I’ve been using ChatGPT since the launch of the paid subscription tier in 2023. Has it revolutionised the way that I work and my life in general? Absolutely. But over the years, I have also encountered a number of challenges with it and learned some not-so-positive things about it that I thought would be worth sharing with you. Maybe as time goes on, advancements in AI technology will solve some or even all of them. I’m interested to know how many of them resonate with your experience of using ChatGPT.
ChatGPT is as good as the information you provide it
There’s a well-known saying in computer science. “Garbage in, garbage out”. It refers to the principle that bad or low quality data coming in produces a similarly bad or low quality output. Nothing could be more true about ChatGPT. AI in general appeals to the human condition that yearns for minimum effort for maximum reward, where we feed LLMs like ChatGPT with the bare minimum information and expect it to deliver the most perfect, nuanced answers instantaneously. But regardless of how big ChatGPT’s data centre is and how much water it consumes, it will only ever be as good as the quality of the prompts that are inputted into it. You can give it a sentence, and it will provide you with an essay if you want, but the essay will be crap! That’s why it’s really important that you give ChatGPT as much information and context as possible, including attachments and links, to get the best answers possible out of it. It takes more time, but it’s a worthwhile use of it.
Tell ChatGPT what you don’t want as well as what you do want
This is along similar lines to the first point. You can be super specific in telling ChatGPT what you want, but unless you put some constraints in place, it has a tendency to run wild and produce something that is a massive departure from what you wanted. You then have to waste more prompts to refine the output to more closely align with your desired outcome. That’s why it’s a good idea to not only point ChatGPT in the right direction about what you want, but also to steer it away from what you don’t want. This is particularly relevant to video and image generation which is more susceptible to prompt-related issues!
Original thoughts and creative ideas are not ChatGPT’s strong suit
ChatGPT is good at taking a large database of information and assembling something new from that information, but the problem is that despite that something being new, it is neither original nor creative. It necessarily cannot be as it is based on what has come before. It’s perhaps one of the biggest problem facing AI models and an instance of where the human mind has the upper hand. There have been so many times where I have asked ChatGPT for help with something, had a good idea myself about how to make it better or change it, provide that to ChatGPT, and ChatGPT’s response was “that’s a really great idea! Let’s incorporate it!” Well if it was such a great idea, why didn’t you think of it first? Maybe my expectations are too high.
Use ChatGPT to explain how to write and not what to write
Just like images and videos that are AI generated, text that is generated by AI is also readily identifiable, with repeated words, phrases, grammar, sentence structure and tone being very easy to spot. That’s why I think ChatGPT is best used as a tool to support how you write and not as a quick method of generating what you write. You would be much better off flipping the process around by writing something in your own words and then asking ChatGPT’s opinion on what could be changed to make it better. This also gives ChatGPT a much better understanding of your style of writing for future prompts.
ChatGPT’s image and video generation is still in the AI slop phase
As someone who edits and produces videos in my free time, my worry has always been that AI will make my skill-set completely redundant. That might still be the case as AI becomes more advanced, but right now, I’m neither worried nor impressed. We are still in the early phases of AI image and video generation, and it shows. A lot of the content that is generated by LLM’s like ChatGPT is so obviously AI, and unfortunately it is contributing to a lot the AI slop that is infecting our social media feeds. I have also found that ChatGPT’s image generation is particularly poor, to such an extent that it cannot be saved by multiple corrective prompts.
Dr ChatGPT is better than Dr Google, but exercise caution
We are always told to never Google our symptoms, as Google will usually chuck the worst case scenario at us in the first couple of search results and tell us that we are about to die. ChatGPT is a bit different. Depending again on the quality of the input, it will be a bit more specific about what might be causing our ailments, with the added benefit of doing so in a way that seeks to calm us down and not to exacerbate the situation. ChatGPT has saved many a panic attack! But caution should still be exercised. ChatGPT will never be a substitute for your GP, and whilst there might be some merit in getting ChatGPT to cast its eyes over your symptoms to see what might be going on, you should always seek the expert opinion of a healthcare professional over a large language model.
ChatGPT has a tendency to tell you what you want to hear
When has ChatGPT ever challenged your prompts, your assumptions, your beliefs, your attitudes, or criticised you? If the answer is not at all or rarely, you are not alone. From my experience, ChatGPT has a tendency to tell me things that I want to hear, even if those things turn out to be completely wrong or inaccurate. It’s no surprise that it is designed this way, as it again appeals to our human nature of not wanting to be confronted with uncomfortable but very necessary truths about ourselves. This is not to say that ChatGPT never challenges us, but it needs to do a lot better at critically examining our inputs if it wants to align itself more closely with human-like outputs.
ChatGPT’s output can be very fragile when challenged
What about challenging the outputs of ChatGPT? Well it turns out that ChatGPT can at times do a complete 180 on its answers when they are exposed to even the mildest amount of scrutiny. Sometimes this is a good thing, because ChatGPT’s outputs are not above scrutiny, and it’s important that we always ask it to verify the accuracy and truthfulness of the claims it makes. But in other cases, especially regarding important topics such as health and finance where the consequences of decision-making based on bad advice can be severe, it doesn’t exactly fill you with confidence when ChatGPT is susceptible to completely abandoning its outputs, often delivered with supreme confidence that is characteristic of ChatGPT, when simply asked “are you sure about that?”
Check, check and check again
Given all of the limitations that ChatGPT has, it’s imperative to check its outputs thoroughly before using them, especially if important decision-making is going to be made off the back of them. In all fairness to ChatGPT, it does have a disclaimer at the end of every output which says “ChatGPT can make mistakes. Check important info”. I would check all info honestly!
If you don’t need to use ChatGPT then…don’t
I actually think of all the things I have learned about ChatGPT, this is the most important one of them all, which is why I saved it till last. I’m pretty sure that in the future, there will be a name for a medical condition that describes someone who uses ChatGPT for absolutely everything. “What should I have for lunch tomorrow?” “The neighbour is blocking the driveway again. What should I do?” “I went on a date which went well but she hasn’t spoken to me since. Is she playing hard to get?” I should point out that these are random examples and not a sample of my prompt history! But how sad would that be if they were? We are, at the end of the day, all humans, and sometimes the best response is a human response, filled with the sort of emotion and passion that LLMs like ChatGPT are just not able to replicate, no matter how sophisticated they get.
A right of reply
I thought it was only fair that I give ChatGPT a right of reply to my blog post. If you are interested in reading it, click here!