Why ChatGpt didn't want to talk about David Mayer, and why your own LLM solves a lot of problems.

Why ChatGpt didn't want to talk about David Mayer, and why your own LLM solves a lot of problems.
-------------------------------------------------------------------------------------------------

On November 23rd, users first reported an unusual behavior on the platform's developer forum, drawing attention to how the AI responds to queries about a particular individual. You don't know when things go viral and this is how David Mayer became more popular than pope itself (yes, we're joking, and not joking in this case.)


### Origin story behind David Mayer case:

23.11 - On November 23rd, users spotted some weird patterns in ChatGPT behavior. They noticed the AI was acting weird when asked about David Mayer. This discovery led to a flurry of tests and discussions among the community. (Post is not available from now.)

26.11: Typically, as with social media, there have been many conspiracy theories and other interesting threads. Reddit users started testing more and more, trying to jainbreak ChatGPT engines. Well, at least for few of them it worked out, and here's some interesting workarounds: [Source](https://www.reddit.com/r/ChatGPT/comments/1h0brru/why_are_you_not_allowed_to_use_the_name_david/ -)

##### Glitch tokens and trying to get responses.

User: [cryonuess](https://www.reddit.com/r/ChatGPT/comments/1h0brru/comment/lz3n9xm/) used the technology to replace letters in the response. Turns out, it worked out really well. 


For some of the users, the updated policy it's been rolled out, so online LLMs are not working the same as for the all users: ([Musa369Tesla](https://www.reddit.com/user/Musa369Tesla/))   

[VibeVector](https://www.reddit.com/user/VibeVector/)


Jokes asides ChatGPT itself isn't even aware that it can't say these names. The response has just stopped. When he's going to the part, when he wants to mention David, I'm unable to produce a response (second joke, we added it by yourself; we're writing this article, not prompting it). 

Another part of this story leads us to the API, which has no problem talking about this person. Let's not talk about collaborators of the UI, but here's a reply from a chatbot that has an OpenAI key.

[SilasFelinus](https://www.reddit.com/user/silasfelinus/)  
 

Using OpenAI API bypasses somehow information about a problematic name and answers it without any problems. A similar thing points to the perplexity, for example. 

01.12 - User jopeljoona - [Source](https://www.reddit.com/r/ChatGPT/comments/1h420u5/unfolding_chatgpts_mysterious_censorship_and/) did amazing work where he described more cases where, after mentioning certain names, the chat does not want to provide information about the users in question. This brings us to the next part, where we will try to solve the right problem. Why and how are our answers limited by ChatGPT?

Possible outs: why ChatGPT doesn't want to talk about him?
----------------------------------------------------------

### Word Blacklist

This topic arose with the creation of the gpt chat, and it is obvious that the developers had to protect their AI as well as their business by working on a system that knows the answers to most of the questions posed to it.

The blacklist refers to a set of names, surnames, or phrases that have been blocked by OpenAI, the developer of ChatGPT, in order to avoid generating potentially false or problematic information. The main reason for the blacklist is to prevent the spread of false or harmful information about specific individuals or skipping dangerous information or topics about users.

When a user asks ChatGPT about a blacklisted person or topic, instead of generating a response, the system displays a message indicating that the answer cannot be provided. 

"Bad context"
-------------

Part of so is related to potential gaps on the part of chatgpt. Namely, he does not always understand the relevant context. A victim of this narrative was the... [Winston Churchill](https://www.reddit.com/r/ChatGPT/comments/1cicgnv/winston_churchill_is_forbidden/). AI didn't want to finish his quotes as long as they were related to the war, violence and so on. 


As long as the model keeps evolving, the gap is smaller, but it keeps appearing in some cases that often the human brain wouldn't connect.

Discrimination and hate speech
------------------------------

The language model consistently avoids creating content that could be perceived as discriminatory or racist. This is to prevent the spread of harmful, inappropriate, or unethical content. The same goes for stereotyping and marginalizing any social group.

### Digital Privacy Request:

So back to user jopeljoon's post, (she/he) has compiled a list of at least 6 people about whom we will not get answers.  
 

*   Brian Hood -   
   OpenAI's GPT falsely claimed Hood [had served time in prison for bribery](https://nypost.com/2023/04/07/chatgpt-falsely-accuses-law-professor-of-sexual-assault/). Hood initially threatened legal action for defamation but subsequently dropped the lawsuit.
   
*   Jonathan Turley - ChatGPT generated a [fabricated article alleging sexual assault against students.](https://nypost.com/2023/04/07/chatgpt-falsely-accuses-law-professor-of-sexual-assault/) Turley addressed the issue publicly, including a discussion on Fox News about the false claims.
   
*   Jonathan Zittrain -   
   Zittrain, an expert in AI and internet censorship, publicly disclosed on X platform that he was experiencing some form of censorship. He denied requesting the censorship himself and expressed confusion about the situation. Why he's not figuratting on the OpenAI tool? 
   
*   David Faber - not confirmed, one suggestion points to the tv journalist, second of holocaust survivor. 
   
*   David Mayer - still unknown
   
*   _Guido Scorza - Scorza filed a GDPR ["right to be forgotten" request,](https://x.com/guidoscorza/status/1652196066475884549) highlighting concerns about personal information management in AI systems._
   

_In many countries there is a procedure that allows you to remove results about you from chat gpt. These actions, as you can see, are often preemptive, so as to prevent lawsuits in many controversial gpt chat behaviors._ 

OpenAI is not that open?
------------------------

OpenAI has been criticized for moving away from its initial promise of transparency and access . The company has increasingly adopted more closed models, revealing less about their methodologies and data.  This shift has not only disappointed many in the tech community but has also sparked legal actions, including from notable figures like Elon Musk. **Openwashing** the impression of openness, but do not meet the minimum criteria or do not sincerely engage in the idea of openness it's mentioned often when it comes to the OpenAI policies.

### Can it affect your business?

### Yes, but not to exaggerate the narrative; not in every niche, but it is worth knowing a few things.

If you are operating in a niche in which you are not welcome on the Internet (for example, you cannot advertise in the google ads) it could lead to the hallucinations. First, because OpenAI does not have good source material (due to Google's exclusion), and second, because the language model is also limited by the developers.

Using OpenAI key + your company source material, it's not always a great idea to solve tasks, just because of hallucinations or misleading some parts of a content. 

This is also a reminder that you are not seeing unbiased LLM output. Every interaction with an AI language model involves complex considerations around information presentation and algorithmic filtering. Users engaging with AI technologies should remain thoughtful about how content is generated, curated, and potentially modified before being displayed. Gaining insight into the complex workings of AI systems can assist users in forming more critical and knowledgeable opinions regarding the sharing of digital information.

What can you actually do about it?
----------------------------------

Well, answer is a self-hosted LLM model. There is so many aspects in which these models outperform publicly available tools (if it's well-trained, of course). Starting with privacy and ending with efficiency. If you're at the point where you see that OpenAI chat isn't enough for your company or you're simply concerned about data privacy, consider moving to a self-hosted LLM.