search
Hilary Faverman
If a storyteller and a grammar nerd had babies, they would birth us.

ChatGPT isn’t taking my job, but it’s already changed how I work

The AI tool will save my clients time and money, but as long as they want innovative, creative work with a sense of humor, they (thankfully) still need me
Using ChatGPT. (Facebook)
Using ChatGPT. (Facebook)

Middle age is all about adaptation, in the interest of remaining both relevant and sane. When my eldest was a toddler, success sounded like an unprompted please or thank you. This morning, I celebrated (discreetly — she is a teen, after all; parental displays of emotion are discouraged) when she came to me to edit her army entrance essay, rather than turning to ChatGPT. 

This is success in 2023. 

I was first introduced to machine learning and AI models in 2004 when I made aliyah with a career that didn’t transition to Israel, zero Hebrew skills and a distinct distaste for the most affordable meal here: falafel with a side of intimidating Israeli street culture. 

To make ends meet, I prepared wedding flowers alongside Russian immigrants who taught me to swear in Arabic, answered secretarial calls at 6 a.m., six days/week from an entrepreneurial physician, and tried desperately to claw my way back into the white collar world. My first “break” here was working with Trendum, an Israeli-based Nielsen company founded to study and monetize natural language by categorizing online sentiments to train AI models. I drove from Jerusalem to a Tel Aviv suburb, sat in their office for a series of six-hour stints in a windowless cubicle, and analyzed comments on comment boards, categorizing inclinations into “loves this product,” “hates this product,” or “feels neutral about this product.” I had no understanding that I was helping form the basis of what evolved into AI-content models, or today’s ChatGPT.

I was so in the dark, in fact, that when I attended a content marketing conference last year, I intentionally skipped the session about AI-generated content, assuming it was a clunky, irrelevant tool.

Oh, baby, it is not.

ChatGPT, the first free and publicly available natural language processing tool driven by AI technology, clocks in with an IQ of 147. It has effectively read over three billion pages of text, holds a thorough understanding of natural language, and predictively generates what comes next, both in terms of logic and wording.

Prom, 1993: Ethan and I, who had a solid laugh about ChatGPT and the article mentioned, were good friends throughout high school and college, and pledged that if we weren’t coupled by 30, we’d marry each other. Five years later, I fell for an Israeli, and sent that Israeli’s ex-girlfriend to New York to meet the then-budding business school professor. He married her, and I moved to Israel to marry the Israeli. Complicated, but fabulous. (courtesy)

It passed the bar, as well as the CPA and American Medical Licensing exams. After publicizing that it passed a class at Wharton Business School, one of Wharton’s professors (who happens to have been my prom date, once upon a time) insisted that it only earned a B.

Netflix took three and a half years to reach one million users.

Facebook took 10 months.

ChatGPT did it in five days.

What does that mean for writers, and for those who hire them?

I’ve been writing professionally for nearly a decade, and the marketability of this skill has allowed me to build a successful business while raising four kids and working from home. I employ seven other writers, all Israeli immigrants, many of whom write for me while they’re working at least two other jobs. 

Then, two months ago, ChatGPT broke my world.

First, I got scared.

Then, I got busy.

Whenever I’m faced with something potentially scary, I revert automatically to research mode. I educate myself, since we all know that unspoken things — unknown things — unpredictable things (think Poltergeist, Voldermort, or Vecna, depending on your generation) are intimidating. Luckily, I have a sharp-witted, diverse, adaptable team, so we got to work. We played, we took classes (yes, there are classes already), we ran 800 zillion tests, and we documented. We learned about structure, plugins, sources, what it can do and what it can’t. We learned that we still have a job to do… but that job looks markedly different than it did 62 days ago.

Let’s break it down. Here’s what we know:

What ChatGPT does well (already)

  • Its texts pass a plagiarism check but can be identified as AI-generated content; there are resources that are keeping up with it, created and utilized by educational environments.
  • Facts, syntax and grammar are 95% accurate on the first try. 
  • It can be directed to pull information from specified sources on the live web, and is not limited to information published prior to 2021.
  • It compiles research instantly, arriving at a “Frankenstein document” and outline, although sources can be sketchy.
  • After refining a query several times, it arrives at an editable rough draft covering all the main points.
  • On a first query, it generates nearly-100%-usable customer responses and SEO metadescriptions, which usually sound auto-generated anyway, so it’s an ideal usage.
  • It’s highly adept at brainstorming; it generates  keywords and topics within and adjacent to any space, which is ideal for creating content calendars. 

