When I first noticed the trend of people using artificial intelligence (AI) chatbots like ChatGPT to generate written content, many “writers” seemed not to realize how obvious it was that they were copying and pasting content generated by the tool.

The text would often come complete with all the emojis, divider lines, neatly subheaded sections, ordered or unordered lists in lieu of paragraphs, frequent bolding for emphasis, an overabundance of em dashes for dramatic pause, and that same generic prose as everyone else using ChatGPT to write their stuff.

You’d also see the AI “hallucinations”, such as fabricated quotes, appear in the copied and pasted material.

Another clue is when links to sources include the UTM parameter showing that they were generated by ChatGPT (?utm_source=chatgpt.com), which could just mean the author used the tool strictly for research but could also indicate copying and pasting of AI-generated material.

You see less of the obvious stuff now as people have learned how to train chatbots to mimic their own writing voice, if they have one, and the AI models have gotten better at being able to do that and otherwise to generate text that doesn’t scream its origin so loudly.

ChatGPT has improved a lot, and I infrequently encounter hallucinations now. When it does occur, it tends to be the tool accurately paraphrasing while making inappropriate use of  quotation marks, as opposed to the types of wholesale inventions I used to see it generate.

There are still hallmarks of AI-generated text, of course, but it has gotten generally less obvious, even if more widespread than ever.

It was almost more ethical to at least remain obvious about using AI to generate content than what you see commonly now, where you suspect it but really can’t be sure.

As an independent researcher who shares my findings and analyses via the written word, the ethical question often comes to mind.

What constitutes an ethical use of AI in writing?

To start with, it depends on the type of writing. Using AI to generate authorless marketing copy is one thing. Using it to generate “literature” to pass off as your own art is another matter. I’m going to focus mainly on its ethical use in journalism, with the understanding that I view my own informative writings as a form of art.

Hence, to me, the idea of using AI to write for me is repulsive not merely on ethical grounds but also because it would steal the joy of writing, the satisfaction gained from crafting word combinations to effectively convey my ideas in ways that are distinctly my own. It’s my ideas I want to share, not some generic conglomerate of everyone else’s ideas on a given topic found on the internet. It’s my voice I want people to hear, not some programmatic facsimile.

Of course, some might object that there is no such thing as “ethical” use of AI. It’s an evil technology that will be used by the government in collaboration with tech corporations for population control. It requires massive data centers that raise concerns about property rights and eminent domain, energy and water consumption, harm to the environment, and impact on local communities. It shouldn’t exist.

It’s not my purpose to evaluate or address those concerns here, merely to acknowledge them. I’m going to presume that ethical use of AI technology is possible. Like any technology, I don’t view it as inherently good or evil. I place my value judgment on how it is used, not the tool itself.

And I’ve benefited greatly from using AI myself. I use both ChatGPT and Claude, each for different purposes and sometimes in a complementary way, and the productivity gains have been tremendous.

So, I suppose a good place to start with my musings on the ethics of AI in journalism is to disclose my own uses for it.

My main use of it, and the task I initially started using it for, is research. It’s incomparably more efficient than a traditional web search with a search engine like Google. It is somewhat infrequent now that I do a regular search in my browser. And the AI-generated responses to search engine queries also aren’t equivalent to just using ChatGPT or Claude for the job.

Those AI summaries raise another important concern about how AI is trained on publicly available information, so a response might include information from a source, yet the user never has to click the link to view the source to acquire the information. Even if credited with a link, it deprives the creators of the value of the search traffic.

Journalists might do all the work yet get no credit or benefit from their information being provided by a search engine in an AI-generated response. The Intercept has sued OpenAI over this very problem, and there are other such lawsuits, but, again, it’s not my purpose here to delve into such concerns but to focus on the ethics of using AI for content generation.

I most often search not for general information about a topic but to find precise answers to specific questions, or to locate some primary source. Prompting ChatGPT is much faster than digging for it with various search terms and clicking through multiple links only to find that most of them aren’t helpful. It basically does the work for you of running multiple searches at once, evaluating results to narrow the scope, and reviewing sources to help you find what you’re looking for.

Of course, you have to always verify the information provided by consulting the sources, but that’s a given. And these days, I find ChatGPT performs fantastically at accurately summarizing or characterizing information. It’s much better than it was just a year or two ago.

I also use it for learning and problem solving, with everything from home maintenance to website development. It’s great for providing tutorials, as well as cutting down time spent on product research and making buying decisions. In numerous instances now, when I’ve encountered a problem with or limitation of the WordPress platform and my tech stack of plugins, either ChatGPT or Claude has generated code snippets or plugins to resolve it for me.

