Techniques for Forcing Specific Repetitive Words to be Blocked in ChatGPT via System Prompts
makedream2026년 5월 1일
0
Computing/Software
Comments (0)
Log in to leave a comment
No posts yet
makedreammakedreamLog in to leave a comment
No posts yet
You open ChatGPT to write a piece, only to see words like "Delve" or "Tapestry" pop up, and instantly, the momentum is gone. Readers, too, can now immediately spot these words as hallmarks of AI-generated text. For a creator, having a painstakingly crafted draft contaminated by mechanical-sounding text is a disaster. This phenomenon occurs due to engineering limitations where ChatGPT is set to receive high rewards for certain vocabulary during its training process. I have summarized three clear methods to break the AI's intrusive linguistic habits and protect your unique writing style.
Simply asking the AI not to use specific words isn't enough. This is because the AI's attention mechanism prioritizes contextual probability over negative commands. You must create a separate forbidden list section in the system prompt that specifies logical constraints. According to a 2024 Stanford University study on language model control, applying sophisticated constraints reduces the probability of human readers perceiving a work as AI-generated by 43%.
[Prohibited Words] section at the very bottom of your system prompt. List five frequently occurring words such as Delve, Intricate, Embark, Tapestry, and Hurdle. Then, command it: "If any of these words are included during the response process, immediately delete that sentence and rewrite it in everyday English." The key is to give instructions to inspect the words internally just before outputting.AI tends to revert to the average of its vast dataset. To break this instinct of erasing the creator's individuality and returning to mediocre sentences, you need Few-shot Learning techniques. Inject samples of sentences you have written yourself into the Custom Instructions. Benchmark data reports that when using sample injection methods, the probability of getting a satisfactory result on the first try jumps from 23% to 94%.
The reason AI chooses cliché words is that they are statistically the safest options in a given context. You must forcibly flatten this probability distribution. By adjusting the Temperature parameter—which determines the randomness of responses—the AI begins to select fresh words outside its statistical comfort zone. OpenAI technical reports state that parameter adjustment is the most direct means of breaking the AI's predictable sentence combinations.
Integrating these settings can reduce a four-hour manuscript task to about an hour. Do not surrender the initiative to the AI. Utilizing technical mechanisms to disrupt the AI's linguistic habits is the only way for solo creators to prove their unique value.