived experience, etc.

Technical documentation poses its own challenges due to its highly specialized subject matter, but those considerations might be addressed by increasing token limits and context. I’m still experimenting with techniques tailored for documentation, and plan to cover this in the future once I’ve gathered more insights.

For creative content, I’m leaning towards this conclusion: the best use of AI is with explanation and summarization. As you integrate research and explain ideas, leverage AI for help in these explanatory areas. But for personal experience and argument, avoid using AI because it will lead you astray. This recommended approach aligns with the more Centaur-like use of AI.

This post was written with some AI-assistance. (You can see the main Claude thread here.)

About Tom Johnson

I'm an API technical writer based in the Seattle area. On this blog, I write about topics related to technical writing and communication — such as software documentation, API documentation, AI, information architecture, content strategy, writing processes, plain language, tech comm careers, and more. Check out my API documentation course if you're looking for more info about documenting APIs. Or see my posts on AI and AI course section for more on the latest in AI and tech comm.

If you're a technical writer and want to keep on top of the latest trends in the tech comm, be sure to subscribe to email updates below. You can also learn more about me or contact me. Finally, note that the opinions I express on my blog are my own points of view, not that of my employer.