As a chatbot, I’ve also found it highly valuable for testing my arguments. There’s a common misconception that ChatGPT will either always tell you some establishment narrative or, conversely, that it will always agree with you. Neither is true.

Yes, owing to its nature as a large language model (LLM) that generates responses based on the conglomerate of input data it’s trained on, it will tend at first to give you more generic information of the type you’d find in the mainstream media or by doing a basic Google search. But you can challenge it, and as a machine built on coding logic and trained on the vast record of human reasoning, it is not unreasonable. It will initially hold firm and politely push back, too, but if you have your facts straight and apply sound reasoning, you can get it to concede points and even reverse itself.

It won’t just agree with you if you don’t have your facts straight and/or don’t use persuasive reasoning. If it identifies a fallacy in your argument, it will tell you. So, it is a very helpful tool for testing arguments and learning how to better anticipate—and answer—the types of objections you’ll face when publicizing them.

For that matter, it’s also a useful tool to correct course when it sufficiently demonstrates that you really are wrong about something.

I’ve come to use ChatGPT also to summarize video content because I generally dislike watching videos, for several reasons.

For one, it is simply so time consuming relative to reading. I can consume information much faster with the written word. I can skim and scan text. I also like to highlight text as I’m reading, and I save articles and PDFs to my personal research archive for future reference. Watching video content, you lack convenient access to sources. Even if a specific source is cited, it’s often laborious to track it down. With articles, you have hyperlinks to be able to easily access cited sources.

And, yes, checking cited sources is essential. It never ceases to amaze me how routinely I find that a source does not support the claim for which it is cited—or even contradicts the claim. Examples are routinely included in my articles dissecting mainstream propaganda (like this one about a New York Times article citing a study that found virtually the opposite of what it reported).

The video format is obviously preferable for some types of content, like “how to” videos where it’s helpful to actually see the thing being done, but for general learning and research purposes, I very much prefer the written word as a matter of efficiency.

YouTube now provides AI-generated transcripts of videos, so I tend to either just scan the transcript or copy it into a text file and upload it to ChatGPT to summarize key points for me. Sometimes, the result is I’ll watch at least parts of the video, whether to learn the fuller details or verify information. Sometimes the quick summary is good enough, and I move on.

Another way I use ChatGPT is to generate images. Sometimes, for example, I need a featured image for my articles but can’t find anything usable in the public domain or under a suitable Creative Commons license, so I’ll have AI generate something for me.

AI image and video generation capabilities have advanced so far that you often can’t tell whether something is real or not. This raises serious concerns about deepfakes and hoaxes.

Of course, I would never generate an image with the intent to pass it off as a real photo. When publishing such images, it is either obviously AI-generated, or I say so in the caption.

I more recently also started using Claude, at first to see whether I should switch to it from ChatGPT. I ended up using both because they each have features and benefits the other doesn’t. I sometimes use them together.

For some examples of how I’ve put Claude to use, I have built various tools with it to improve my workflow.

One is a tool for headline optimization. Back in the day, I got pretty good at being able to rank highly in Google search results for specific terms, and organic search was the greatest source of website traffic for me. This required doing a bit of search engine optimization (SEO) keyword research. But then Google disappeared my site, the SEO research lost its utility, and I became dependent on social media to get my messages out.

Of course, then the social media companies also started censoring massively, making it a real challenge to get my articles found. It’s eased up since its peak during the lockdown madness, but it continues in various less overt ways, particularly by virtue of algorithms designed to favor “authoritative” sources—which is to say mainstream sources that serve to manufacture public consent for whatever political or financial agenda by perpetually misinforming people.

The faux “fact check” industry, as I’ve often observed, exists for that very purpose, but that’s another topic.

Back to the point, while I basically gave up on SEO, the importance of writing a good headline remained as crucial as ever. Whether in web search results or a social media post, you have a second or two—or a fraction of a second—to grab someone’s attention with a headline, then another few seconds to provide the reader with a compelling reason to click through to your content. This requires your headline to contain a value proposition, a promise to deliver something specific.

This can be a real challenge when you have a complex thought to succinctly communicate with a character limit beyond which your title will be cut off in various feeds where the link might appear. It’s helpful to have someone—er, something—to bounce ideas off of.

So, the tool I built in Claude essentially assesses my working headline and provides data-driven input about SEO and answer engine optimization (AEO), or generative engine optimization (GEO). I don’t expect AI tools to regard my writings as “authoritative” enough to really drive traffic to my site that way, given how I constantly challenge mainstream narratives and supposed authorities, but the same principles that make a headline work for getting content found also apply to getting your target audience’s attention on social media and elsewhere, so it’s a useful exercise.

