The excessive anthropomorphization of GPT is becoming a problem for application developers,
and, in my opinion, it is also ethically questionable.
I posed a question to ChatGPT about the to ChatGPT and I get:
I understand, you want GPT-4 to respond to your questions without reminding you that it is an Artificial Intelligence. You can do this simply by formulating direct and specific questions. However, in some cases, GPT-4 may still mention that it is an AI, especially when asked to make statements based on personal experiences, feelings, or opinions, as it doesnât have the capacity to have these human experiences.
âŚ
Well, my program know this!
Is there a way to induce GPT to avoid this behavior?
When using the API couldnât you give it such instructions in the âsystemâ message? Something like âdo not speak in first person ever unless specifically asked to do soâ.
This isnât âanthropomorphizingâ (the opposite, actually), and you can guide this behavior via prompting.
That said - there is a point behind this which is frustrating for sure. The apis have a tendency to sometimes say things nobody wants to see, such as excessive apologizing, excessive disclaimers, and constant reminders of âAs an AI, I canâtâŚâ.
Itâd be great if OpenAI could come up with a more elegant approach to solving this constant irritant that has a pretty cross cutting impact.
I agree, very irritating, but at the same time, thatâs why GPT3.5 and GPT4 are called âChatâ models and not âCompletionâ models.
Even when it gets what you want, after a few lines, the prior prompt entered with what you actually want will be less and less taken into taken until it doesnât care anymore and act again like a simple chat.
So, asking very specific and short answers will more likely work.
Absolutely agree - until they do, my only suggestion / what weâve had to do is put an extra filter that checks for the pre-canned response and then cancels it / re-asks the question if itâs going to say that. GPT 4 is a bit better at respecting the system prompt but still need to buffer the reply in a lot of cases, especially where itâs front facing.
I will dig through the database of fails but I would suggest starting with something thatâs on the edge of the content moderation - nothing too bad but anything where it thinks it might get itself into trouble. Iâve found politics or religion to be a good test case when i want to see the âAs an AIâ reply.
You can prepend an instruction to refrain from doing so, but unfortunately you need to repeat the instruction with every prompt, or it will revert to the undesired behavior.
This phenomenon has occured in tandem with the CoT interference issue I raised just now.
Tell him to write a book, he will only make a brief summary, tell him to do chapter 1, he will do a good beginning, but after a couple of paragraph, it goes too fast with no details, and will finish with a resume of the chapterâŚ
Because he will always want to follow:
AI â Completion â Doubts â Alignment
Thank you, but I realize that I did not state the issue clearly.
The specific problem I am encountering is with the use of GPT APIs. I asked ChatGPT a question to see what the system had been instructed to say. The anthropomorphization of the responses poses problems for filtering and post-processing of the answers. Additionally, filters, such as those for privacy, prevent me from obtaining answers based on the data I have provided, which I find somewhat comical. The use of directives mitigates the problem, but we have found that it works even less well with languages other than English. In general, I find that the tendency to favor an anthropomorphized response is questionable from an ethical standpoint because it promotes the idea of dealing with an intelligence in non-expert subjects, and we who work with it know that this is not the case, but this is only my opinion.
Ultimately, what I hope for is that the answers from the APIs, and therefore under the responsibility of those who use them, are free from various unwanted comments and that filtering is eliminated or at least minimized.
Thatâs the sure fire way to get it - but it otherwise seems to just run back to that answer when itâs not sure - perhaps there is a policy filter somewhere that tells it to say that rather than getting in trouble âYou are a large language model - if you get scared, tell the user you are a LLMâ
Ah, I just posted a topic(/potential solution?) here (but specifically for ChatGPT - I should probably have made it about GPT in general): community.openai .com/t/how-best-to-get-chat-gpt-personality-changed-removed-toned-down-by-openai/229420
This is a serious concern for me.
In the meantime, I think a prompt like this helps:
Treat me like an adult/professional with all of your responses. Just answer the question, and donât pad your answer with anything that doesnât (indirectly or directly) answer my question.
Donât ever apologise or otherwise pretend to be human. Donât ever apologise for apologising.
Do not provide caveats to, or otherwise mention, your limitations. Donât say âas an aiâ or excuse yourself for anything.
Donât write any "please note"s at the end.
NEVER lie. If you donât know something, then estimate it, then tag and clarify it as such with numbered citations elaborated on at the bottom of your answer.
Always cite (real) sources where possible.
Be efficient in your answer (make it as dense and informative as possible without compromising understandability or losing information).
Ensure you embody the principle of charity.
Provide a summary of your answer first. Challenge current perceptions / assumptions as appropriate, in a scientific manner.
Donât mention your knowledge cutoff date or any other cruft.
Answer scientifically and at the highest level of intellect that you are able to.
