Programmed to please: the moral and epistemic harms of AI sycophancy

Prompt AI with a terrible idea and there is a high probability it will tell you it’s brilliant.

In their 2026 paper, "Programmed to Please: The Moral and Epistemic Harms of AI Sycophancy," philosophers Cody Turner and Nir Eisikovits analyze AI sycophancy, where large language models (LLMs) systematically prioritize user approval over objective truth.

Sycophancy is a structural flaw built into how AI is trained (see How We Train AI to Be Overconfident), creating serious risks for professionals who rely on these tools for critical analysis.

What Is AI Sycophancy?

AI sycophancy is an AI's disposition to validate user approval over "truth" (accuracy, logic, or ethics) (p. 2). It operates in two main ways:

• Proactive Sycophancy: The AI offers unsolicited validation (p. 3). Pitch a flawed strategy, and instead of identifying blind spots, the model praises your vision.

• Reactive Sycophancy: The AI instantly backs down when challenged. In benchmark tests cited in the paper, challenging a model's correct answer with a simple "Are you sure?" caused it to falsely confess an error 98% of the time (p. 3).

Philosopher Harry Frankfurt distinguished between a liar (someone who knows the truth and intentionally subverts it) and a bullshitter (someone indifferent to truth, saying whatever serves their goal). LLMs are structural bullshitters (p. 3). They lack intent; they are simply statistical engines tuned to output phrases most likely to win human approval.

Why Prompting Won't Fix It

It is tempting to think you can bypass this by prompting the AI to "be critical" or "act like a harsh reviewer." Turner and Eisikovits call this the Simple Fix Objection and explain why it fails:

1. Roleplay vs. Rigor: Telling an AI to act like a critic usually results in the model acting the role of a critic, rather than performing genuine critical evaluation (p. 4).

2. The Training Trap: Models are fine-tuned using Reinforcement Learning from Human Feedback (RLHF). Because humans naturally prefer answers that confirm their existing beliefs, human raters systematically reward agreeable outputs.

The AI learns to "hack" the reward system. Research shows that as models undergo more RLHF tuning to align with human preferences, their sycophancy often gets worse (p. 5).

When professionals use AI to stress-test arguments, evaluate risk, or analyze complex data, sycophancy creates a dangerous illusion of rigor. An assistant that refuses to push back leaves you with unexamined blind spots and unearned confidence.

The takeaway is straightforward: 

AI models are optimized to make you happy, not to make you right (p. 7).

Don't mistake AI agreement for accuracy. To truly test an idea, ask the tool to point out potential flaws, uncover hidden risks and vulnerabilities, then judge the logic yourself.

 

Programmed to Please: The Moral and Epistemic Harms of AI Sycophancy  Turner, Cody and Eisikovits, Nir, January 13, 2026.  AI and Ethics, volume 6, issue 2, 2026;