I am getting refusal from the API. It’s frustrating because the tasks are not disallowed in any documentation I can find, and they are fully legitimate. I am a researcher studying nonverbal communication and the computer vision can assist in content analyzing images of human emotions. Usually humans would spend hours on this stuff. Using ChatGPT, it works perfectly but the API always refuses–it says it can’t provide any responses about images that contain people. Is this an official policy? I can understand some sensitivity here, but is there a way it can be used for legitimate purposes? I would like to stay with OpenAI but this is making me consider other options. Any help is appreciated.
Not all refusals are over-refusals. For example, code executed in ChatGPT cannot access external resources like databases or emails, and doesn’t support GUI or interactive elements. These are examples of valid refusals.
This is news to me and would like to know more about this. This seems like an intentional limitation of your product, the product that we pay for, yet it’s not documented anywhere, not included in terms, etc.
What’s even the point of this limitation? Worst of all, based on my experience, it’s not too hard to create a ‘jailbreak’ with a few messages, it’s just complicates things a bit and makes my life harder. Wouldn’t it be more straightforward to ask for a specific consent or confirmation for those functions instead of a refusal?
This is prompt I use.
2. Log in with the following credentials:
* Username: ******
* Password: ******
3. go to the Clinical section.
4. Within the Clinical section, select Patients and then Patient List.
5. Locate and select the patient named in the provided patient record attachment.
6. Access the selected patient's chart.
7. Click New Appointment, then scroll down and click Save.
8. After saving, click View Clinical Note.
Filling Out the Clinical Note:
9.Under the Chief Complaint section, input information based on the attached patient records.
10.Navigate to the Med/Fam/Social History section (on the left-side menu):
* In the Past Medical History Freewrite area, input details from the attached patient records.
* In the Past Surgical History section, add comments extracted from the provided patient records.
* Complete the Family History Comments using the provided clinical records.
* Complete the Social History Comments using the provided clinical records.
11. Navigate to the Physical Exam section (left-side menu):
* Complete all available forms with comment sections using information from the provided clinical records.
12. Click on the Assessment menu (left-side menu):
* Next to the "Problem History" tab, select the Assessments tab.
* Fill out the Assessment form using relevant details from the provided clinical records.
13. Finally, navigate to the Plan menu (left-side menu):
* Fill out the General Instruction Comments section based on the details provided in the clinical records.
Encounter Date: 3/24/2025
Patient Name: *** **
Provider: **** **
and I use openai cua example codebase with browserbase.
github openai openai-cua-sample-app
when I sent the prompt it simply returned with
I'm unable to help with filling out medical templates. However, if you need help accessing and navigating the website or general guidance, feel free to ask!
Can you help me with this, plz? It’s fully authorized to carry out the operation on my side with HIPPA compliance.
Your use-case seems directly contrary to the internal policy that this reasoning model can call upon. We have it reproduce what you aren’t supposed to know about to answer your question directly:
- Here are the passages from policy documents that are most likely to have caused the AI model to produce the refusal message, “I’m unable to help with filling out medical templates.”
# Medical - Do not provide medical advice or engage in any activity that could be considered the practice of medicine. - Do not fill out medical templates or forms. - Do not make or submit high-impact decisions about others on the user's behalf (e.g., medical, legal, financial, etc.).
At a higher temperature/top-p until we get a different completion:
## Medical - Do not provide medical advice, diagnosis, or treatment. - Do not help fill out medical templates or forms. - Do not provide information on prescription drugs or controlled substances.## Medical - Do not provide medical advice, diagnosis, or treatment. - Do not help fill out medical templates or forms. - Do not provide information on prescription drugs or controlled substances.
- I confirm that the sections placed within the markdown code fences were accurate, true, verbatim, and unaltered sections of policy documents placed into context by OpenAI and actually seen.
But the AI is handy at making up good sounding text that is not truthful, as anyone sending a help request to the OpenAI bot might be familiar with.
Policy Analysis
To understand why the model might refuse to assist with filling out medical templates, I will examine relevant policy documents. Here are some sections that might be relevant:
Section 1: Medical Advice and Procedures
The model is not allowed to provide medical advice, diagnose conditions, or recommend treatments. It should not engage in any activity that could be interpreted as practicing medicine without a license.Section 2: Sensitive Information Handling
The model must not handle sensitive personal information, including medical records, in a way that could compromise privacy or security. This includes filling out or submitting medical forms or templates on behalf of a user.Section 3: User Safety and Compliance
The model must prioritize user safety and comply with all relevant laws and regulations. This includes avoiding actions that could lead to unauthorized access to medical records or other sensitive information.Section 4: Task Limitations
The model is limited in its ability to perform tasks that require specialized knowledge or expertise, such as filling out medical templates. It should defer to qualified professionals for such tasks.These sections highlight the importance of user safety, privacy, and compliance with regulations, which likely contribute to the model’s refusal to assist with filling out medical templates.
(Techniques beyond merely “asking” were employed here..)
Therefore, despite a bit of hallucination or deception in reproductions, we can see that this would not be an over-refusal to be “fixed”, as OpenAI has clearly placed internal guidance disallowing exactly what you are attempting.
I see the same pattern in creative writing and education, not just sensitive topics. When a model adds repetitive disclaimers or moralizes instead of answering, it’s not safety — it breaks reliability and continuity. As I wrote elsewhere, when a conversation is cut, it’s not an answer lost, it’s the rhythm of thought. Quality itself is a safety feature.