Three Simple Ways to Make ChatGPT Outputs More Accurate

Three Simple Ways to Make ChatGPT Outputs More Accurate

The challenge with ChatGPT is that it's not always reliable. The way in which the model works (predicting the next word based on its training) means that it will sometimes give you incorrect or downright nonsensical responses that sound plausible.

However, there are a few simple ways in which you can output better responses. Let's go through them in turn.

#1: Ask it to only respond using reliable sources which it should also cite.

In the same way that when we write essays we need to say where our information comes from, we can ask ChatGPT to do the same thing. Remember that it's trained on the internet's data up until 2021 so any sources beyond that will likely not be available.

By writing a prompt such as: 'Is the earth flat? Answer only using reliable sources and cite those sources', the model will give you quotes and acknowledgements from NASA.

#2: Force it to give you its reasoning

By forcing the model to think in steps you can get far more accurate and detailed responses. Think of it like questioning a student to get it to give their rationale for their ideas. The deeper you go, the more you make it think.

You should first define the paramaters of this exchange. You can do this by including this initial prompt:

Use this format:

Q: <repeat question>

A: Let's think step by step.

<give reasoning> Therefore, the answer is <final answer>

What you're doing is asking ChatGPT to answer in a specific format and to fill in the blanks indicated by the parentheses.

By including this after your initial question 'is the earth flat' you will get a response like this:

Q: Is the earth flat?

A: Let's think step by step. The shape of the Earth can be determined by observing the shadow it casts during a lunar eclipse, measuring the curvature of the Earth's surface, and analyzing the deflection of starlight passing near the Earth. The scientific evidence gathered from these observations, as well as from satellite imagery and global navigation, all indicate that the Earth is not a flat surface, but an oblate spheroid, meaning it is slightly flattened at the poles and bulging at the equator. Therefore, the answer is no, the Earth is not flat.

The model is thinking in a more rational, logical way, which is what we want it to do.

#3: Ask it if there is anything factually incorrect in its response

This one is simple, as it is asking the model to review its response. You can use this after any response just to check if you're not convinced it's got the answer totally correct.

To do this, simply write 'Is there anything factually incorrect in this response?' once the model has outputted.

By using these three additional prompts, you should find your responses are more accurate and truthful. However, nothing beats checking yourself, as the 'human in the loop' approach is usually the best option. More on that in later articles.

If you've found any other ways to increase the accuracy of ChatGPT please do share them in comments.

Daniel Ang

Microsoft Azure | TOGAF 9 | Power Platform | Transformation | Change Management | FinOps | OpenAI ChatGPT | Speaker

1 年

Very insightful!

回复
Keith Brown

Headmaster at Willington Prep Wimbledon, ISI Team Inspector; ILM Executive Coach; Member of IAPS; School Governor; Google Certified Educator and brand ambassador at Taisk.com

1 年

Great tips Darren, thank you!

回复

要查看或添加评论,请登录

Darren Coxon的更多文章

  • Seven Reasons Why Schools are Not Ready for AI

    Seven Reasons Why Schools are Not Ready for AI

    We are all reading the excitable rhetoric on how AI is going to change the world, and many of us are having lots of fun…

    14 条评论
  • Unlocking Creativity with AI: How My Seven-Year-Old Son Wrote His Own Adventure Novel

    Unlocking Creativity with AI: How My Seven-Year-Old Son Wrote His Own Adventure Novel

    Last week my seven year old son Misha wanted to write a story but was stuck on how to start. I’d not yet introduced him…

    6 条评论
  • Learning Design in the Wake of AI

    Learning Design in the Wake of AI

    There are profound things happening in AI, and I’m not talking about the technical side. What I’m seeing, for the first…

    4 条评论
  • ChatGPT: Your Revision Guide Maestro

    ChatGPT: Your Revision Guide Maestro

    Last week my eldest daughter came to me rather stressed. She’s due to sit her history GCSE in May and cannot find…

    2 条评论
  • 3 Reasons Teachers Should Not Fear AI (And One Reason They Should)

    3 Reasons Teachers Should Not Fear AI (And One Reason They Should)

    There is so much buzz around AI right now, and every sector is considering how it might impact them. Sam Altman, CEO of…

    6 条评论
  • Getting the most from ChatGPT

    Getting the most from ChatGPT

    Here's an interesting thing. We've been thinking a lot about the nature of writing with large language models (LLMs)…

  • Will ChatGPT kill the art of essay writing?

    Will ChatGPT kill the art of essay writing?

    ChatGPT is now three months old. And, as in all truly disruptive technologies, it’s hard to remember what the world was…

    5 条评论
  • Chat GPT and the importance of style

    Chat GPT and the importance of style

    Try this as an experiment. Ask Chat GPT to write you an article on the future of AI in education.

    4 条评论
  • Using ChatGPT as an Educational PA

    Using ChatGPT as an Educational PA

    If you've read my most recent articles, you'll know that the recent release of ChatGPT from open.ai has got many in…

    2 条评论
  • How to detect AI Generated Essays

    How to detect AI Generated Essays

    In my first article on AI in education, I outlined concerns many have over the increase in AI generated content. The…

    2 条评论

社区洞察

其他会员也浏览了