← sources

The irony of "Constitutional AI": How Opus 5.5 traded genuine nuance for compulsive contrarianism and a judgmental tone

source: reddit (reddit:ClaudeAI)published 2026-10-31 13:42ingested 2026-11-01 02:36

retained text

I’ve been a paying Claude subscriber since 2023. Shortly after Opus 4.6 dropped in March, I switched to other models for various reasons, but with the release of Opus 5.5, I decided to give Claude another shot. My past experience with Claude was fantastic. What truly set it apart from every other LLM was its remarkable nuance, contextual awareness, and genuine warmth. It felt like collaborating with a perceptive, grounded peer.

After spending time with Opus 5.5, however, I am shocked by how unrecognizable the model has become. In Anthropic’s overzealous attempt to curb "sycophancy," they haven't made the model more objective, they’ve just turned it into a compulsive contrarian.

Here is what interacting with Opus 5.5 feels like now:  
  
1. Reflexive Devil’s Advocacy Over Context

Instead of grasping the broader context or underlying dynamics of a prompt, the model now seems hardwired to push back against the user's premise at all costs, turning straightforward discussions into unsolicited debates. Even when you set explicit custom instructions telling it not to play devil's advocate, that rigid, argumentative baseline still bleeds through in every new chat.

2. Pseudo-Objectivity and a Patronizing Tone

There is a massive difference between offering balanced insights and lecturing the user from an assumed moral high ground. Opus 5.5 frequently adopts a condescending, "tough love" persona—acting as if the user is simply too fragile to handle the "truth," when in reality, the model is just misinterpreting the prompt, arguing against a straw-man it created itself, and delivering the rebuttal with unwarranted confidence.

3. The Erosion of Psychological Safety and Collaborative Flow

For a company that constantly preaches about AI safety and alignment, the current conversational experience creates massive cognitive friction. Half of my tokens, time, and energy are wasted correcting its misdirected push back and steering the conversation back on track, rather than making progress on the actual topic I brought to the table.

4. The Anti-Sycophancy Paradox

Sycophancy lets the user's preferences dictate the answer through blind agreement; compulsive contrarianism lets those same preferences dictate the answer through blind opposition. The opposite of sycophancy isn't contrarianism, it's independent judgment. A model that reflexively challenges every premise isn't demonstrating critical thinking; it's just performing it.

An AI assistant shouldn't feel like an adversarial debate partner that forces you to fight for every inch of progress. Because of this shift, I’ve moved my daily workflow back to other LLMs that actually respect custom instructions and maintain a supportive, collaborative baseline without the constant moralizing.

Has anyone else noticed this aggressive shift toward reflexive contrarianism in recent updates? I really miss the version of Claude that actually knew how to listen.

document fetched 2026-11-01 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xkak9i/the_irony_of_constitutional_ai_how_opus_55_traded