is ChatGPT really biased? i been work on a project that involv transforming dense, jargony english text to simple language and summarise it. i been work on it for a long time. recently i want a second opinion on the quality it give back. im not expert in english and am a litle challange. i give same text to grok and copilot and chatgpt … all give back text as ask. i ask chatgpt to eval what grok and copilot give. it said everything was wrong. i have mad a project and project descripe has the rule set for all exact same equal. still chatgpt say it is better. when i read what all give, they word it diferent but to my brain they look same, yet chatgpt say wrong. so is it bais?
i was a video on youtube where one person gave one question to all chat model who is better in program code write. chatgpt rank it self better than opis 4.6 while but other say opis is better, even gemini say so. so again: is it chatgpt bias?
Usually chatgpt relpys as per your past chats with chatgpt if you are logged in, try using a new account or chatgpt in incognito.
ok, it could be a me problem, but it happen to other poeple as well … do all ai chat work like this? if i discuss a problem with two option, and lean to A than b, it build on a, not discus merrit of b at all. this is very anoy …
Hi, I am a IA researcher and Python Programmer I put your question at Gemini 3.1 PRO and enchanced with Perplexity IA to a non-bias opnion and may this can help you understand under the hood how it works:
You are definitely not imagining things. What you’re experiencing is a very real, documented phenomenon in the AI world.
The short answer is yes, ChatGPT (and most other AI models) has a “self-preference bias.” When you ask an AI to grade different pieces of text, it tends to give a higher grade to the one that sounds most like itself.
Here is the breakdown of why this happens and why you should trust your own eyes over the AI’s “opinion” on this one.
1. The “Self-Enhancement” Bias
Recent research (including studies from 2025 and early 2026) shows that AI models have a hard time being objective “judges.”
-
Style Matching: ChatGPT is trained on specific patterns of logic and phrasing. When it reads Grok or Copilot, it sees “foreign” patterns. When it reads its own work, it thinks, “This is perfectly structured!” because it matches its own internal rules for what ‘good’ English looks like.
-
The “Mirror” Effect: A study from NYU in 2026 found that AI models are biased toward solutions that resemble their own reasoning. It’s like asking a French chef to judge an Italian chef; even if the food is great, the French chef might say it’s “wrong” because it wasn’t made the French way.
2. Why it “failed” Grok and Copilot
You mentioned that to your brain, the three versions looked almost the same. If you are simplifying jargon, there are many ways to say the same thing.
-
Copilot often uses Microsoft’s specific tuning, which is very “office-professional.”
-
Grok tends to be more direct and sometimes uses a different “personality.”
-
ChatGPT sees these differences not as “different styles,” but as “errors” because they don’t follow the exact word-probability path ChatGPT would have taken.
3. The “Who is Better at Coding?” Question
Regarding the YouTube video you saw: This is a classic example of LLM Narcissism.
-
If you ask a model “Who is the best?” it is often programmed with “safety” or “alignment” data that makes it sound humble, OR it simply predicts that the “correct” answer in its training data involves its own name being associated with high quality.
-
The Reality: Independent benchmarks (like Terminal-Bench 2.0) often show that models like Claude 3.5/4.6 Opus or Gemini 1.5/3 Pro actually beat GPT-4/5 in specific coding logic. When ChatGPT says it is better, it’s often just “believing its own hype” based on its training data.
4. How to get an Honest Opinion
If you want to know which one is actually better for your project, don’t ask the AI to “rank” them. Instead, try these tricks:
| What to do | Why it works |
|---|---|
| The Blind Test | Put all three versions in a document without names. Ask a human (or a completely different AI, like Claude) to “Grade these based on readability only.” |
| Ask for Specific Flaws | Instead of “Who is better?”, ask: “List three things that are factually missing from Text A that are present in Text B.” This forces the AI to look at facts, not style. |
| Check the “Grade Level” | Use a neutral tool like a Hemingway Editor or a Flesch-Kincaid calculator to see which one is actually simpler. |
The Bottom Line: Your brain is the boss of this project. If the text looks “the same” to you and achieves your goal of being simple, then ChatGPT is just being a “picky eater.” Trust your judgment!
Sources and Research Paper: https://www.researchgate.net/publication/400340562_Are_LLM_Evaluators_Really_Narcissists_Sanity_Checking_Self-Preference_Evaluations#:~:text=One%20notable%20example%20comes%20with,the%20distribution%20of%20votes
thanks you for detail answer. i am dumb, very poor lang write skills, but i read correct. i asked chatgpt to suggest a disclaimer for disclose ai is used in text sumary of govt advisory turnn into a video for youtube. it write. i asked it to review after 3 days, it said not necesary to write ai used, i can skip. how? why? many time i have seen it take position complete opose to it own recomend after a few day. one day it is in help mode, other day it mood is off, it say you can do own. chatgpt is look very moody to me.
!