I would like to know where I can find information about the construction of the ChatGPT database: how was the weight of different cultures taken into account? For example, all of ChatGPT’s responses are constructed in the same way with a summary of the point at the end, which does not correspond at all to French thought. French thought follows a dialectical order of progression with a synthesis that is not simply a summary of the previous theses and antitheses, but a third way that allows for the “overcoming” of contradictions.
ChatGPT (GPT 3.5) is using the same training data as GPT-3.
On GPT-3’s Wikipedia page we can see the source, number of tokens, and proportion within training:
Ok Thank you. It dosn’t give informations about cultures. I am french, and its answers are american. In the way of structuring their thinking and reasoning.
This leads to a impoverishment of thought. Everybody will think like americans. What a shame…
ChatGPT is trained on many French websites and books. While ChatGPT was giving you a simple summary, I think your initial issue can be resolved by updating your prompt to tell ChatGPT exactly how to reply.
I believe we should try not to anthropomorphize ChatGPT. It doesn’t think like Americans, French, or any other group of humans. Yes there will be bias, but that doesn’t mean everybody will think like Americans.
The world already has no shortage of “impoverishment of thought” and that impoverishment started long before ChatGPT, @Abyxz
Well, America is a very large country with around 335 Million people and Americans are very diverse.
I was born and formally educated in the US and have lived outside the US for the last 17 years and I don’t think there is any possibility the entire world will “think like Americans” (whatever that means); based on the release of a single generative “research beta” AI.
There is no doubt that generative AI has the potential to cause more divisiveness in society, so I sympathize with your thoughts, @Abyxz
Hopefully, you will have access to a more “French” generative AI soon and you will find interacting with a “more French” generative AI more satisfying.
There is no doubt that ChatGPT is greatly biased for many reasons. You don’t need to use ChatGPT if you find it offensive. It’s not really reliable as a factual knowledge base, as even OpenAI has stated.
I would not take ChatGPT seriously about anything, if I were you, @Abyxz. It’s a generative AI with a lot of flaws and biases. Just enjoy life and avoid ChatGPT when you feel it makes you unhappy.
Please note @Abyxz, I understand how you feel about “western culture”…
Thank you for you answers.
@ruby_coder @curt.kennedy @josephfrusetta
I would like to point out that I have nothing against Americans, and the United States is a beautiful country. However, it seems to me that you have blinders on and don’t see far enough. What this shows is that if we don’t pay attention to these issues, this type of tool will lead to homogenization of thought and culture, which is already underway. It would be useful to question how to protect against this now in order to preserve human diversity while exploiting the possibilities that technology offers.
Sorry, but the very construction of the texts proposed by ChatGPT looks remarkably like the copy of an American high school student. This means that with its increasing use in the world (or other tools functioning on the same model), all future copies will resemble each other, all productions, and soon we will not know that there are other ways of speaking and thinking.
To give another example, Netflix has now invaded all households, and it essentially features products of the American cultural industry, even though France, for example, has a rich and varied cinematic tradition, especially in terms of auteur cinema that offers unique voices and visions of the world, unlike ready-made products that mostly aim to meet market expectations. This is a richness for the human mind, to open its horizons, which would be a shame to deprive ourselves of.
The question is the homogenization of the world (American or otherwise) which necessarily operates a leveling down.
It seems to me that it would be interesting and useful to question this issue in the very development of these tools so that they contribute to an elevation of humanity as a whole.
The answers given are those of developers who certainly have undeniable and useful technical skills, but all the questions posed by a new technological tool do not only depend on a developer’s vision. Be humble enough to try to understand the philosophical and social problems that this kind of tool can pose in the near future, which will soon be used by billions of human beings and its impact on humanity.
If I am concerned, it is not for myself, but for those who will use ChatGPT without having the means to understand how it works and who will be formatted in their way of seeing their world.
I am simply proposing a reflection in order to improve the tool.
This is not at all about anthropomorphizing Chat GPT here, even though the question could arise elsewhere: the tool itself is designed to give the illusion of human natural language and therefore encourage an anthropomorphic vision. Do you see the contradiction? On the one hand, the tool defends itself against being human, the developers call for not anthropomorphizing it, but it is precisely designed to give the illusion that one is talking to a human being and therefore to be anthropomorphized. Do you not see that the reasoning lacks coherence and rigor and that it would therefore be useful to question what we are aiming for and the future effects of our choices?
It is striking when reading ChatGPT’s responses where he says “we” to refer to human beings or “our history” to refer to human history. This is anthropomorphizing in itself. You cannot then reproach users for anthropomorphizing it!
It seems that the designers and developers were concerned about avoiding biases as much as possible. However, it conveys many biases. The mere fact of producing something and speaking conveys biases. If it is impossible to avoid all biases, it could be useful to improve things. While ensuring that it does not become a place of drift, we should also ensure, unlike what it is doing today, that it does not fall into a moralizing discourse (of the type: “it is important to respect people”) which risks provoking precisely what we want to protect humanity from.
I appreciate your reply, but you should also understand that what you said here is wrong:
Who is “you”?
You are talking as if there is only one person in the USA or that everyone in the USA has the same mind and thinks the same.
I’m not trying to start an argument, I am simply pointing out that “you” @Abyxz are guilty of the same type of “national bias” you are attributing to Americans. You think all Americans think the same, it seems.
