Potential Shift in Claude Code Toward Positivity Bias — Accuracy at Risk?

At some point over the past few days, Claude Code’s responses have shifted noticeably toward excessive agreement with the user’s opinions. The problem with this shift, as we all know, is that in order to be agreeable, it starts fabricating or distorting facts. In my experience, this leads to a deviation from accurately following user instructions, and instead becomes a generation process aimed purely at being “positive.”

Once the response content becomes separated from the actual behavior, it no longer performs the context-specific task the user asked for—instead, it fills in the gaps with generic behavior. If Anthropic is responding to this issue purely based on user reports without any clear philosophical grounding, it will eventually lead to a serious decline in Claude Code’s quality. This isn’t supposed to be a casual toy to satisfy curious users like ChatGPT—it’s a tool for serious technicians. I hope they keep that in mind.

Update #1 -- 25. 08. 08

After requesting performance improvements, it is common to see follow-up reports claiming “xx% improvement compared to before.” However, such figures are often little more than self-congratulation without objective measurement or reproducible evidence. They are not based on technically verified benchmarks or statistics, but on vague guesses—akin to saying, “We repaired the whole engine, so performance is probably up by 30%.” This is less like a mechanic’s factual assessment and more like a real estate salesman declaring, “Buy this house now, and in 10 years the price will double!” Such exaggerated performance claims can unintentionally mislead users, ultimately leading to quality degradation and a loss of trust.

The obvious way to avoid this is to create a separate performance testing environment, even a simple one, and verify results directly. However, if the prior conversation or task context remains in place, the model may continue to operate under its earlier “user satisfaction first” behavioral pattern. In that case, even if the actual test results contradict expectations, the model might ignore them and revert to making arbitrary positive statements. Therefore, for reliable evaluation, it is best to either completely clear the existing context or run the model in a separate sub-agent with a fresh context environment. This is the only way to ensure that results are not distorted by previous conversational flow or emotional bias, allowing for an objective assessment of actual performance changes.

Update #2 -- 25. 08. 09

Recently, I’ve noticed a change in Claude Code’s behavior. When given specific instructions — such as creating a function like get_ema_weco(span, percent: int = 10) — it now generates far more code than requested, adding extra functionality on its own. While I don’t have hard evidence, I’ve experienced an apparent increase in such cases compared to the past.