To work round these guidelines, the Humanizer ability tells Claude to switch inflated language with plain info and gives this instance transformation:
Earlier than: “The Statistical Institute of Catalonia was formally established in 1989, marking a pivotal second within the evolution of regional statistics in Spain.”
After: “The Statistical Institute of Catalonia was established in 1989 to gather and publish regional statistics.”
Claude will learn that and do its greatest as a pattern-matching machine to create an output that matches the context of the dialog or job at hand.
An instance of why AI writing detection fails
Even with such a assured algorithm crafted by Wikipedia editors, we’ve previously written about why AI writing detectors don’t work reliably: There’s nothing inherently distinctive about human writing that reliably differentiates it from LLM writing.
One cause is that despite the fact that most AI language fashions have a tendency towards sure forms of language, they can be prompted to keep away from them, as with the Humanizer ability. (Though generally it’s very troublesome, as OpenAI present in its yearslong struggle in opposition to the em sprint.)
Additionally, people can write in chatbot-like methods. For instance, this text doubtless accommodates some “AI-written traits” that set off AI detectors despite the fact that it was written by knowledgeable author—particularly if we use even a single em sprint—as a result of most LLMs picked up writing strategies from examples {of professional} writing scraped from the net.
Alongside these traces, the Wikipedia information has a caveat value noting: Whereas the record factors out some apparent tells of, say, unaltered ChatGPT utilization, it’s nonetheless composed of observations, not ironclad guidelines. A 2025 preprint cited on the web page discovered that heavy customers of enormous language fashions appropriately spot AI-generated articles about 90 % of the time. That sounds nice till you understand that 10 % are false positives, which is sufficient to probably throw out some high quality writing in pursuit of detecting AI slop.
Taking a step again, that in all probability means AI detection work may have to go deeper than flagging specific phrasing and delve (see what I did there?) extra into the substantive factual content material of the work itself.
