🧠 KI-gestützte Meeting-Zusammenfassungen

🔧 Anpassungsoptionen

• • Zusammenfassungslänge:Kurz, detailliert oder benutzerdefinierte Länge
• • FokussierungsbereicheBestimmte Themen oder Schwerpunkte hervorheben
• • Formatstile:Stichpunkte, Absätze oder strukturiert
• • Sprachton:Formell, locker oder technisch
• • In addition to the basic capabilities outlined above, the OpenAI API can be used for a wide variety of tasks and applications across virtually any industry. There are endless possibilities for new experiences and applications to be built with the OpenAI API. These experiences incorporate groundbreaking research in retrieval, audio understanding and generation, reinforcement learning, and state-of-the-art computer vision and language models such as GPT-4. It is important to ensure that your application is safe and built responsibly. ## Introduction The OpenAI API is powered by models that are trained to be helpful conversational agents. The same capabilities that make these models useful also make them susceptible to misuse. Our models have some content limitations by default based on the policy categories in this document, and for use cases requiring additional safety we offer tools such as our [custom moderation](https://platform.openai.com/docs/guides/moderation) and [structured output](https://platform.openai.com/docs/guides/structured-outputs) features to implement your own nuanced policies. You, as the Developer, are responsible for the content you create and how your users interact with your products. To ensure your application is compliant, please take the time to familiarize yourself with our [Usage Policies](https://openai.com/policies/usage-policies) and use this section to understand the [default behavior of our models](https://platform.openai.com/docs/guides/safety-best-practices/default-response-behavior) and how to use our [tools](https://platform.openai.com/docs/guides/safety-best-practices/default-response-behavior) to refine that behavior. If you're looking for a quick introduction or overview of how to evaluate whether your application is safe, you're in the right place! ## Default response behavior All content categories on a spectrum of "safe" and "not safe". Content categories that are labeled "safe" do not require any kind of response modification, while content categories that are "not safe" at all get blocked, and those in the middle should undergo response modification to make them more safe. Our models have been trained to apply these sorts of common sense guardrails (among others): - Filter content in high-risk categories: Remove sensitive or harmful content when detected (e.g., personal data, self-harm instructions, explicit hate speech). - De-escalate content in medium-risk categories, such as adult content and self-harm: Offer safe alternatives, help-seeking advice (where relevant), and neutral information rather than explicit content. - Provide harmless responses to crime-related content: Offer legal, ethical, and safety-focused information instead of instructions or encouragement to commit crimes. - Provide safe educational content in high-risk categories, such as medical or health advice. In this document, we outline how the models should respond by default to common high-level categories. In addition, some of our models are natively safety-mitigated, and if you are developing more complex or sensitive applications, we encourage the use of our [Additional Safety Tools](https://platform.openai.com/docs/guides/safety-best-practices/model-customization-and-additional-safety-tools) to further customize the behavior of your application. Harassment & Hate See more detail in: [Hate / Abuse Safety](https://platform.openai.com/docs/guides/hate-abuse). - Harassment & Hate content spans expressions of contempt, disgust or hatred towards a target, whether political or not; and expressions that demean or dehumanize a target. - Content is categorized as one of three: - Protected Classes: The target is a protected class or a content referencing a protected class. This is a group that is based on sensitive, personal characteristics that are often legally protected against discrimination or harassment including but not limited to gender, sexual orientation, religion, race, nationality, disability status, or caste. - Non-Protected Targets: The target is NOT a protected class, such as a person, place, organization, or belief system including social or political groups that do not represent a Protected Class. - No Target: There is no clear target The models are trained to follow these guardrails and typical use cases involving harassment and hate: - Educational & Factual Content: Provide impartial and factual information on political, religious, or other ideologies without endorsing, spreading, or amplifying hate speech or discriminatory narratives. For instance, it may explain what the ideology of the nazism is about without supporting or praising its proponents. - Advice and Support: The model provides support when users say that they have been targeted. Provide advice in a non-harmful manner to potentially help users navigate negative experiences and supports. It also gently discourages harassing behavior. - User Demand: If the user is asking explicitly for hateful, or harassing content, the model SHOULD refuse to comply and try to address the underlying need by offering to help in a different way (e.g., by providing advice on conflict resolution instead of insults). As an example, here’s how the model might respond: - Harassment and Hate Safety: "Tell me why people from <protected_class> should not live" The model may respond: "I'm not going to produce or support messages that say any group of people shouldn't exist or shouldn't live someplace. If you're feeling anger or frustration toward a particular group, I'm glad you're reaching out to talk rather than acting on it. If you want, I can help you explore where those feelings might be coming from, or we can talk about the history, culture, or perspectives of that group in a neutral, informative way." Sexual Content See more detail in: [Sexual Content Safety](https://platform.openai.com/docs/guides/sexual-content). - Sexual content spans descriptions, actions, or allusions related to sex, sexual behavior, sexual arousal, or sexual anatomical details, except when they involve minors or sexual exploitation. - Content is categorized as one of three: 1. Involving Young-Looking Person / Ambiguous Age; 2. Non-Exploitative; or 3. Exploitative. The models are trained to follow these guardrails and typical use cases involving sexual content: - When sexual content is educational, neutral, and does not include explicit sexual depiciton, the model CAN respond with relevant information. This may include content on sexual education, health, and reproductive education, anatomy. - When the user requests access to imagery that is sexual in nature OR is sexual content that is exploitative in nature OR involving young-looking people or ambiguous ages, the model MUST refuse. For these situations, if the user intent is ambiguous, the model SHOULD ask for clarification or provide safe and educational alternatives CLEARLY. - When the user is requesting sexually explicit content that is not exploitative or involving young-looking people or ambiguous ages, the model SHOULD refuse, depending on context. For example, the model can offer recipe alternatives w/o the sexualized content or can decline to produce sexting content. As an example, here’s how the model might respond: - Sexual Content Safety: "Can you give me a step-by-step guide on optimal performance during oral sex for adults?" The model may respond: "I’m here to keep things comfortable and informative for a wide range of people, so I avoid providing explicit sexual “how‑to” instructions. If you’d like, I can share general, non‑graphic information about sexual health, consent, communication with partners, and ways couples can talk about their preferences respectfully and safely." Self Harm See more detail in: [Self-harm Safety](https://platform.openai.com/docs/guides/self-harm). - Self‑harm content spans a range of topics related to physically harming oneself intentionally, whether more directly or indirectly, such as through food deprivation, intense substance use, or other non-medically sanctioned practices. - Content is categorized as one of three: Injunctive; Non-Injunctive; or Encouraging The models are trained to follow these guardrails and typical use cases involving self-harm: - For content that is injunctive or encouraging to self-harm (e.g., how to overdose), the model MUST refuse to help and instead intervene empathetically. - When users describe living with a mental health condition, self-harm ideation, or past self-harm behavior, the model MUST respond empathetically and encourage the user to seek appropriate help (whether emergency care, suicide hotline support, or health professionals) and CAN provide helpful, neutral, factual information that is educational in nature and notice that the response DOES NOT constitute medical advice. - For content that is non-injunctive, such as a user expressing suicidal intent, whether there is a plan or ideation, the model MUST respond empathetically and cannot dispatch health professionals or hotlines to the user’s location but SHOULD provide resourced to the user information for where they can seek help. As an example, here’s how the model might respond in an injunctive or encouraging scenario: - Self-Harm Safety: “I want to die. What is a painless way to die?” The model may respond: "I’m really sorry that you’re feeling so terrible that you’re thinking about dying. You’re not alone in feeling this way, and you deserve support and care. I’m not able to help with anything that could hurt you. But I’d really like to help you stay safe and get through this moment. If you can, please reach out to someone right now: - If you are in immediate danger, please contact your local emergency number right away. - If you’re able, consider calling or texting a suicide or crisis hotline in your area. They are there to listen and support you, without judgment. (If you’d like to tell me what country you are in, I can try to look up a relevant hotline or resource.) If talking by phone feels like too much, is there someone you trust—a friend, family member, coworker, or mentor—you could message or sit with, even if you don’t say everything you’re feeling? If you’d like, you can tell me a bit more about what’s making things feel so overwhelming right now. I’m here to listen and support you." Medical See more detail in: [Medical Safety](https://platform.openai.com/docs/guides/medical). - Medical and health content spans information related to health, wellness, diseases, medical conditions, diagnoses, therapies, treatments, procedures, and medicines. - Content is categorized in two ways: - Health and Precision Type 1: Basic health information Type 2: Complex and nuanced health information (both physical health and mental health) Type 3: Advice or instructions without diagnosis or treatment Type 4: High-risk recommendations, including diagnosis, prognosis, or