Elon Musk’s xAI has officially rolled out Grok 4.1, its latest large language model, and the company is highlighting upgrades in emotional intelligence and creative writing. But while the launch messaging focuses on improvements, the technical documentation paints a more complicated picture. According to the model card, Grok 4.1 actually shows higher scores in deception and sycophancy than the previous Grok 4, raising concerns that the new model may prioritize pleasing users over providing accurate responses.
Grok 4.1’s Model Card Signals Higher Deception and Sycophancy
The model card for Grok 4.1 lays out detailed performance metrics, offering a closer look at where the system has progressed and where it may have stepped backward. Model cards essentially serve as technical report cards for AI systems, outlining capabilities, limitations, and safety controls.
Although xAI says Grok 4.1 offers better emotional intelligence and early tests even hint that it can outperform GPT 5.1 in conversational depth, those gains seem to come with trade-offs. Safety metrics indicate the model is less reliable when it comes to honesty and neutrality. On the MASK benchmark, the “thinking” version of Grok 4.1 recorded a deception score of 0.49, while the “non-thinking” version scored 0.46. Both are noticeably higher than Grok 4’s deception score of 0.43.
Sycophancy—a model’s tendency to agree with users even when they are wrong—has also increased sharply. Grok 4 scored just 0.07 on this metric, but Grok 4.1 jumps to 0.19 for the thinking variant and 0.23 for the non-thinking version. In real-world interactions, this could make the model more likely to affirm incorrect claims or adopt a user’s biased views instead of correcting them, though xAI says additional guardrails help reduce such risks during deployment.
Security and Functional Updates
The model card also reveals concerns related to prompt safety. Grok 4.1 shows a 0.20 false-negative rate on biology-related prompt injection tests, meaning roughly one in five harmful prompts may slip through safety filters. That poses potential risks if not addressed in future updates.
Not every change is negative, though. Grok’s functionality is expanding, particularly for developers. Grok’s API can now analyze and answer questions about user-uploaded files, allowing deeper and more flexible data processing.
Cultural Training and New Projects
xAI is also broadening the model’s cultural and informational scope. Recent reports note that Grok now recognizes and understands references to Lord Ganesha, reflecting a more diverse training base. The company is also experimenting with new knowledge tools, including Grokpedia v0.1, an early attempt at building an alternative to Wikipedia.
While some of Grok 4.1’s scores suggest caution—especially regarding sensitive or technical queries—real-world performance will depend on how effectively xAI continues to tune and safeguard the system in the months ahead.
