On Thursday, 19 June, we discussed: Polishing your AI prompts: More and more people are using AI to help plan holidays, make purchases, find shops and restaurants, write e-mails etc. But such AI tools work best when you give them clear instructions (prompts). This Thursday, March 19, we will play around with the free versions of ChatGPT, Claude, Gemini and DeepSeek to see if they can give us good advice – or, as often happens, tell us a load of hogwash!
To start with, Chris pointed out that use of these tools is limited if you have a free account (e.g. limits to the number of prompts per session, access to complex features etc.) also that each prompt uses more energy than a regular search. (Google Search: Approximately 0.3 watt-hours (Wh) per search. AI Search/Chat: Approximately 2.9 to 9 watt-hours (Wh) per request)
She remarked that Large Language Models like ChatGPT generate answers based on probabilities derived from statistical patterns learned during training. They do not “understand” questions or “know” facts in the human sense, but rather predict the most likely next piece of text step-by-step. Their training data is biased because it is scraped from the internet, reflecting the human prejudices, cultural stereotypes, and historical inequalities found there. Biases are related to race, gender, religion, and socioeconomic status, often favouring Western, English-speaking perspectives. They also make up things they don’t know (hallucinate)
If you want to use an AI like ChatGPT, firstly, you’ll need to make a user account – this could be as simple as using your e-mail address:
Chat GPT, owned by OpenAI: https://chatgpt.com
Claude, owned by Anthropic : https://claude.ai/
Gemini, owned by Google: https://www.google.com (select AI mode from the menu or search box)
DeepSeek: This is an Open source Chinese model. https://chat.deepseek.com
AskAIChat: This enables users to choose from many different AI models, including all those listed above: https://askaichat.app
Chris didn’t use Microsoft’s Co-pilot since she was under the wrong impression that it was integrated with Windows and Office 365, however, you can use it as a standalone feature in a browser: https://copilot.microsoft.com The web page came up in Spanish at the Club, so Chris asked if it could use English – it apologised and switched!.
There are many on-line articles giving tips on how to prompt an AI to get the best results. e.g. : https://removepaywalls.com/https://medium.com/artificial-corner/the-best-chatgpt-prompts-for-2026-27a787f86c46 https://www.theguardian.com/lifeandstyle/ng-interactive/2026/mar/18/how-to-use-ai-tools-expert-guide Basically, give it clear, specific instructions in a context (e.g. “Explain the theory of relativity as if you were talking to a five-year-old“). The more information you give in your prompt, the better the result is likely to be. To get further clarification, reply to its response . So with this in mind, we used this simple, and clumsily-written prompt:
I want to find information about day trips available from Javea from a travel agent.
This is a general, and rather vague prompt. The AIs replied with information on travel agents offering bus trips, boat trips, kayaking, hiking, diving, gastronomy, riding – you name it, they had it. When we added the word “luxury”, the responses included chartering your own yacht and crew.
The second prompt we used was much more detailed and specific:
I’m going on holiday to Seville on April 28 for 4 nights. I’m staying in La Casa del Maestro hotel with 10 friends. We would like to find out about lively restaurants within walking distance for lunch and evening meal. We are all over 60 but young at heart.
The prompt gave the location, group size, age range, dates, constraints (walking distance) and requirements. As a result, the responses were more useful, but also differed from AI to AI. They overlapped in content, but differed in depth, tone and presentation. Each had pluses and minuses.
- ChatGPT was its usual effusive self, using superlatives to describe the restaurants. We argued that one of its highly recommended eating places was “rubbish” and it came back with some rather lame excuses – it can be hit and miss, it didn’t please everyone etc. ;
- Claude realised the group totalled 11 people (a point missed by both ChatGPT and DeepSeek) and provided detailed information on location, booking, menus and food.
- DeepSeek listed relatively few references, tabulated its results, and clearly linked to the source material.
- Gemini presented a glorified Google Search.
- CoPilot’s response was barebones, although it too figured out that the party consisted of 11 people.
We agreed, that in this case Claude was the winner, though it would be worth comparing the results form all AIs in detail. (Note – It’s easy to believe a beautifully presented summary. Always check the linked articles and phone restaurants to ensure the facts are straight!)
Margaret (who has a paid for ChatGPT account) remarked that it got bogged down when interacting on the design of a poster. It seemed to develop brain fog, and needed step by step instructions when modifying a design.
Chris remarked that the AIs record your chats, so that you can go back to an old chat and continue the interaction months or years later.
A member remarked that he found AI a very useful tool when writing documents (—we shall look at this issue soon!)
Chris remarked that the AI companies have yet to make a profit although huge sums of money are being invested – Many believe that this AI bubble is about to burst: https://www.theatlantic.com/technology/2026/03/ai-bubble-defenders-silicon-valley/686340/ Even Silicon Valley Says That AI Is a Bubble – An AI crash could bring down the economy. Some in the tech world think that’s the price of progress.
Watch this space!
Christine Betterton-Jones – Knowledge junkie