Can customer service chatbots go rogue? ft. Getting started with prompting!

✨Welcome to untechnical✨
This is tech content made for the untechnical.

For those who are new here, this weekly newsletter is an easy and approachable way for you to incrementally build your knowledge and skills in all things new tech, AI and automation, while also staying up-to-date with a curated summary of the week’s AI/tech news.

So let’s get into it!

Tech Tip of the Week 🍎 

Prompting

AKA the most important skill to learn when adopting generative AI.

If you’ve tried ChatGPT and have felt a bit disappointed with what the tool’s produced, or if you can’t seem to get the output you want, it’s probably because you need to work on your prompting skills.

Up-leveling your prompting skills is the most effective way to get better results out of ChatGPT or similar AI tools.

Over the coming weeks, you’ll learn to prompt like a pro, but first things first, let’s start with the basics.

What is a prompt?

• Prompts are the input you feed into tools like ChatGPT, Claude, DALL-E, Midjourney, etc.

• A prompt includes instructions or a question, along with other details such as context, and examples, which help ChatGPT generate better and more accurate responses.

• For example, a prompt could be anything from ‘generate a pancake recipe for me’ to ‘rewrite this cover letter for me to sound more professional.’

Getting started with prompting:

There’s a lot to learn about prompting, but you’ll see a massive uplift in your ChatGPT outputs by mastering the basics. Let’s take a look below at some of the key elements that go into composing a solid prompt.

Building on the above tips, you can string these concepts together to come up with a basic prompt template to follow:

• RTF: Role, Task, Format

• CTF: Context, Task, Format

Later in the Playground section of this edition, we look at the immediate difference prompts make to ChatGPT’s responses ⬇️

Weekly News Digest 🗞

There’s no shortage of news about tech, AI and automation but it’s hard to know what you should tune into. Here, I summarise noteworthy news from the week and why you should care.

1: UK courier company’s customer service bot goes rogue

What happened? 

• A UK parcel firm, DPD, had to disable its AI customer service chatbot after a frustrated customer coaxed the bot into telling a joke, swearing and composing two poems about how bad the company’s customer service was. (Eek!)

Tell me more:

• The disgruntled customer, Ashley Beauchamp, a pianist and conductor posted his exchange with the bot on X (Twitter) last Thursday. Since then, his post has been viewed 2 million times.

• DPD has said the company had been using an AI element in its online chatbot system successfully for several years, alongside human customer service but an error had occurred after a system update.

Why is this important/interesting? 

• Just because we can introduce chatbots to support customer service, doesn’t mean they’ll be loyal servants.

• When implementing AI solutions, these need to be rigorously stress-tested to ensure they can withstand people trying to break and corrupt them 😜

2. Math geeks have competition with Google Deepmind’s Alpha Geometry

What happened? 

• Google’s Deepmind (AI research and development company) has built AlphaGeometry, an AI system that can solve complex geometry problems from Mathematical Olympiads; the world’s most prestigious high-school mathematics competitions in the world.

Tell me more:

• By completing 25 out of 30 Olympiad geometry problems within the competition time limit, not only could AlphaGeometry be a contestant at an Olympiad, but the system is approaching a human Olympiad gold-medallist level.

• The system utilises two types of AI algorithms to solve these mathematical problems:

  1. Language model: great at identifying general patterns and relationships in data

  2. Symbolic engine: based on formal logic and uses clear rules to arrive at conclusions

• Using two types of algorithms means AlphaGeometry can use the strengths of each algorithm to more quickly and accurately solve the problem put in front of them.

• Google tested the system’s step-by-step problem-solving process to answer questions that when analysed by a former Olympiad gold-medallist, Evan Chen, AlphaGeometry was praised for how it could work through problems ‘just as [human] students do.”

Why is this important/interesting? 

• While AlphaGeometry can only be applied to geometry problems at present and couldn’t compete across all subject-matter areas of the real Olympiad competitions, the achievement demonstrates the impressive applications of this type of solution and how AI can mimic more human-like reasoning.

3. Think you can tell the difference between real and AI-generated images of people? Think again.

What happened? 

• The New York Times has released a quick 10-question quiz this week prompting readers to guess whether images of people are real or AI-generated. 

Tell me more:

• It turns out I’m shockingly poor at differentiating between what’s real and what’s not. I scored 40% 😬

Why is this important/interesting? 

• The quality of images included in the NYT quiz illustrates how far generative AI images have come in a frighteningly short space of time.

• Last week, we talked about the prevalence of deepfakes and this week’s NYT quiz has reinforced how difficult it is getting to determine what’s ‘real’ and what’s not. It’s another reminder that we mustn’t blindly believe everything we see online!

Playground 🪁

Experimenting with new tech, AI and automation is one of the best ways to learn. But it can be hard to know where to start or what to play around with. Here, I provide an example of what I’ve been playing around with recently to give you some inspiration.

ChatGPT: Prompting in Practice

Earlier in this edition, we looked at the basics of prompting (i.e. how we instruct tools like ChatGPT to do what we want them to!)

So, let’s see how much of a difference a ‘good’ basic prompt makes when interacting with ChatGPT…

Prompt 1 (Weak prompt): Can you tell me about common cloud structures

Observations: While it’s a very decent output given the vague input, it’s a lot of information to digest. So I think we can do better with some more tweaks!

Prompt 2 (Better prompt): Act as a meteorological expert. Give me the top 6 most common cloud types. Produce a table to share your answer with columns [cloud name], [characteristics], [produce precipitation], [frequency in UK].

Observation: ChatGPT willingly assumed the role of a meteorological expert and the response is more concise. The tabulated format is also much easier to read and absorb all the information!

Conclusion: While we could keep tweaking the prompts to produce even better results, it’s hopefully clear that by making simple changes to our prompts, we can see immediate improvement and results.

So why not give these tips a go this week?

• Be specific

• Act ‘as if’

• Use examples

• Impose constraints

We’ll be diving into some more prompting tips next week 😃 

Before you go, if you think anyone else would benefit from reading untechnical newsletters, please share far and wide!

Feedback 🎯

This newsletter is a two-way street. Learning more about what you’d like to see more/less of and what content is going to support you best is incredibly important.

💬 Do you have any feedback that may help improve future editions of untechnical? Any questions? Or content requests?

Get in touch by replying to this email.

That’s all for this week! Until next Tuesday…

✨ Thanks for tuning in! ✨