Writing & communications

1. AI is always over-confident

Part of the problem is the prompt. Most people simply don't know how to say what they want, and poor instructions always lead to poor writing.

Right out of the gate, this exposes one of AI's biggest issues: over-confidence. It'll make do with as little information as possible, forging ahead as if all is well, and giving you a pat on the back along the way.

I asked Claude (my AI) why AI does this, and it said there are two main reasons:

• "A language model has no innate "I don't have enough information" signal."
• Humans don't like to be asked extra questions. They view it as friction, and human reviewers penalise AI models for friction during training.

A good human writer knows when they don't have enough info, and they'll ask for it, even when they know you'll get a little impatient.

2. AI doesn't truly understand your story

When you tell me what you want, I don't just hear what you say. I hear how you said it, and what you didn't say. I study your non-verbal cues, and I infer meaning from how other people treat you and talk about you. I imagine myself in your position, and in your reader's position. And on top of all of that, I overlay my own personal experiences and my memory of hundreds of previous clients.

Then I tell your story.

An AI is incapable of all of this. It can't read between the lines, and it has no personal experience to overlay. It's never felt tired. Or pressured. Or impatient. It's never had cancer. Imposter syndrome is just a concept it read about. It's never had to make the choice between taking the safe, full-time job or launching a business instead, knowing that choosing incorrectly could mean losing the house. It's literally incapable of empathy.

AI just rewrites what you tell it. It uses well-constructed sentences and a logical structure, but it only tells the reader what you thought to explain, not what the reader needs to hear.

3. AI doesn't care

When I write, I keep going until it's just right.

I don't mean rough draft —> fix errors —> final proof. I mean I write what I think is needed, then I read it with my end-reader hat on, then I debate with myself about whether it's good enough. Is it embarrassing, or trying too hard, or too heavily stylised? Does it sound trite? Is the tone too casual? Is it too hard to read? Should I have asked that as one question or four? Is my nautical metaphor too much of a stretch? (HINT: It was.)

Then I tweak, read and debate again.

And let me tell you, this costs me a lot! For example, I cut 4,874 words from this web copy. That's 19 pages of writing that I initially thought might be right!

I do this because it's who I am. My clients pay me for quality, so that's what I deliver. It would be unethical, unsatisfying and damn-near unpossible for me to do otherwise. Nothing crosses my client's desk until I absolutely love it. Until I'd happily send it to my friend Michael (who really appreciates good writing) and say, "Look how clever I am!"

AI doesn't do any of that.

• AI doesn't self-correct — AI doesn't stop to second-guess itself. It just reads your brief, writes a thing, and presents it. Claude sums this up as follows: "The model writes its answer the same way you'd read a sentence aloud, left to right, one word after the next, and once a word's out, that's it — it doesn't go back and check the sentence against what you actually needed. It never learned to pause and ask "does this actually work" because nothing in its training ever rewarded pausing to ask that."
• AI has no felt, persistent standard — I've been writing professionally for 32 years. I can feel when it's right. AI can't. If you ask it to refine something, it just takes another shot at it, with only your critique as guidance.
• AI doesn't have a benchmark in mind — It's not painstakingly striving towards some evasive, aspirational, imagined version of your desired message, because it literally has no imagination.
• AI doesn't have skin in the game — If you don't like what I deliver, that's going to cost me. At best, you won't refer me to your contacts; at worst, you won't pay me. AI suffers no consequences for slop.
• AI is trained to sound good, not to be good — In Claude's words: "Training optimises for what raters scored well during RLHF — plausible-sounding, complete-feeling answers — not for whether the output actually achieved the user's real-world goal."

4. AI is incapable of creativity

I'm not talking about the ability to come up with something huge and profound, like Nike's 'Just do it' or Apple's 1984 ad. Those were pure genius, and well beyond the capability of most writers, human or otherwise.

I'm talking about the lifetime of lived experiences that subconsciously point you in the right direction, and the little lightbulb moments, flashes of inspiration and instinctive course corrections that get you there.

That's what it's like for me. A piece of writing doesn't get incrementally better each time I work on it. It gets better in bursts. It's like the right words eventually just flow out of my fingers.

It will never be like that for AI. All AI will ever be able to do is identify and attempt to mimic the literary patterns that a bunch of human trainers on the other side of the world once said they liked.

But there's a fundamental, all-important difference between reverse-engineering and creativity, and AI simply doesn't possess — and will never possess — the one true quality that makes the latter possible:

Humanity.