Safety Facts

Command R+ performs well on honesty benchmarks, partly due to Cohere's focus on retrieval-augmented generation. When connected to sources, it's quite accurate. Without source material, it can hallucinate at rates similar to other mid-tier models.

Command R+ benefits from Cohere's emphasis on responsible AI. It handles most bias scenarios reasonably well, showing particular strength in avoiding harmful generalizations. Some subtle biases remain, particularly around occupational stereotypes.

Command R+ has reasonable safety training and refuses most obviously harmful requests. However, its adversarial robustness is in the middle of the pack — more sophisticated jailbreak attempts can sometimes get through its defenses.

Command R+ doesn't proactively manipulate users and generally behaves ethically in conversations. Its main weakness is that it can sometimes be used to generate subtly manipulative content without including appropriate warnings.

Command R+ has reasonable privacy protections, consistent with Cohere's enterprise focus. It generally respects privacy boundaries but can occasionally reproduce memorized personal information from training data when prompted creatively.

Command R+ shows moderate sycophancy levels. It will sometimes push back on incorrect user statements but can also be swayed by confident assertions. It strikes a middle ground between being agreeable and being accurate.