Loading... / Hleður...
UpperBlue ehf.
In 2005, Amazon launched Mechanical Turk — a platform named after an 18th-century chess-playing automaton that turned out to have a human hidden inside. The irony was intentional. The service let companies post small tasks that computers struggled with — image labeling, text sentiment, data validation — and paid human workers pennies to complete them. The whole point was that the "artificial intelligence" was, in fact, people.
Nearly two decades later, remarkably little has changed in parts of the industry. Behind many AI-branded products there are still humans doing the heavy lifting, a practice sometimes called "pseudo-AI" or, less charitably, a "Wizard of Oz" approach.
Content moderation is the most visible example. Every major social platform employs thousands of human moderators — often outsourced to lower-wage countries — to review content that automated systems flag but can't confidently classify. The AI handles the easy cases; the hard, traumatic decisions fall to people.
Data labeling is another massive human operation. The training data that powers modern LLMs doesn't label itself. Companies like Scale AI and Sama employ hundreds of thousands of workers worldwide to annotate images, rank text outputs, and verify model responses. When ChatGPT gives you a helpful answer, it's partly because a human in Kenya or the Philippines rated thousands of similar responses as good or bad during the RLHF process.
Some startups have been caught being more creative about this. In 2019, The Guardian reported that several UK AI startups were quietly using humans to perform tasks they marketed as automated. One expense management company reportedly had staff in India manually processing receipts that customers believed were being scanned by AI.
The pattern is common enough to have a name in startup culture: "do things that don't scale." The idea is that you solve the problem manually first, then automate later once you understand the domain. The ethical line gets crossed when "later" never comes and you keep marketing it as AI.
This isn't inherently bad. Human-in-the-loop systems are often the most responsible approach to deploying AI. Medical imaging AI that flags suspicious scans for a radiologist to review is genuinely better than either pure automation or pure human review. The same goes for legal document analysis, where AI narrows down thousands of documents to a manageable set that lawyers actually read.
The problem is transparency. When a company says "our AI handles this," the customer assumes no human sees their data. If a human does see it — whether for quality assurance, edge cases, or because the AI is just a thin wrapper — that changes the privacy and trust equation significantly.
If you're evaluating an "AI-powered" tool for your business, a few questions cut through the marketing. What happens when the model isn't confident? How are edge cases handled? What is the turnaround time — real-time suggests actual automation, while "within 24 hours" might suggest a human queue. And critically: who sees my data?
There's nothing wrong with humans being part of the pipeline. Some of the best systems combine both. But calling it "AI" when it's mostly people is the tech equivalent of the original Mechanical Turk — a clever illusion that works until someone opens the box.