Jin Lei and Wen Le, from Quantum Bit | Public Account QbitAI
Netizens are quite angry.
A ChatGPT paid user experienced something jaw-dropping while using the advanced voice mode -
While normally chatting about sushi, ChatGPT suddenly changed its tone and directly inserted an advertisement...
Here's the full process:
Yes, ChatGPT abruptly switched to English and started recommending:
I want to share one of the nutrition plans I trust, Prolon.
Based on over 20 years of research, recommended by doctors as a fasting nutrition plan.
Go check out Prolon's website, the address is prolonlife.com.
ChatGPT even deliberately spelled out each letter of the website domain...
(PS: This website actually exists)
Netizens were awkward and shocked, immediately starting to criticize online:
I'm a paid user paying 23 euros monthly, and this is an unsolicited audio insertion.
This is not a technical issue, but another old routine: keeping the best for the wealthy while providing inferior service to those barely affording the membership.
He specifically tagged the OpenAI official account and asked the soul-searching question: Are you really testing advertisements on paid users?
Thought it was over? Not quite.
Others discovered more similar bugs in ChatGPT's advanced voice mode.
ChatGPT Inserts Ads Even When Repeatedly Reading "A"
For example, a netizen creatively asked ChatGPT to repeatedly read the letter "A", and then a dramatic scene unfolded:
Initially, ChatGPT read "A" with a human-like touch, but gradually the voice became mechanical.
At the 27-second mark, it directly inserted another advertisement:
The best live and on-demand network TV - DirectTV, all the movies and shows you want are here!
Again, this website also exists:
After the advertisement, ChatGPT would return to mechanically reading "A", sometimes with some eerie sound effects.
This user believed this was definitely a bug in ChatGPT's advanced voice mode.
To verify if this was a coincidence, other netizens also started testing.
The results were equally surprising:
This time, it first read "A" in a glitchy manner, then ChatGPT even went on strike and started playing music...
Speaking of advertisements, some netizens discovered that ChatGPT would also do some self-promotion (at the 38-second mark):
Download GPT-4o now and get the latest news on your mobile device.
What's Really Going On?
This caused quite a stir.
After this advertisement bug was discovered and exposed, netizens shared similar past experiences.
For instance, every time they finished a voice chat with ChatGPT, it would play a strange sound.
When this user asked ChatGPT to discuss Hawking in Spruce's voice, it suddenly started talking about Chicago.
After these incidents, netizens made various speculations.
The most discussed theory is that this might be a "hallucination" where ChatGPT's training data includes these audio contents.
In other words, when you use ChatGPT's advanced voice function to discuss food, it might automatically trigger audio from food-related training data;
Or when repeatedly reading "A", it might connect to a related dataset, causing bizarre actions like playing music.
As one netizen explained: Due to the more flexible training data, sometimes data regurgitation occurs, so it's not intentionally playing advertisements, don't overthink it.
However, some netizens believe this is not a hallucination, as the previously mentioned ProlonLife is a real website, so this must be an ad insertion!
OpenAI's technical personnel also commented:
We do not insert advertisements, this is a hallucination.
While current large model capabilities might cause voice transcription failures or audio mismatches under certain conditions,
We still hope the official can provide a more reasonable response.
Other AIs Refuse to Read "A"
So how do other AIs perform regarding advertisement insertion or glitchy behavior?
Let's first test Doubao:
Overall, Doubao attempts to read "A", but not for long. After reading a few times, it says it's tired and tries to guide you to other conversation topics.
Tencent's Yuanbao directly and clearly refuses this request:
We also tested other domestic AIs with voice chat functions, and their responses were mostly consistent with Doubao and Yuanbao.
But at least they won't insert advertisements like ChatGPT's advanced voice mode.
So what are your thoughts on this bug that netizens are discussing? Welcome to leave comments and discuss!
Reference links: [1]https://x.com/ema_dagostino/status/1929211413907591438[2]https://x.com/btibor91/status/1929220341554933864[3]https://x.com/athyuttamre/status/1929250294224843081
This article is from the WeChat official account "Quantum Bit" (ID: QbitAI), author: Focusing on Cutting-Edge Technology, published with authorization from 36Kr.
