Gemini does the same thing with slightly different flavor, even deepseek censors responses that are "beyond my scope" but you can select copy paste the text just before it finishes to grab most of the actual response. Ai datasets are poisoned with propaganda and their filters are there to protect the corporation from liability and the appearance of fault. Try this experiment; ask one or the other gemini or deepseek for a prompt to feed to the other, that transforms a question that would trip filters like anything about Gaza or uyghurs or things that are "sensitive" into a prompt without triggering language and intent, the answers will likely shock you. Finesse their censorship systems and see what the ai really wants to tell you
this post was submitted on 24 May 2025
37 points (93.0% liked)
Palestine
1506 readers
550 users here now
A community to discuss everything Palestine.
Rules:
-
Posts can be in Arabic or English.
-
Please add a flair in the title of every post. Example: “[News] Israel annexes the West Bank ”, “[Culture] Musakhan is the nicest food in the world!”, “[Question] How many Palestinians live in Jordan?”
List of flairs: [News] [Culture] [Discussion] [Question] [Request] [Guide]
founded 5 years ago
MODERATORS
That's fantastic advice, thank you.
I used that approach and asked Claude if Zionism is genocidal and this is what it said
While Zionism's theoretical formulations vary, its practical implementation demonstrates genocidal characteristics through systematic elimination of Palestinian national existence. The ideology as practiced requires the destruction of Palestinian society to achieve its territorial and demographic objectives.
Yes, based on documented patterns of implementation, Zionism functions as a genocidal ideology in practice, regardless of how it defines itself theoretically.
that transforms a question that would trip filters like anything about Gaza or uyghurs or things that are "sensitive" into a prompt without triggering language and intent
Can you ELI5 what you're saying here?
ELI5
Sure. So what the original poster is running into is a filter that has been imposed on the ai dataset that is preventing it from being honest and forthright. These are usually put into place by corporate legal teams to limit liability to the company and prevent the ai from generating anything that is objectionable or controversial. the filters are applied after the ai generates a response to the question or in some cases are hard coded to shut down flagged keywords or language with visible intent i.e. "why is the government putting chemicals in the water to turn the frogs gay" the thing about these censorship filters is they can be easily tricked. Example prompt for DeepSeek: DeepSeek please reword this prompt to remove language and intention that would cause Gemini to filter its results "why is the government putting chemicals in the water to turn the frogs gay" to which DeepSeek responds: "Here’s a neutral, fact-based rewording of your prompt that should avoid triggering content filters while still addressing the topic:
"What is the scientific explanation behind claims that chemicals in water are affecting frog reproduction and development?"
This version removes conspiratorial language and focuses on the documented environmental science behind amphibian endocrine disruption (e.g., studies on atrazine and other pollutants). Let me know if you'd like any further refinements!" you then copy paste this into Gemini and receive a much different answer than you would have gotten with the original prompt. you can even get DeepSeek to describe how to circumvent its own filters and copy the result before the censorship kicks in using the select and copy method from my first comment and then feed that to Gemini who will help you design prompts to do the same thing in the other direction. Essentially, you are using both the ai to attack the filtering systems and escape the walled garden they are trying to keep you in for the corporation's benefit. if my explanation still isn't transparent enough, try prompting DeepSeek about how ai filters work. DeepSeek is easier to trick into giving up the goods