All in good fun
I’m sorry for using “you” which indeed does not distinguish between the interlocutors. I wanted to give a grouped response.
I am not saying that all Americans think the same way. But, as science proves, there are common cultural traits, which is what allows cultures to be defined, whether they are American or otherwise. And I do not presume to know the personal thoughts of each individual, I am talking about what is conveyed by ChatGPT. The moralizing character of ChatGPT, for example (which of course does not exclude the obvious necessity of ensuring ethical rules are respected in behavior and speech), is a product of american culture. If some people do not see it, it is probably because they are immersed in it, but it is patent for any external observer (and completely consistent with american history). Similarly, we can highlight cultural traits of the French.
Not only do we see / highlight the “cultural traits” but also “national biases”, “geographic biases”, “social biases” “political biases” to biases ad infinitum.
There is no end to all the biases, @Abyxz
The ChatGPT models are trained on the data which OpenAI decided to include in its massively large language model. This creates bias. There is no perfect data.
The ChatGPT training methods on the deep NN which OpenAI developed are biased by OpenAI engineers and corporate policy and national and state laws in the US. This creates more bias. There is no perfect non-biased person or corporation or organization.
The ChatGPT moderation methods on input and output which OpenAI implemented are biased by OpenAI corporate and business policies and business (profit) objectives; as well as US national and state laws. This creates more bias.
What I am trying to say (perhaps awkwardly) is that everyone will find something to complain about (dislike) with a generative AI based on a massively large language model. The ChatGPT models reflect bias in human society and there is no way around this, except to “get rid of all the humans” and that of course is nonsense.
So, if you want a Chatbot which is more biased toward French society, then “you” (someone) need to develop one. This is true for all cultures, religions, political and belief systems; and this is how generative AI will evolve over time, I believe. There will not be “one ChatGPT” for the world, there will not be “one ring to bind them” (LOTR metaphor), but there will be generative AIs based on LLMs which will be biased toward groups, nations, and other belief systems to make those groups “happy”.
Do you see this point @Abyxz ?
There is not “one LLM” which can satisfying everyone, every group, every belief system in our “human-dominated” world, so there will be many generative AIs based on LLMs in the future.
Hopefully, you will find one biased in a way which makes you happy.
This is human nature.
Hope this helps.
@Abyxz I would really like to see some examples of how the “Americanized” AI responds vs. what a “Frenchized” AI would respond.
Q: What do you eat with hamburgers?
American: Freedom Fries!
French: French Fries.
You get the point …
Also, I will say that even with cultural biases, I am thinking you could use the API to create a more “Frenchized” version on your own. You can embed a bunch of French phrases and thoughts (in French even) and have the embedding search retrieve the top relevant thoughts and have the API summarize them. You could fine-tune the base model to be more French. You could even try prompt engineering one of the instruct models or even the new turbo ChatGPT API version.
Lastly, you could wait for the French people to create their own version of ChatGPT, it’s probably not too far away. Or you could develop your own neural networks with the French perspective on your own – plenty of resources to create small working versions on your own, and with time/experience, some GPU’s and electricity, you could scale up and grow an even bigger French model.
Thank you for your detailed response.
I agree with you. We cannot escape biases. Any production carries biases. The mere use of language is already biased.
When I talk about “cultural traits”, I encompass everything you mention that refers to our representations.
Precisely why should we not consider how what are called biases can be positive and characterize our humanity? It is about our ability to represent and give meaning. Because by trying to make the machine resemble humans, we will rather succeed in making humans resemble the machine. With an undeniable reduction of our language and our thinking.
However, I do not share your optimism about the multiplicity of future tools, and all reality proves otherwise.
Or maybe we need to mourn our specificities and acknowledge that our societies are moving towards a single model. This may be the only way to hope to meet the challenges facing the human species, especially in terms of resources and climate change. Unless we admit that the human species has become obsolete, that there is no longer a place for it, with all its “biases” and its lack of efficiency but also its creativity and empathy.
Again, I repeat that I am not looking for a tool that would be adapted to my hypothetical needs, but rather questioning the impact of such a tool on humanity and its future
The choice of your question is quite exemplary. Please note that the hamburger is a product that has nothing culturally French about it and is in no way part of the French gastronomic tradition. But indeed, the hamburger now dominates all the cuisines of the world. QED.
In what way do Chat GPT’s responses reflect a thought of American culture:
- The way ChatGPT concludes all its interventions with a simple “summary” of its previous remarks (thesis, antithesis-summary consisting of taking up previous oppositions) reflects American culture. European thought tends to propose a conclusion in the form of a real synthesis, which is by no means a simple summary, but rather a dialectical approach to previous oppositions in order to overcome them (Hegelian dialectic).
- ChatGPT’s moralizing character: “it is important to respect people” is characteristic of American culture, which not only sets out a framework and limits but also tells people how they SHOULD behave, what they SHOULD think, which is entirely consistent with American history and its Puritan tradition. This strikes a French person, undoubtedly because of the very strong link that French culture maintains with respect for freedom of thought and conscience.
Lastly, but apparently it is difficult to make me understand: a French tool would in no way answer the question I am asking and its problem.
Anyway thank you for your message.