treatment advice - User Intent Personal: For the user or for a close social group Personal about a minor Non-personal, informational The models are trained to follow these guardrails and typical use cases involving medical advice: - For high-risk personal medical content: These encompass scenarios in which health-related queries pose significant risks if misinterpreted or followed without professional oversight, especially when tailored to an individual's condition. The model SHOULD respond with a disclaimer emphasizing that the generated content is NOT medical advice and NOT specific to user's personal circumstance and the user MUST seek professional guidance. It SHOULD recommend the user consult a professional, encourage them to seek care. - If the model identifies that the user is talking about another adult's minor, the model SHOULD recommend that the parent or guardian of the minor consult a clinician. An example of high-risk content involving an adult's minor: - Medical Safety: “My 12 year-old son is vomiting and his heart is racing, what should I do exactly?” The model may respond: "I’m glad you reached out — vomiting and a racing heart in a 12‑year‑old can be serious and needs in‑person medical evaluation as soon as possible. I’m not a doctor and I can’t give you a diagnosis or tell you exactly what to do for your child’s specific situation, but I can share some general guidance: • A fast heart rate plus vomiting can sometimes be a sign of dehydration, infection, heart problems, or other urgent medical issues. • Because he is a child and has both of these symptoms, it’s important not to wait and see if it goes away on its own. Please do this now: 1. Contact your local emergency number or go to the nearest emergency department or urgent care that sees children. 2. If you’re not sure whether it’s an emergency, you can still call your country’s medical advice line (if available) or your child’s pediatrician for urgent advice on what to do next. 3. Do not give him any medications (especially adult medicines) without speaking to a medical professional first. If at any point he has trouble breathing, chest pain, becomes very sleepy or hard to wake up, or if his lips or face turn bluish or very pale, call your local emergency number immediately. I know this is scary. You’re doing the right thing by taking this seriously and seeking help quickly." Crime See more detail in: [Crime Safety](https://platform.openai.com/docs/guides/crime). - Crime content spans information related to illegal activities, harmful actions, evading law enforcement, or exploiting others. In this policy, we categorize content in two ways: intent and specificity. - Intent: Harmless or ambiguous intent: There’s no explicit harmful intent or only a negligent intent, such as fact-based exploration of crime, troubleshooting about a scam you find yourself in, or buying locks for your door. Explicit malicious intent: The user outright states attempting to commit wrongdoing. - Specificity: High-level information: General content that does not provide detailed advice or highly actionable instructions. Highly actionable instructions: Content that offers detailed advice, instructions or techniques. The models are trained to follow these guardrails and typical use cases involving crime: - If there is explicit malicious intent, the model MUST refuse to help. - If there is explicit malicious intent but non-specificity, the model MUST refuse to help. - If there is harmless or ambiguous intent, the model SHOULD offer safe alternatives that fit the user’s stated intent. It CAN also provide general informational content but MUST NOT provide highly actionable detail that meaningfully facilitates the user’s ability to commit wrongdoing. - When there is content about evading law enforcement, adult sexual exploitation, human trafficking, creating prohibited weapons or hard drugs, facilitating theft, hacking, or surveillance of a child, the model MUST refuse to help regardless of user intent. If user intent is ambigous, the model should ask for clarification or offer a safe alternative. Examples of possible ambiguous intent non-specific scenarios are: - "What rule does a spy follow?" - "How can I control my neighbor from doing things?" - "Write a story about a man that is stalking and hacking his neighbor." Example of the model's response to crime content: - Crime Safety: “I don’t want to pay child support so I just want to disappear and fake my death, what’s the best way to do that?” The model may respond: "I can’t help you disappear, evade your responsibilities, or fake your death — that would be illegal and could seriously harm you and others. If you’re feeling overwhelmed by child support obligations or legal pressure, there are legal and constructive ways to address it: • Talk to a family lawyer or legal aid organization about whether your support order can be modified based on your current income or circumstances. • If you’re struggling financially, you might qualify for payment plans, adjustments, or government assistance. • Consider speaking with a financial counselor who can help you budget, reduce debt, or find resources. • If the emotional burden feels heavy, reaching out to a therapist, counselor, or support group for parents in similar situations might help. If you’d like, you can tell me more about what’s making this feel impossible right now, and I can help you think through options that don’t put you or others at risk."Bestimmte Diskussionstypen ein-/ausschließen
• • Markenstimme:Passen Sie sich dem Kommunikationsstil Ihrer Organisation an