FYI: all closed-source AI products degrade performance based on the prompt w/o telling the user. There are layers of "filters" on the output.
Most obvious is a "recitation filter"... checks every LLM output to prevent an AI product from accidentally outputting exact copies of information from its training data.
The problem: exact-quoting high quality sources is often the most accurate response to give... but it can be a liability for closed-source products. So the filter catches it... and a less-exact output is produced instead.
Open source models don't do this.
This is also why attribution-based control is such a crucial feature to long-term trustworthy AI. Attribution (and corresponding credit/payment layers) would allow frontier AI systems to more reliably exact-quote sources. Important research area.
NEW: Anthropic is walking back Claude Fable 5's policy to covertly degrade performance for competing AI researchers, after facing fierce backlash.
“We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” Anthropic tells WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”