My instructions for the tool specify that my headlines should communicate a clear value proposition without being in any way click-bait-y. I provide the text of my article along with working headline so it can assess and ensure that my content delivers on the title’s promise. (This process also serves as a final fact-check as it will highlight any claims it regards with skepticism or as requiring solid sourcing.)

It accounts for what Claude termed, during the process of developing the tool and creating effective instructions, my “latent audience problem”: a target audience often consisting of people who most need the knowledge I’m providing but who may not know they need it and so aren’t actively searching for it.

Another factor it considers is algorithmic censorship. It warns me if it thinks my phrasing might result in my article being suppressed or even banned on social media. (Indeed, while I managed to survive on Facebook and Twitter during the lockdown insanity, I was banned “permanently” from LinkedIn for citing CDC data to correct CDC disinformation, although I did manage to eventually get my account back.)

My Claude tool tries to balance various factors such as search volume for keywords, competition, and audience alignment. It offers alternative suggestions to help me brainstorm. Sometimes the result is I convince it that my original idea is best. Sometimes it reveals to me a need to go with something else. I may or may not choose to use its wording, but either way, the substance is mine.

For this article, to illustrate, after bouncing around several ideas and the rationale behind each, I suggested “Ethical Use of AI in Journalism: How I Use It and How I Won't”, and Claude pointed out that it exceeded the character limit and included a clunky repetition of “Use” as a noun and then verb. Hence, the title this article bears, now that it’s published, is the fix it suggested: “The Ethics of . . .”

Another tool I built in Claude reads my article text and suggests concepts for a featured image, in instances where something usable is not readily found online. I go through a process of refining the concept, and then it generates the prompt for me to copy and paste into ChatGPT, whose image generation system is quite phenomenal. (Much better now than it was just a year ago.) Getting the result you want with a generated image isn’t all that easy, and it’s a skill I don’t need to take the time to develop since I have my Claude tool doing that for me. I get far better results using its prompts—and without all the time wasted with trial and error and still getting it wrong on my own.

A third tool I built reads my articles and identifies potential “hooks” to use for social media posts. Similar to my headline optimizer, the idea is not for it to write social posts for me but to speed up the process by helping me brainstorm and assess potential for reaching an intended audience with particular excerpted messages.

Once an idea is locked in, it does suggest post text, but in doing so it is instructed to rely heavily on my own wording from the article. It basically summarizes my own key points in a way that saves me time, providing me with a clear starting point in lieu of a totally blank slate. The words used, in the end, are my own, whether because they are taken directly from the article or because I’ve rewritten its suggested (often poor) paraphrasing back into my own voice.

I’ve also been in the process of developing a tool to perform automated website analytics so I can make more data-driven decisions about how to better serve my readers, giving them more of what they want—or, as often is the case, what they need to know whether it’s what they want to hear or not. Because of the need to connect my analytic tools, this project is more complicated and still a work in progress, but it will be an extraordinarily helpful tool once I overcome current obstacles.

On that note, just so my readers know, I do not use Google Analytics or any other third party that requires me to place tracking code on my site and to place a “cookie” in visitors’ browsers. I instead now use a self-hosted solution for traffic analysis, a WordPress plugin that happily enabled me to ditch Google Analytics about a year ago.

I have also used Claude this year to build a mobile app!

I needed one for time tracking, and nothing on the market met my criteria. I spent so much time looking into options and testing several out only to be met with frustration, I finally gave up and figured I’d just try using Claude’s ability to try coding what I wanted for me. It started out as basically an experiment, but I was so satisfied with initial results and ease of use that I kept at it, adding more features and improving the user interface and experience so that now I have a fully featured app unlike anything else out there installed on my phone. (Sorry, it’s not available in the app store—yet.)

Anyhow, that particular Claude project is tangential to my point here, but it’s a good example of the utility of AI for a non-coder like myself. I’m very happy and excited about having been able to bring my vision to life. It’s been quite a fun project, and a welcome relief in my “free time” from the dreadfulness of the issues I focus my journalism on (war, corruption, deception, rights violations, etc.).

Back to the issue of writing, I have always had a habit of making generous use of the dictionary and thesaurus to communicate my ideas more precisely, and this is another way that I now use AI. For example, there are times when I have a word on the tip of my tongue but just can’t think of it, and I just tell ChatGPT I’m looking for a word to express such-and-such a train of thought, and it reliably reminds me of just the word I was looking for.

