VI #003: Maximizing the Benefits and Safety of ChatGPT: A 3-Step Practical Guide For Users

Read time: 6 minutes

ChatGPT, a powerful AI tool created by OpenAI, has taken the world by storm, taking only 5 days to reach 1 million users following its launch on Nov 30, 2022. An alleged 500 million people have already experimented with its capabilities and many are writing about what can be achieved with it, however, most people have no clue how to leverage it in a way that maximizes benefits while avoiding accidental misuse. OpenAI CEO Sam Altman and many AI experts have warned about the potential dangers of misusing AI whether accidental or otherwise. From my experience working in AI/ML, particularly including my time building and running a Machine Learning group handling sensitive data for one of the world’s most admired companies, protecting from misuse of AI is something I care about deeply as well.

To ensure proper and safe use of an AI tool, it is important to establish a framework and use best practices such as pre-processing techniques, human moderation, and keeping track of model performance. It's also important to be aware of training data, biases, and legal and ethical considerations. However, since ChatGPT is already available via open beta to the general public, it’s in everyone’s best interests to learn and adopt such best practices. Here, I attempt to collate a practical guide that I’ve personally found helpful as a starting point for using ChatGPT effectively while minimizing the potential for accidental misuse.

Quick Intro To ChatGPT

In case you’re unfamiliar with it, here’s how ChatGPT describes itself:

Alongside other large language models such as Anthropic’s Claude, ChatGPT is one of the most advanced chatbots in the world right now. It can offer answers to everything from creating online content to parenting advice and writing computer code.

What’s the concern then?

ChatGPT is improving all the time, however, current known limitations are below and some of these are active areas of research that are very hard to solve:

• Sometimes writes plausible-sounding but incorrect or nonsensical answers
• Is sensitive to tweaks to input phrasing or attempting the same prompt multiple times
• Is often excessively verbose and overuses certain phrases
• Guesses what a user intended, rather than asking clarifying questions
• Sometimes responds to harmful instructions or exhibits biased behavior

As with any technology it can be used for good or for bad, whether intentionally or otherwise.

So then… how can we maximize the good and minimize the bad?

A 3-step framework for maximizing ChatGPT benefits and safety

At the risk of stating the obvious, using ChatGPT well can be boiled down to a simple process:

1. Provide appropriate inputs, precisely and completely
2. Sanitizing outputs thoroughly and consistently
3. Use the outputs appropriately

By following these steps with care, thoughtfulness, and common sense we have a better chance of achieving high quality results from ChatGPT repeatably, in a way that provides benefit rather than harm.

1. Providing appropriate inputs, precisely and completely

Guidelines for providing input to ChatGPT:

1. Be specific and clear in questions or prompts
2. Use complete sentences and proper grammar
3. Provide sufficient context or background information
4. Try rephrasing if needed to optimize results
5. Avoid asking multiple questions at once
6. Use specific keywords if looking for specific information
7. If needed, specify output e.g. "write a story about...", “write code to…”, etc
8. Be aware of legal or ethical considerations when providing input, notably protecting sensitive information or perpetuating harmful biases
9. If using the API:
   1. Be familiar with the documentation, use the appropriate model endpoint/version for the needed task and complexity, and use completion options to fine-tune output
   2. Follow OpenAI's safety best practices
   3. Consider using the model with other tools, techniques, or models to gain better results

2. Sanitizing outputs thoroughly & consistently

Verify and edit ChatGPT outputs in a structured, repeatable way to sanitize before then using them. Here are some items to watch out for in its output:

1. Bias:
   1. Look for underrepresented or marginalized groups, stereotypes, offensive or insensitive language, and regularly update any fine-tuned models
   2. Test with different inputs, get a second opinion, and review any fine-tuning data sets to further understand and identify biases
2. Mistakes, inaccuracies, gaps in knowledge, coherence:
   1. Cross-reference with reputable sources, check for inconsistencies, lack of context, typos
   2. Test with different inputs, get a second opinion, and understand the training data for errors and gaps
   3. Note that at the time of writing this, ChatGPT has only been trained on data up to around 2021, so its output should be considered in this context
3. Repetition:
   1. Look for repeated phrases or sentences and consider using a tool to detect repetition, such as a plagiarism checker
4. Lack of context or personalization:
   1. Update inputs with more context relevant to your use case for more relevant results
5. Lack of creativity, social awareness, actionability, and style:
   1. Review output and edit it with a human touch to make it sound more natural, engaging, original, and actionable as appropriate
   2. Experiment with the inputs, such as specifying desired tone, branding, or format. More diverse input data can lead to more varied and creative output

3. Using outputs appropriately

Text from ChatGPT should always be reviewed by a human before being used in any official capacity. Appropriate use depends on the specific context and use case. However, in general, it is considered appropriate to use the outputs from ChatGPT for:

1. Language understanding and language generation tasks, such as natural language processing, text summarization, and text completion
2. Research and development of new AI technologies, such as machine learning and NLP
3. Automation of repetitive tasks, such as for helping with data entry and customer service
4. Content creation such as generating text for articles, stories, or scripts
5. Educational and training purposes, such as language learning or teaching students about AI

It's considered not appropriate to use the outputs from ChatGPT for:

1. Spreading misinformation or disinformation
2. Automating tasks that require decision making or critical thinking
3. Generating content that is offensive, harmful, or discriminatory
4. Misrepresenting the model's outputs as your own work, or using it to plagiarize content
5. Using the model to gain unauthorized access to personal information
6. Breaking any laws or regulations

That’s it for today.

In Summary

ChatGPT’s capabilities are constantly evolving and it is continually learning from further training and feedback it receives, so it's important to keep up-to-date with its features, limitations, performance, ethical and legal considerations, and best practices. A practical starting point I’ve found helpful for using ChatGPT effectively and safely is below, I hope you may find this helpful too:

1. Provide appropriate inputs, precisely and completely
2. Sanitizing outputs thoroughly and consistently
3. Use outputs appropriately

ChatGPT is a powerful tool that is supercharging human creativity around the world at an incredible rate. Though it is not a substitute for human intelligence, creativity, or judgement. It's important to use it in conjunction with human oversight, review, and editing, to ensure that the output generated is accurate, professional, polished, and minimizes or eliminates misuse and harm.

See you next Sunday.

Whenever you’re ready, there are 2 ways I can help you:

1. Work with me 1:1 to build your team, product, platform, or career.
2. Book a free discovery call with me to explore if your business needs would be a good fit for my advisory services. If we’re not a good fit, rest assured I’ll kindly let you know.

Disclaimer: Please of course use your own judgement, any reliance placed on the info here is at your own risk. And I am neither advocating for nor against use of AI such as via ChatGPT, rather recognizing that such tools are here to stay and are only going to become more advanced, ubiquitous, and pervasive. Change can be hard, yet I believe it’s in our interest to adapt and learn how to adopt and shape new technology in a way that can be beneficial and safe for all. This is one small attempt to help towards this, I hope you may find it useful.

And thanks to ChatGPT for helping ideate and organize this article.