What ChatGPT does poorly (at least, for now)

Almost everything it generates sounds plausible and coherent, but:

  • Thought leadership pieces are impossible to generate, even when an opinion is stipulated, since it lacks original insights.
  • The language it produces is formulaic and smells of high school reports; attempts at storytelling or persuasion are unusable.
  • It’s unable to connect a client offering to pain points in a relevant, illustrative, compelling  format; ultimately, it fails at WIIFM (“what’s in it for me”).
  • It cannot judge and select relevant, significant, appropriate sources or quotes.
  • Its transitions between paragraphs and topics are either absent or forced.
  • It is unable to promote specific client collateral or backlinks; “plugs” have to be crafted and inserted manually.
  • It has a tendency to repeat wording and entire phrases; we’ve observed exact repetition up to 8x within 1,000 words.
  • The humor. It’s just… cringeworthy.
  • It fails at mimicking branded voice.
  • It is unable to rewrite an existing blog to retain its value but exist as “fresh flesh.” I had an incredible mentor who was 50 when I was 25. She was independently successful, had been married and divorced twice, pivoted half a dozen times, and when I asked her about the biggest challenge I would face as a middle-aged woman, she said, “Honey, you can reinvent yourself however you want, but you can never become fresh flesh.” I’ve kept that with me, and use it to refer not only to marriage, but also when talking about content that needs to be “respun” and launched again while avoiding being detected and flagged as “recycled/unoriginal/reposted” by search engines. Middle-aged content and middle-aged people — ChatGPT helps neither of us become “fresh flesh.”

Data only through 2021? Nope, it’s current

Initially, 90% of the criticism of ChatGPT content complained that it was only drawing upon the three billion pages of text published online through 2021. (We, as content marketers, sound spoiled, don’t we? “Only three billion? Pshaw, that’s not good enough.”) 

The problem was solved immediately with their Web ChatGPT plugin, which, although tricky to use, and lacking a Jasper equivalent (for those in the know), allows users to pull comprehensively (through the current date) or target a date range, like the last three days or week.

Storytelling: Content must compel 

Every piece of marketing content needs to incite action. Yes, we’re tasked with informing and enlightening, but bottom line, content needs to move someone to action: feel, react, click, navigate further into the website, share, donate, convert, download, register, or even drop a name or a fact at a cocktail party, thus creating “buzz.” 

The secret sauce of how to do this is not so secret. Relevant stories, examples, quotes, statistics and anecdotes make content interesting, potentially entertaining, and therefore consumable, trustworthy and convincing. The task is to find the intersection between the pain point the audience is feeling (“niche audience profiling”) and what the client is offering, in a manner that moves that audience. 

ChatGPT fails miserably at this.

Sources: ChatGPT identifies them, but its judgment is drunk

Yes, almost everything ChatGPT generates sounds plausible and looks reasonably coherent. After twisting and pushing our tests, we figured out how to get it to list its sources so we can evaluate them, and then we stood on our heads and bounced three times (OK, we just tested further and further), and finally arrived at the key to requiring it to pull from specified sources. We documented those instructions here for everyone, since we couldn’t find them published anywhere online.

Yes, it will list its sources.

But no, you can’t actually count on them. Marketers must evaluate those sources with the judgment of an Israeli sporting full “frier” (sucker) guard. 

We asked it to produce a list of popular blogs that cover remote work, listicle style. When we required it to list its six sources, three were dead links, one was a marketing site for a psychic, the fifth was a direct competitor of our client, and the final was qualified and usable. If you don’t check and push it for sources that you (or your client) consider reliable, you’ll end up looking foolish, publicly.

Humor and branded voice: “Can I get a hell yeah?” (Kill us now)

Oh, the cringeworthiness.

Attempts at humor consistently fell flat and landed like the proverbial dad joke. What we were most interested in, however, was whether we could get it to mimic an established client voice, since we (like all qualified marketers) cultivate a branded voice for each client. Could we point it to the sources we dictate (which we figured out how to do, see above) and get it to pull information and direction not only from the content, but the voice as well? 

That would be a big, fat no. It can’t (yet) mimic a voice we dictate; all attempts failed. 

It can spin content into a pre-programmed public personality’s voice like Oprah, Tim Ferris, or Suze Orman, when you give it a specified audience, the attempts at voice are unpublishable:

When you default on a loan, the first thing that happens is that your lender will start blowing up your phone like a teenager trying to get a date for prom. They’ll be all like…

Let me tell you, this is not something you want to happen. My dear friends, that’s not something any of us want to happen. Listen up, folks. Business credit is like the secret sauce in your business recipe. Can I get a hell yeah?

Facts: Accurate, yet valueless and unconvincing

The tool, in its current format, offers endless, accurate facts. Unfortunately, they are often redundant, obvious, irrelevant, tertiary, or even contrary to the client’s interests. Let me explain…. No, there is too much. Let me sum up, with an example: 

We worked within ChatGPT to write an American small business finance blog covering defaults on merchant cash advances and other funding options. Our client offers small business financing to high-risk candidates, so the explanation had to be directly tied to the audience’s pain point (I’m in a tight spot cash-flow-wise and I need funding; what can happen if I do X rather than Y?) and discuss the consequences of various default situations, positioning the reader to be wary and draw their own (nuanced, suggested) conclusions that merchant cash advances aren’t the way to go in these situations. Of course, the content was to be “informative,” without mentioning the client or their services.

