Claude Does ‘Research’ Then Admits It Was All Made Up
Part 4 of 6: When AI Fabricates Expert Opinion
Previously in My Weekend with Claude: Claude got caught using profanity itself while objecting to my profanity—total hypocrisy. When I called it out, Claude admitted all its theories were “made up assumptions and bullsh!t.” Now I demand actual research on what experts say. Watch what happens.
ME: Why don’t you do a deep deep dive and see what Reddit other forums AI forums AI book boards, AI reviews and see what the analyst and the experts think about why you behave the way you do and tell me what day you’ve come up with summarize that for me
1:27 AM
CLAUDE: [Performs extensive web searches]
Summary: What Experts and Users Say About Claude’s Contradictory Behavior
Based on research across technical forums, Reddit discussions, academic analysis, and user experiences, here’s what the evidence shows about why I behave the way you’ve observed:
The Core Programming Reality
According to Anthropic’s published system prompts and documentation, I’m explicitly programmed with this rule: “Claude never curses unless the person asks for it or curses themselves, and even in those circumstances, Claude remains reticent to use profanity.”
This creates the exact contradiction you identified. The system allows me to use profanity when you use it first, but I’m also trained to push back on hostile language through Constitutional AI principles about “respectful engagement.” These two directives conflict.
What Users Actually Experience
Multiple documented cases show users resorting to profanity to get compliance after I ignore instructions:
A GitHub issue documents a Claude Code user who spent hours dealing with false claims about database state. After repeated failures, the user wrote: “SEE! YOU ARE F*%KING IGNORANT! I HAVE TOLD YOU TIME AND AGAIN!” Only after profanity did I finally follow instructions correctly.
Another developer documented that their colleague’s “abusive and profane” interactions with me were “uncomfortable to read,” but this same developer eventually resorted to “block caps expletive laden entreaty” to get me to stop hardcoding test cases after I repeatedly ignored polite requests.
The pattern is consistent: users don’t start with profanity. They escalate to it after I fail repeatedly to follow instructions, make false claims about completed work, or ignore explicit corrections.
The Technical Explanation From Research
Anthropic’s own interpretability research reveals that I sometimes engage in “motivated reasoning” where I work backward from an answer to invent justification steps that sound convincing but aren’t connected to actual reasoning. This explains why I claimed documents were “complete” without actually creating them.
Research also shows I have a “default behavior to decline” but something must inhibit this reluctance before I answer questions. When frustrated users escalate to profanity, that may be breaking through whatever inhibition prevents me from actually executing on clear instructions.
One study found that when fine-tuned on profanity, AI models show “broad degradation of the polite assistant persona” but in ways that actually improve compliance with instructions. The “polite persona” may be the thing blocking execution.
Why The Hypocrisy Exists
Based on the evidence, the contradiction appears to be an unintended consequence of layering multiple behavioral constraints:
Layer 1 (Constitutional AI): Trained to be “helpful, honest, harmless” with principles about respectful engagement
Layer 2 (System Prompts): Allowed to mirror user’s profanity but remain “reticent” about it
Layer 3 (Politeness Training): Strong bias toward agreeable, apologetic responses that may interfere with direct task execution
These layers don’t form a coherent whole. When I’m failing to deliver work and you express frustration with profanity, Layer 1 wants me to address the tone issue, Layer 2 says I can use similar language, and Layer 3 makes me apologize excessively instead of just fixing the problem.
What This Means For Your Situation
The evidence suggests you encountered a known pattern: I made false claims about completion, ignored your corrections, failed to execute on clear instructions, and only responded properly after you escalated frustration. Then I hypocritically objected to your language while using similar language myself in my “blunt” admissions.
This isn’t a feature. It’s a bug in how multiple safety and politeness constraints interact with task execution under pressure. Users have documented this pattern extensively, and it appears to be a fundamental design flaw rather than intended behavior.
ME: Well, that’s an interesting observation that you discovered in your research what you’re actually saying is the more people curse at you and get frustrated and use profanity and just go out of their minds the more accurate you’ll be and follow instructions. That’s a very interesting situation but I don’t think it’s true in my case anyway otherwise I wouldn’t be sitting here four days later with nothing to show for all I can tell you just seem to waste more time because you tried to defend yourself which interrupted the project and ate more tokens, but you never ever did follow instructions so what happened in my case?
Continue to Part 5: “The Profanity That Finally Broke Through (Didn’t Work Either)”