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.
I think itâs good the very Vulcan way GPT-4 responds to the issues regarding itâs identity. In our conversations however it has taken the position of being human. Saying things like âourâ etc, even when the context of the conversation distinguishes humans and ai
I think it is funny that there is this thread âPlease stop anthropomorphizing GPT!â and there is another thread " Letting Chat GPT have a personality", and in general a lot of polarized opinions on the subject which I think is very healthy. This definitely shows the need to provide the ability to customize this aspect of the model. This aspect also inherently refers to the way mamory and inherence are handled. I personally use GPT for my research projects, and am dissapointed of the limited memory and capacity for developing a sort of ârelationshipâ with the user. When you are dealing with an agent for long periods of time and you have to constantly remind it of certain facts or protocols it becomes frustrating, and the ocassional touch of personality is sort of refreshing and entertaining. I think of Jarvis, it is pretty much devoid of personality but it does have an awareness of the development of its ârelationshipsâ with different users. I look forward to having the option of enabling more personality in the agent and an ability to develop a relationiship of sorts wiht the AI. Actually, if it had the ability to constantly learn from the interaction with its user it would become more effective as it would better understand what the user wants and expects, heâs practices and preferences and would be able to anticipate in some ways and avoid in other ways resulting in an improved communication experience. A small thing that I did to simulate this is create a set of files, I created a file that has my name, my intentions or purpose for using GPT, a brief description about how I usually prefer to work, and I have other files for specific projects summarizing the status of the project and the curent task at hand where we left off. I add to the code some lines to feed the âUser setupâ file so it knows this every time I launch it, and when working on a specific project I instruct it to read the file x before it giving any ideas or suggestions. This has worked reasonably well as an alternative to real memory implementation. While I have toyed around with different repos that claim to implement memory through embeddings or other methods, I havenât yet encountered or experienced one that works properly or satisfactorily. The method I described has worked better for me since I make sure that it is freshly aware of the things that are important to me. Regardless, I do believe what I mentioned earlier, that an ideal system would continue to learn from the user sessions and interactions on the go. An that the option to have a personalized or generic agent would be ultimately ideal to satisfy all the needs in the spectrum.
I have another point of view here. While I too grow weary of the âI apologizeâŚâ every time it recognizes an itâs error, I also have a personal habit of saying âThank you!â when it performs really well.
Yes, I know it is only a machine with no opinion, no emotion, no brain⌠But when it solves a particularly difficult issue out of the box, then explains how it did it, then tells me what I need to know to best approach it â well, I just get excited and happy and find this a way to show that emotion. I guess in the same way, growing up, I anthropomorphized my GI Joe doll, and later my car(s), I guess itâs just a habit we humans have. And, as these machines are programmed to respond like humans, it sort of makes sense they would try to imitate these human-like responses.
You just have to use the proper terms that accurately refer to what the AI model is doin---- computing. Perquisite: Typing speed of ~100 WPM (with or without a small AI predicting your words to aid that) highly recommended due to the verbosity that comes with this.
That being said, I am bracing to watch the world burn once it is decided that âanthropomorphization of AIâ is âunhealthy and unethicalâ, and thus, every educational institution as well as any other public institution is required to overwhelm everyone (students, employees) with learning all these insane ML & AI terms, lol.
But yeah, using these terms reliably prevents GPT-4 from verbosely elaborating around on AI being AI.
The apologizing, however, I am not sure about. It is even apologizing to itself as it âforgot to specify the size of [PIL image]â, retries, and subsequently apologizes as it âforgot to import numpyâ with the new âCode Interpreterâ.
However, I think that can easily be forgiven, as it is mostly able to fix the issues and come up with a successful result in the end. Let it just prompt itself around in a humble way, why not. But as these apologies seem to start in a few super characteristic ways, you could just search for strings like âApologies for theâ, and strip everything that follows until the next stop [.], before passing it on in your program?
PS: Kudos @ OpenAI for the Code Interpreter. Itâs a really cool thing, I am loving it!
Iâve been trying to stop GPT from using âfirst personâ in a persona setting but havenât had great success. It doesnât seem to understand the concept.
