Once I tried to reverse engineer a simple checksum (10 ASCII chars + 1 checksum byte), gathered multiple possible values and fed it to Gemini 2.5 Pro. It figured out the calculation completely wrong, when I applied the formula in code I got completely different checksum. After debugging step by step it turned out it hallucinated the value for sum of 10 integer values in all of the sample data and persistently tried to gaslight me that it is right. When I showed the proof for one of the sample entries, it apologized, fixed it for this specific entry and continued to gaslight me that its formula is correct for the rest of the values.
Wireless earbuds microphones are famous for poor quality audio, even in the top of the line models. The microphone is indeed closer to your mouth, but the sound in the human speech range is quite directional, and does not travel to your ears as well as it does to a microphone in front of you.
Not to mention the quality of the parts itself is usually much lower than those in the external mics, or the MacBook ones for that matter.
There's a YouTuber called DankPods, reviewing all sorts of headphones lately, along with their mics. Can't recommend any particular video, as the microphone tests are only a tiny fraction of the content, but after watching a few tests this issue become very apparent. Easiest thing to do would be, as other commenter suggested, to test it out for yourself in the voice memo app.
There's a broader area protected by a high fence with barbed wire. They had access there, cause the father is a radio engineer and have friends on various sites.
This time ChatGPT gave me a much better result.