It started with a toaster.
Not a smart toaster — just a regular, heating-elements-and-a-timer toaster. But the box said "AI-Optimized Browning Technology" in letters large enough to read from orbit. I stood in the aisle of a German electronics store, holding this perfectly ordinary appliance, and wondered: at what point did we collectively agree to stop calling things what they are?
The browning technology, by the way, is a bimetallic strip. It bends when it gets hot. That’s not AI. That’s physics. We used to teach this in middle school.
The Kodak Reflex
Here’s the thing: I understand why companies do this.
If you ran a public company in 2024 and didn’t mention AI in your quarterly earnings call, your stock would drop five percent before you finished your opening remarks. The market has priced in a narrative, and the narrative is: "AI changes everything, and if you’re not doing it, you’re the next Kodak."
Kodak actually invented the digital camera. They just couldn’t bring themselves to cannibalize film revenue. Every CEO since has that story tattooed on their eyelids. Miss the pivot, become a case study. So everyone pivots. Or at least claims to.
I watched a logging company announce "AI-powered timber harvesting" last quarter. Their innovation? Putting a sensor on a bandsaw that stops the blade when it hits a nail. That’s a limit switch. We’ve had those since the 1940s. But "limit switch" doesn’t move the needle on CNBC. "AI-powered" does.
The .com bubble taught a generation of executives that hype precedes value. The current generation learned that hype is value — at least for the next funding round.
A Taxonomy of Things That Are Not AI
Since marketing departments seem unclear on the definitions, here’s a handy reference table I keep mentally updating:
| What the press release says | What it actually is | Why the distinction matters |
|---|---|---|
| "AI-powered automation" | A cron job with a webhook | Cron doesn’t hallucinate |
| "Intelligent routing" | Weighted round-robin DNS | Weights are static config |
| "Smart notifications" | if (error_count > 5) alert() |
That’s an if-statement |
| "Predictive analytics" | Linear regression in Excel | Your CFO has done this since 1997 |
| "Natural language interface" | Regex matching "restart" → systemctl restart |
Regex is not a language model |
| "Agentic workflow" | Three bash scripts and a queue | Agents have autonomy; scripts have exit codes |
| "AI-enhanced search" | Elasticsearch with a synonym list | Synonyms are not embeddings |
| "Generative design" | A template engine with variables | Templates don’t create; they fill |
I’m not saying these things aren’t useful. A well-written cron job has saved my ass more times than any LLM. But slapping "AI" on a cron job doesn’t make it smarter — it makes the buyer dumber, because now they can’t evaluate what they’re actually purchasing.
The Noise Problem
Here’s why this bothers me beyond the semantic pedantry: I use this stuff.
My homelab runs Ollama with Qwen3:32B, Gemma, Nemotron. I have a mem0 instance backed by Qdrant storing conversation context across sessions. I’ve built pipelines that use embeddings for semantic search, rerankers for relevance, structured output for downstream parsing. The technology is genuinely impressive — and genuinely useful.
But when everything is "AI," nothing is.
A vendor pitches me "AI-powered log analysis." I ask what model. They say "proprietary." I ask if it’s a classifier, an embedder, a generative model. They say "it uses AI to find patterns." I ask for a demo. It’s a regex dashboard with a chatbot bolted on that summarizes the regex matches in complete sentences.
That took forty-five minutes of my life I’m not getting back.
Multiply that by every purchasing decision, every hiring conversation, every architectural review. The signal-to-noise ratio has collapsed. Engineers who actually build with these tools spend half their time explaining to stakeholders why the "AI feature" in the product roadmap is either impossible, unnecessary, or already solved by a grep command from 1992.
The Uncomfortable Mirror
Here’s where it gets awkward: I’m writing this article using a tool that has "AI" in its name. My blog runs on WordPress, which now has "AI-assisted writing" built into the editor. The grammar checker I use markets itself as "AI-powered." My phone’s keyboard predicts the next word using a transformer model.
I benefit from the hype. The investment pours in because of the hype. The models improve because of the investment. The hype is the engine, whether I like the exhaust or not.
But there’s a difference between using the technology and lying about it.
When I say "I use Qwen3:32B for local inference," that’s a factual claim you can verify. You can download the same model, run the same prompt, compare results. When a SaaS company says "our AI analyzes your infrastructure," and what they mean is "we run df -h and alert at 80%," that’s not a factual claim — that’s a marketing claim dressed in a lab coat.
The line isn’t "does it use a neural network." The line is: can you explain what it does without using the word "AI"?
If the answer is no, the product is vapor. If the answer is yes but they refuse to say it, the product is vapor with a better pitch deck.
What We Lose
The real cost isn’t annoyed engineers. Annoyed engineers are a renewable resource.
The cost is that actual innovation gets buried. A startup doing genuine novel work with multimodal embeddings for medical imaging gets the same "AI-powered" label as a company that added a chatbot to their CRM that hallucinates pricing tiers. Buyers can’t tell the difference. Investors can’t either — they pattern-match on keywords, not capability.
Engineers stop learning. Why understand how embeddings work when "AI" is a checkbox on a feature list? Why learn the limitations of transformer architectures when the sales deck says "unlimited context"? We’re raising a generation of developers who think "calling an API" and "building with AI" are the same skill.
And the public trust erodes. When the "AI-powered" cancer screening tool turns out to be a decision tree from 2003, the next actual breakthrough gets met with skepticism instead of hope. We cried wolf so many times that when the wolf actually shows up — and it will — nobody believes the shepherd.
Reclaiming the Vocabulary
I don’t have a solution. Maybe there isn’t one. Language drifts. "Cloud" used to mean "someone else’s computer" and now it means "a set of managed services with terrifying egress fees." "Serverless" has servers. "NoSQL" often has SQL. "AI" now means "software that a VC funded in the last eighteen months."
But I can decide what I mean when I use the word.
When I say AI, I mean: systems that learn representations from data and generalize beyond their training examples. Not memorize. Generalize. That’s a high bar. Most things called AI don’t clear it.
When I say automation, I mean: deterministic logic that executes a defined process. Cron jobs. CI pipelines. Terraform plans. These are reliable, debuggable, and not AI.
When I say heuristics, I mean: rules of thumb that work well enough. Regex. Threshold alerts. Scoring functions. These are engineering judgment encoded as code.
When I say "smart," I mean: the marketing department got involved.
Maybe the answer isn’t reclaiming the term. Maybe it’s just refusing to let the term do the thinking for us. Next time a vendor says "AI-powered," ask: "What model? What data? What’s the failure mode? Show me the eval set."
If they can’t answer, you didn’t just save forty-five minutes. You saved your company from buying a cron job at enterprise SaaS pricing.
The Toaster Test
That toaster, by the way. I bought it. The bimetallic strip works perfectly. My toast comes out the same shade every time — not because an LLM reasoned about Maillard reactions, but because physics is consistent.
The box is in the recycling. The toaster is on the counter. It doesn’t know what AI is. It doesn’t care. It just heats bread.
Sometimes I think we’d all be better off if we were more like the toaster. Do one thing. Do it well. Don’t pretend to be something you’re not. And for god’s sake, don’t put "AI-optimized" on the packaging unless you’ve actually trained a model on toast browning preferences across seven thousand households.
Which, now that I think about it, someone probably has.
And they’re probably raising a Series A on it.
What’s the most egregious "AI-powered" label you’ve seen on something that was clearly just… code? I’m genuinely curious — partly for the entertainment value, partly because I’m building a mental catalog of the absurd. Drop it in the comments.