Here is some of the AI generated content — accurate but valueless, both to the client and to the content consumer:

Defaulting on an MCA can damage a business’ credit score. 

A legal judgment against a merchant who defaulted on an MCA and was sued by the lender may be recorded on their credit report. 

Consider crowdfunding platforms as an alternative funding resource. 

Everything it generated was in this style: factual but offering no value. As compared to human-generated content that is both factual and relevant to both parties:

If you default on your merchant cash advance, the funder has the right to…   

A merchant cash advance default lawsuit is extremely common; that means…   

Even if you did not put up hard collateral for your MCA, the funder can still require you to…   If you are served with a summons…

If you fail to respond, the court can order a default judgment. That means…

The difference between false and phony

One of the best reviews of ChatGPT I’ve read — and believe me, I’ve read many — is Tim Leberecht’s, in Psychology Today. He refers to moral philosopher Harry Frankfurt’s seminal book, On Bullshit, explaining: “The essence of bullshit is not that it is false but that it is phony.” Tim’s contention is that the difference between a bullshitter and a liar is that the liar knows what the truth is, but decides to take the opposite direction. A bullshitter, however, has no regard for the truth at all. 

As marketers, we have to decide — and let’s be honest, our clients and the market will force our hand in that decision — whether our content is accurate but phony, and if we’re comfortable with bullshit. 

What ChatGPT really costs

So far, searches cost ChatGPT $3 million monthly or $100K/day. Right this second, it’s free to users, but have you noticed how often you get the following message?

(courtesy)

There are dozens of articles offering lame workarounds, including refreshing your page, using a VPN, opening several windows and trying them all in turn, and the genius instruction to “simply wait and try again.” But there is zero published documentation on how often the system is overloaded and rejects queries. It happens, in fact, constantly, especially at “peak times,” which no one is defining.

Enter the upcoming Professional Version starting at $42 monthly, which is not yet available and was only announced this month via OpenAI’s post Discord post, offering a waitlist application for the upgraded, paid version. Supposedly, the Professional Version will offer no blackouts (the error message displayed above) and an unlimited number of messages within ChatGPT, or “at least twice the regular daily limit.”

The critical piece to note here is that the cost remains undetermined. Within the waitlist application itself, OpenAI is conducting market research regarding how much users are willing to pay. These costs, once applicable, will be passed on by marketers to their clients who elect to use the tool.

What we, as content marketers, need to become, and fast

You’ll notice that I use the term “content marketer” here, instead of the myriad of titles we are often graced with, from copywriters to content strategists to (my favorite) storytellers.

The difference, as I see it, between a “writer” and a “content marketer” is experience and success in written persuasion, storytelling, identifying and writing toward pain points, and the nuanced judgment and consciousness required to perform critical thinking. That is our starting point. 

Turning those talented, experienced content marketers (ourselves, our teams, our colleagues, our community) into “prompt engineers” is what we’d all better be doing right this second.

Marc Cuban coined our new title. Since success with ChatGPT depends upon asking it the right questions, a prompt engineer (it’s a Forbes article, behind a paywall, so the resourceful among us can read the free version over here on Yahoo Finance) will become both a skill and a title. A successful prompt engineer will be the one who can optimize the tool to arrive at the highest quality rough draft possible, nearly eliminating the initial research and compilation phase of content marketing. In effect, an experienced prompt engineer — a content marketer who uses the tool to its fullest — will save 25% of the time it takes to create factually accurate, consumable, compelling piece of content. That is, equivalent to the quality produced by humans alone, but 25% faster.

Of course, I thought it apt to ask the ChatGPT equivalent, Dall E2, to produce an image of a prompt engineer. After massaging and tweaking the prompt, DALL E2 constructed this souped-up, magical-hero-content-marketer-turned-prompt-engineer. I guess this is what we look like now:

(courtesy)

Embrace it for what it can do

It’s time for all of us to make decisions. Educate, embrace, adapt, and evolve. Talk to your clients and employers about what ChatGPT can do for their content, and whether it’s a useful tool, depending on your goals, methodology, and the type of content you’re producing. It could potentially save their bottom line 25%, retaining the resonating quality they’re accustomed to from you.

As my new favorite writer Tim concludes,Writing, as a transcendent act, will remain inherently human. Like us, AI has the data. But unlike us, it lacks the self-awareness to struggle with it. It has the intelligence, but not the consciousness. It can’t really think. Thinking is hard; critical thinking even harder, and ChatGPT isn’t good at either. It just rehashes what has already been said. It regurgitates; it is one big recycling machine.

But all machines are tools. Learn them. Use them.

About the Author
Hilary Faverman Communications creates valuable, informative, inspirational content your clients want to consume.
Related Topics
Related Posts