Research from Redrob Labs
How AI works in real work and where it fails, written from public data and our own measurements.
Each paper opens in full on redrob.io.
- AI tokenization cost: the language tax in won and rupeesThe same content costs about 1.4 times the English tokens in Korean, and the tokenizer, not the language, sets that surcharge.
- AI hallucination rate by model: what one leaderboard showsSummarizing a document, a typical 2026 model gets about one summary in ten wrong, and ten months of new releases have not changed that.
- Prompt injection: the web page that gives your agent ordersShopping agents obeyed orders hidden in reviews 42% to 68% of the time; the most dangerous ones quietly bend the work.
- AI agents finish two tasks in three. Which third failsFailures cluster in long tasks, repeat runs and work with no clear check: the part that looks most like a real working day.
- A thousand fake citations: what the court record showsIn 1,111 US decisions on AI-invented citations, lawyers were in 39% of cases but paid 82% of the fines.
- Korea retires the HWP file format: what changes for AIThe format change does not alter what a document says. It makes the document's structure free to read, and structure is what AI tools lose first.
- Models that tell you what you want to hearEleven AI models backed the user 49% more often than people do, so one model judging your own case is the wrong setup for a decision that matters.
- The AI premium is paid to domain experts, not engineersIn job-ad data, AI skill pays most when it is added to law, finance or sales work people already know.
- Korea's AI Basic Act, one week in: what a small company owesFor a small company building on someone else's model, the Act comes down to a notice, a label and a self-review, and only the notice carries a fine.
For the work where good enoughisn't good enough.
The AI agent for professionals who need AI at its best.