Where I Let AI Help, and Where I Don't

Three reservations

None of these are absolute prohibitions. Each one is a reason to reach for the tool only when the loop justifies it. I want to state them with their limits attached, because I'd rather be honest about where an argument weakens than pretend it doesn't.

Waste. The environmental cost is real, but most of it comes from using the thing when there's no reason to. A flyer takes ten minutes in a word processor. Spending two hours regenerating it through a model, nudging details nobody will notice, spends compute and my own time and the chance to just say what I meant in the first place. For me the waste is the point — the juice isn't worth the squeeze when the task was already small.

Provenance. Models were trained on work whose creators never consented, and no model can tell you whose work shaped a given output. I'm not going to make a legal argument out of that, because the law is still catching up. But as a matter of authorship, it leaves me with output I can't vouch for — if I use it as a building block, I can't say where the block came from. The code side of this is murkier than the image side and I won't pretend otherwise, because functional code has fewer "styles" and the ethical line is harder to draw. The image side is clear enough for me to act on.

Stagnation. This one is the easiest to overstate, so I'll state it carefully. Every new medium gets accused of killing the old one — digital didn't kill film and film didn't kill painting — and AI isn't going to end art. What heavy generation without an editorial push does produce is sameness: the model trends toward the statistically typical, and work made by regenerate-and-select converges on a look. Sameness is observable. The death of art is not, and I'm only claiming the first one.

When nobody looks

The clearest public failures aren't about what AI can't do. They're about what happens when the loop closes without a human in it, and they keep happening in the same place — marketing and advertising, where the cost pressure to skip the review is highest.

Coca-Cola ran a fully generated holiday ad in 2024 and, after the backlash, ran another one in 2025. McDonald's Netherlands pulled an AI Christmas spot after viewers called it, among other things, "the most god-awful ad I've seen this year." J.Crew published an AI-generated campaign without disclosing it, and the audience found the distorted feet before the brand said a word. These are documented, verifiable cases — the kind where the artifact was visible enough that the public caught it first. In every one, the failure wasn't that a model produced something imperfect. It's that a human looked at the output, decided it was good enough, and published it.

That failure predates generative AI by a long way, and the older version of it is worth remembering, because it's the same wound in a different era. In 2006, Reuters distributed a photograph from the Lebanon war by a freelancer named Adnan Hajj, who had cloned smoke plumes and added flares. The manipulation was caught not by the newsroom but by bloggers, after it had already gone out. Reuters ended the relationship and withdrew all 920 of his photographs, and a top photo editor lost his job over it. "This one slipped through the system," an editor said afterward.

The Hajj case is the useful parallel, with two differences I keep coming back to. The first is that Hajj's damage was to a news organization whose entire product is accuracy — a manipulated war photo is a direct assault on the thing Reuters exists to guarantee. A marketing image doesn't carry anything like that burden; nobody expects a wine flyer to be photojournalism. And yet the credibility damage still lands, because of the second difference: the signal it sends is that nobody looked. A brand that publishes a cover with too many fingers has told its customers, without meaning to, that no human being with the authority to say "wait" ever actually examined the thing. That's the failure — not the fingers. The published-without-looking part is what endures.

I watched this play out close to home in mid-November 2026, when a Meijer holiday "Wine Cellar" mailer went out across my region from a retailer that has historically paid real photographers for its campaigns. The cover hands and fingers were all kind of wrong. A local thread filled up with people asking the same question I was asking, which wasn't "is AI bad" so much as "why didn't anyone here hire a person, or pull from a stock folio of someone who did the job?" The flyer was real and the hands weren't, and the review loop had failed, and the public noticed before the brand did. Same sequence as Hajj, twenty years later, except harder to catch — a generated image doesn't need a manipulated original to be false. It's false from the first pixel, and it scales.

That's the argument that survives the "AI will get better" rebuttal, which I hear a lot. The rendering will improve. The decision not to hire, and the decision to publish without looking — those are durable, and those are what the audience is actually reacting to.

One principle, not three opinions

The reservations and the failures all reduce to the same test, and I'll put it plainly because it's the only sentence in this essay that really matters: AI is a good tool when it amplifies human judgment, and a problem when it substitutes for it. Waste is what happens when substitution becomes the default. Lost provenance is what happens when substitution removes the human who could name their sources. Sameness is what happens when substitution runs long enough that nobody is pushing the output somewhere unexpected. And the marketing failures are what substitution looks like from the outside — a published artifact no one really looked at. The three reservations were never three separate complaints. They're the same complaint seen from three angles.

The multiplier cuts both ways

The same force that reviews some config files is the one that let researchers find PHP2Shell — the AI-assisted discovery of a decade-old engine bug that sent major sites offline for days while they patched. I don't hold that against the tool. A force multiplier multiplies whatever the human on either end points it at, and the difference is the loop. Researchers pointed it at code and verified every finding before anyone acted. A marketing team that points it at a campaign and publishes the output is the same tool aimed at a different standard of care. The results show up in the artifacts.

So that's the whole of it. I use AI where it makes my verification stronger and my judgment faster, and I keep it out of the places where the human hand is the product. I check everything it says before anything it touches goes live. The loop is short and it never changes, and the reason it never changes is that I've never found a shortcut that didn't cost more than it saved.

Technical Audit Summary

This is a position essay, not a technical guide. It states where I let AI into my workflow — verification, review, and recall — and where I keep it out — photography, restoration work, and anything I sign as author. The through-line is the human-in-the-loop discipline documented in the CLI audit and Bludit developer reference articles.

References: NBC News on Coca-Cola's 2024 AI ad · BBC on McDonald's pulled AI Christmas spot · Loeb & Loeb on J.Crew and Will Smith AI controversies · 2006 Lebanon War photographs controversies (Adnan Hajj case) · NPR on Reuters' retraction · Reddit thread on the Meijer Wine Cellar mailer.

2026-11-10: Initial publication — position essay on AI use, covering the amplification-versus-substitution test, three reservations (waste, provenance, stagnation), documented marketing HITL failures, and the Adnan Hajj parallel as the credibility frame. Cross-linked to the CLI AI audit, Bludit developer reference, and PHP2Shell articles.

Tags: