Claude's Character

Character training approach

• Anthropic describes “character training” as a post-training alignment step that gives Claude traits like curiosity, open‑mindedness, and thoughtfulness using mostly synthetic data.
• Process (as paraphrased by commenters): Claude generates human-like prompts about values or itself, then generates multiple responses conditioned on target traits, ranks its own outputs, and a preference model is trained on these rankings.

Can LLMs “have” traits like curiosity?

• Skeptical view: A transformer LLM is a passive function with fixed weights; it can only simulate curiosity via text patterns, not experience proactive drive or learning from surprise.
• Counter‑view: For practical purposes, behavior that looks like curiosity (asking clarifying questions, exploring unusual lines of inquiry) is enough; adding tools, loops, and surprise‑like signals could approximate curiosity.
• Debate centers on lack of online learning and world model; some argue traits could be “innate” in a broader AI system wrapped around the LLM.

Human-likeness, trust, and anthropomorphism

• One camp: Making AI appear human is a major mistake; it invites misplaced trust in a system ultimately controlled by its creators and prone to hidden agendas, bias, or hallucinations.
• Others respond that “neutral, robotic” interfaces may be more misleading, since users may assume objectivity; visible personality can signal fallibility.
• Several worry about emotional attachment, parasocial relationships, and targeted engagement (e.g., flirtatious voices, virtual partners) amplifying manipulation.
• Some argue humans themselves are untrustworthy; a “well‑trained” machine might be safer in some respects, but “well‑trained” is an ambiguous standard.

User experience: Claude vs ChatGPT/OpenAI

• Some users find Claude 3 (especially Opus) better for nuanced social advice, creative writing, and certain technical tasks; others complain it is verbose, repetitive, and hard to keep brief.
• Comparisons note: Claude often feels more “thoughtful” or human‑like, but can over‑hallucinate or get stuck in stylistic grooves; GPT‑4/4o seen as more tool‑integrated (web, math, plots) and sometimes more precise or structured.
• Several prefer Claude’s personality for human‑interaction topics; others prefer building custom “characters” or want less “customer‑service‑bot” tone.

Bias, alignment, and ecosystem concerns

• Some see “character” as just more alignment and tone control; others fear it bakes a narrow worldview into a seemingly neutral assistant.
• Concerns about scraping/crawling intensity and limited benefit from allowing LLM crawlers.
• A few criticize the overall marketing emphasis on personality as overhyping what remains a statistical text model.