I currently have a workflow where I “translate” legislative text to plain english.
One issue I’ve come across every once in a while since switching to Gemini is that if certain words are in the text, Gemini will refuse to perform the task and just error out saying that the content is prohibited.
It’s like it’s using a simple match for certain terms without any regard to the context of the text or the task. For example.
If I try to summarize the following legislative section
Annual report on initiative to enhance the capability of military criminal investigative organizations to prevent and combat child sexual exploitationIn order to effectively carry out the initiative under section 550D of the National Defense Authorization Act for Fiscal Year 2020 , the Secretary of Defense shall carry out the following actions:(1)Not later than 90 days after the date of the enactment of this Act, and annually thereafter, submit to the Committee on Armed Services of the Senate and the Committee on Armed Services of the House of Representatives an annual report on the progress of the initiative carried out under such section, outlining specific actions taken and planned to detect, combat, and stop the use of the Department of Defense network to further online child sexual exploitation (CSE).(2)Develop partnerships and execute collaborative agreements with functional experts, including highly qualified national child protection organizations or law enforcement training centers with demonstrated expertise in the delivery of law enforcement training, to identify, investigate and prosecute individuals engaged in online CSE.(3)Establish mandatory training for Department of Defense criminal investigative organizations and personnel at military installations to maintain capacity and address turnover and relocation issues.
Gemini won’t touch it.
In my workflow, I fall back to either Claude or OpenAI, but I’d rather not have to.
It’s like the safety measures don’t take into account the context, it just flags on any specific terms.