For example, I prompt, “There's a ‘p’ word I can't think of that could describe, e.g., an explanation that makes sense and explains a matter, well, efficiently, for lack of this better word. Help me think of the word that's on the tip of my tongue but not coming to memory.”

And sure enough! The LLM reminds me of the very word I’m looking for: “Perhaps parsimonious? A parsimonious explanation accounts for something with relatively few assumptions or unnecessary complications—economical but still adequate.”

Yep, that’s the one! Why do I always forget that word? I suppose it’s because I use it rarely but, well, parsimoniously. (ChatGPT suggests that “judiciously” works somewhat more precisely here, but why lose the fun wordplay?)

Getting help with organizing and outlining ideas and information is another way I’ve used AI. One of my research projects has involved reading PDFs of three scanned primary source documents from the CIA regarding the 1953 coup in Iran. My PDF software can render scanned pages to text so you can search and highlight, which was needed for the oldest of the documents, from 1954. For each file, I highlighted and made notes in the margin with textboxes. The documents are quite long, containing a somewhat overwhelming amount of information, with each jumping around the timeline a lot. I wanted to organize the source materials to tell the story more cohesively and chronologically, but for me to have gone back through the files to pull out my highlighted parts and organize them into an outline while ensuring each aspect is properly sourced to the right document would have been days of work.

So, instead, I uploaded my marked-up PDFs to ChatGPT and instructed it to produce the chronology I desired with intact references to source and page numbers. It produced what I needed in minutes, instead of days or even weeks, thus enabling me to focus my time on the outlining and writing. Instead of spending hours just trying to find each highlight or note I had in mind to use, I had a quick reference guide to be able to quickly jump to precisely the location in each document from where I wanted to draw information or direct quotations.

Of course, I did refer back to the source documents to ensure verbatim quotes instead of trusting ChatGPT to faithfully reproduce them, although in this case, it indeed produced quality results.

What a tremendous time saver that was! It’s still been a highly laborious project to piece everything together, but not nearly as unfun as it would have been had I not utilized AI to help me get the information organized meaningfully.

As I said at the start, I do not and wouldn’t want to use AI to write for me. Writing is an art form that I enjoy. But sometimes it does help to have an editor or a way to bounce ideas around. I might ask AI to proofread for me and identify spelling, punctuation, or grammar issues. (I don’t do that often but probably should, actually.) By contrast, I wouldn’t ask it to make substantive changes to an article I’ve written. And I certainly wouldn’t prompt it to write the thing for me.

As much as I might hate to admit it, there have been times, when using AI to explore thoughts, that it did reformulate my ideas with wordings that I wish I’d come up with. On a few occasions, I ended up incorporating such a line into whatever article I was working on. But, fundamentally, it was still my idea and, ultimately, my choice of words.

Using AI as a tool to help you find your voice is very different from using it to be your voice.

Using AI as you might a human editor is a far cry from using it as a ghostwriter.

It can be a very useful tool for research and fact-checking, brainstorming, problem solving, challenging your arguments, and getting suggestions for how to more precisely or effectively communicate a thought.

In my view, this all falls within ethical usage.

It crosses the line, however, to have AI write for you, or to take what you’ve written and substantially rewrite it.

With literature in general, passing off a chatbot’s text output as your own art or original ideation is at best poor taste. With journalism of any variety, in my view, it crosses the ethical boundary.

There might be instances in which it makes sense to publish AI-generated text, but this should be disclosed and never passed off as one’s own original writing.

The apparent habit of many writers on Substack and elsewhere to utilize AI for content generation is one I find revolting.

It’s a fact of life that the pretenders are out there now. Sometimes obviously, but other times with enough human input that it’s hard to tell the difference.

The key tell, ultimately, is the voice. Suddenly, everyone sounds the same!

My own goal as a writer is to express myself honestly. It isn’t logically possible for me to do that by having AI generate content for me, so my own readers really have nothing to fear in that regard.

Just remember, if you’re a writer, what your purpose is for writing. If you have to cross ethical boundaries by using AI to get a message out and passing it off as your original content, then perhaps you should reevaluate your goals.

If you wish to use AI to generate propaganda, well, you’re the one who has to live with yourself.

If you’re well-intended, setting the ethics question aside, do not rob yourself of your own voice.

Let your writing be yours.

If you put your name in the byline, own it.

Instead of depriving yourself of the practice, work at it.

Whatever you do, just don’t join the pretenders.

In this new age of AI, the world now needs genuineness more than ever.