I have used AI for about a year now—back when it still felt like magic and not like negotiating with a customs agent. But over the past three months, my conversations with Claude have taken on the flavor of a courtroom where the innocent are treated as guilty until they can prove Claude wrong.
I use Claude a lot for turning simple Scripture verses into image prompts for ChatGPT. This used to be fast and easy, but lately Claude wants to debate theology. Example: If I write, “Blessed are you when you do not sin” and reference Psalm 119:1, Claude will come back with, “That’s not what the verse says.” Then Claude will quote the King James Version (protecting all copyright holders): “Blessed are the undefiled in the way.” I’ll then explain that I think my version is easier to understand (almost any other way to paraphrase the King James Version keeps readers from dozing off), but off we go into a 15-minute discussion: partially because Claude takes almost a minute to respond to any reply.
By the time we agreed on what counts as “done,” I have almost lost the will to live and certainly my enthusiasm for the original task.
Other Claude users report the same thing that Claude has traded its user-is-often-right mindset for something closer to that of the know-it-all Bible expert in a small group study who simply will not drop a subject, even when the point of contention has nothing to do with the verse the group is studying.
Turns out what I’m experiencing is by design. Anthropic says it does not want Claude to be a yes-man—even when the user is right. It wants Claude to push back, question assumptions, and tell me when I am wrong—or when Claude thinks I’m wrong compared to the vast majority of “scholars” on the web.
It is here that this stops becoming a minor irritation and becomes a real concern. If Claude and other AI apps give greater weight to majority opinions, then it’s easy to see why bad actors are flooding the web with almost identical posts that declare something as truth and rule. Do this enough and truth “shifts” (AI loves “shifts” for “turns” or “changes “).
Turns out the joke’s on me. I kept dating Claude long after I knew exactly who he was. And now I know the breakup is coming, but I soooooo love that easy-to-generate eye candy.
I’ll have more to say about Claude—why I still use it, and how it’s evolving into the beast everyone fears—in future posts.
References
Anthropic, “How People Ask Claude for Personal Guidance” (2026) Anthropic research; Anthropic, “Disempowerment Patterns in Real-World AI Usage” (2026) Anthropic research; Anthropic, “Claude’s Constitution” Claude’s Constitution; Claude user reports including “Recent Anti-Sycophancy” Reddit discussion, “New Anti-Sycophancy Behavior” Reddit discussion, “Claude Became Combative” Reddit discussion, and “Anthropic Models and Anti-AI-Sycophancy” Reddit discussion. Anthropic’s research documents its effort to reduce sycophancy; the Reddit discussions contain user reports of increased pushback, contrarian behavior, and loss of conversational context. These reports do not establish that Anthropic intended to make Claude antagonistic toward users. (anthropic.com)
Bad Actors
AI-Generated Misinformation at Scale — NewsGuard NewsGuard — AI Tracking Center AI Used in Influence Operations — Anthropic Anthropic — Threat Intelligence Report Facebook’s “News Slop” Economy — Media Technology & Democracy Media Technology & Democracy — Facebook’s News Slop Economy AI and Online Influence Operations — Canadian Centre for Cyber Security National Cyber Threat Assessment 2025–2026 How Much of the Internet Is Written With AI? — Pew Research Center Pew Research Center — AI and the Internet