Tips, Use Cases and Limitations

SCORE's Perspective on Generative AI
- Generative AI offers tremendous opportunities for mentors to enhance their service to clients. It is essential, however, to approach it with the following beliefs:
  - It is useful for certain tasks, but human judgment is essential.
  - The inherent limitations and risks require governance and oversight.

Permissible AI Usage

Data Type	What You Input into an AI Tool	AI Tools
Sensitive	Personally identifiable information	None
Strategic	Internal information, client communications, non-public performance data	ChatGPT (training mode turned off)
Public	Public information, general industry information	Bard, Bing Chat, ChatGPT

It is important to recognize that different AI tools handle data we input differently. We have to consider how data can be accessed and how it might be used to generate future output for users outside of SCORE. These guidelines establish a framework for responsible AI Use based on data sensitivity so we can protect our clients and SCORE in accordance with SCORE’s Personal Information Protection Policy.

Permissible Usage Examples
- Sensitive:
  - "Gary Irvine is struggling with his business and I need some advice on how to coach him"
  - Routing number 21194201 Account Number 21002001001001
- Strategic:
  - "Summarize the following transcript of a meeting I had."
  - "Give me a response to the following email I got from my client."
  - "Use the following information about this business to write an About Us page for the website"
- Public:
  - "Access the online presence of Gary's Towing in Granville, OH, and compare it with their top 3 competitors."

Turning Off Training Mode or Opting Out of Training in ChatGPT
- Unless you turn off or opt out of training mode in ChatGPT, it may use your inputs to train the data model. SCORE recommends turning off training mode.
- Register an account with your SCORE email address
- Once logged in, click the three dots next to your email address or name
- Click Settings
- Click Data Controls
- Toggle Chat History & Training off
- Alternatively, you can use this “User Content Opt Out Request” form.

Limitations and Considerations
- Knowledge cutoff: Many tools use knowledge that is not current. For example, the current version of ChatGPT only knows about events and information up to September 2021.
- Inaccurate information: Output may contain information, that is incorrect, commonly called "hallucinations".
- Preparation counts: Output is often only as good as the input you give it. The better context and more information you can give an AI tool, the more likely you will get useful output.
- Watch out: Be aware that using Plugins, Advanced Data Analysis, or other beta features will automatically turn Chat History & Training data capture back on.
- Not infallible: ChatGPT can make mistakes. ChatGPT requires double-checking. Some of its answers can sound authoritative and current but are wrong. GPT-4 makes about 60% fewer mistakes than GPT-3.5.
- What it considers: AI lacks human insight. It only "knows" what it has gathered from the past.
- Possible updates: Web browsing plugins are available in ChatGPT-4 to search for current information.

Limitations and Considerations Questions and Answers
- Is the information I am getting current? Many tools use knowledge that is not current. For example, the current version of ChatGPT only knows about events and information through September 2021. Web-browsing plugins are available in GPT-4 to search for current information. Remember: AI only "knows" what it's seen in the past.
- Is the answer I'm getting always right? Sometimes output can contain information that is incorrect. Those mistakes are commonly called "hallucinations". Everything output you receive requires double-checking. Some answers sound authoritative and current but are wrong.
- How detailed should my request be? Output is often only as good as the input you give it. The better context and more information you can give an AI tool, the more likely you will get useful output.
- What do I need to worry about technically? Be aware that using Plugins, Advanced Data Analysis, or other beta features will automatically turn Chat History & training data capture back on.
- Anything I can do to increase accuracy? Chat-GPT-4 makes about 60% fewer mistakes than its 3.5 counterpart.

Tips and Techniques
- Output is only as good as the input you give it. Using effective "prompts" (what you input into an AI tool) is a great way to make AI tools more useful to you.
  
  Good Techniques to try:
  - Be careful when using client-confidential information: If you do not have their permission, do not use it. This includes proprietary, personal client information, or other innovative ideas. See SCORE’s Personal Information Protection Policy.
  - Be clear and explicit: Provide specific details in your prompt to get a precise response. Example: "What were the key milestones in the development of artificial intelligence during the 1990s?"
  - Think of ChatGPT as an expert assistant that needs supervision.
  - Verify the "factual" information ChatGPT provides.
  - Ask the AI for the references it used, including URLs.
  - You can ask the AI to limit its knowledge sources and references to a specific domain, e.g. score.org.
  - Require that qualified people use ChatGPT. Example: If you're not an attorney, ChatGPT won't make you one.
  - ChatGPT might use your input as training material unless you opt-out – be careful about confidential information (See - SCORE’s Personal Information Protection Policy.
  - Wholly AI-generated content cannot be copyrighted.
  - Treat any output you get from an AI tool as a first draft. It can be useful in getting you started on a task, but it will need you to look it over and likely make changes.
  - If someone else saw your input and it would be embarrassing or damaging to a client or SCORE, don't put it into an AI tool.
  - If you're unsure what tool you are allowed to use in any given situation, err on the side caution and ask your Chapter Chair or email help@score.org

Ethical Issues
- Social
  - Racial, gender, and social biases can be created inadvertently by ChatGPT. Again, everything should be double-checked.
- Intellectual
  - Plagiarism
    Chat GPT results may contain the work of others.
    It's best to check for plagiarism using a tool like Grammarly.
- Copyright issues:
  - Wholly AI-generated content cannot be copyrighted.

Acceptable Use
- OpenAI limits some of the ways ChatGPT can be used, such as for political campaigns. https://openai.com/policies/usage-policies

Mentoring Use Cases
- Researching business "unicorns"
  "How do I market and sell my NFTs?"
  "What are local regulations for using my barn as an event venue?"
- Drafting a client email using score.org resources
  "How do I start a food-truck business and please use only score.org URLs"
  See Bing result below
- Finding an answer to a technical question out of my state
  "How do I move an LLC from VA to OH?"

Common Generative AI Terms
- Few-shot / One-shot Technique: AI learning from a single or few examples.
- Generative AI: AI that creates new content similar to its input data. It can make things like text, images, or music.
- Hallucination: When provide information that is incorrect.
- Large Language Model: An AI model trained on a lot of text. It generates sentences by guessing what word comes next, like GPT-4.
- Prompt: The input given to an AI. It's what you ask or tell the AI to get a response.
- Prompt Engineering: Tweaking the prompt to get better results from AI. It can involve changing the question or adding more details.

Comments

Address:

Phone:

Technical Support:

Related articles