5 Failings of ChatGPT and Generative AI (and How to Fix Them)
- By Bilal Hafeez
- Sign-up to discuss on Slack
- 4 min read
Interacting with ChatGPT is exciting, scary and even magical. It really does feel like there is another person on the other side of the interface. No wonder people are talking about the rise of AI robots to take over the world.
Yet the debate is polarised. Tech bros say this will revolutionise the world for good. Meanwhile, doomsayers think the end of the world has been brought closer.
What is missing is a realistic take on generative AI – the catch-all term for things like ChatGPT and Midjourney. Notably, many people are missing some of the clearest failings of the current set of generative AI models. So here are five that stand out to me.
1. ChatGPT Cannot Understand Like a Human
At their heart, ChatGPT and other Large Language Models (LLMs) are amazing pattern recognition and data-mining models. Ask them a question, and LLMs will use a huge corpus of human generated text (the internet), shove it into a neural network model (a so-called transformer), and spit out the most reasonable-sounding answer.
It does this by guessing the best first word, then running that word through its dataset (text from the internet) and model (neural network) to see what the typical word is to follow. It adds that to the sentence, then repeats the process to find the next most likely word and so on.
Yet it does not pick the most probable next word – it adds randomness (or ‘temperature’, in AI speak) to make the answers more interesting. If you say the answer is not right, it will cleverly use your feedback to improve on the answer. In technical terms, it is using reinforcement learning from human feedback (RLHF). It sounds like a human because it is trained on human text.
LLMs’ ability to create ‘new’ sentences and answer questions naturalistically is a genuine breakthrough. Its underlying ‘thinking’ engine is pattern recognition, which humans also do a lot whether consciously or not. It will be able to do that much faster than any person.
But humans do more than pattern recognition. Humans can understand why things happen. They can come up with laws and theories that can explain yet-to-be-seen phenomena (i.e., science). And they come up with creative leaps that have never been seen before.
That is where LLMs and generative AI fall apart. You can ask it something, and it can come up with a convincing answer. But it does not really understand what it is talking about.
It is a bit like if I wrote an essay on quantum field theory. I could probably write a convincing essay, especially if I could plagiarise (i.e., copy from the internet). But if a real physicist checked the essay or even interrogated me further, they would discover my answers were full of errors.
It is no wonder ChatGPT and other LLMs do well in standardised tests. Those can be beaten through pattern recognition and plagiarism. Ask any student who had to take online exams at home during the COVID lockdowns. A consequence of this is that ChatGPT can be wrong but extremely convincing. This hallucination is a major challenge for LLM models.
How to Fix:
- The first thing is pair to ChatGPT with human domain experts – that combination is super-powerful. The human saves time on ‘grunt’ work, but checks the output.
- The second is to refine your ‘prompts’ (the questions you ask). Already, we have seen an explosion in suggestions on the best ways of wording prompts to get the right answers. This has created a new job of prompt engineering.
- Third, we can use more trusted datasets than the unknown training set of LLMs to rework the models.
- Finally, the biggest advance will likely come from combining ChatGPT/LLMs with computational/theory engines. That engine would know the laws of physics or maths or accounting identities and would ensure the LLM spits out the ‘right’ answer.
2. Explosion in Fake Content, Fake People, and Fake Sites
Many expect some kind of flourishing of original and creative content thanks to generative AI, whether text-based like ChatGPT or image-based like Midjourney.
But remember when the internet first started?
It was a cesspool of websites of dubious moral content, ugly pop-up windows and malware. The same thing will likely happen with generative AI. The ability to create websites, write content, and create images has just become exponentially easier. Expect an explosion of fake websites, fake emails and fake people.
Anyone can now ask ChatGPT to write copy for travel, finance or any topic you name. This can then be converted to websites, emails and social media posts. And using Midjourney, people will create very realistic fake celebrities to entice the unsuspecting user. You think you will be interacting with someone real only to be defrauded.
How to Fix:
- Do not trust what you read or see unless you know the source.
- Smart companies will be transparent about their use of generative AI and become safe places for people to consume content.
3. You Will Be Hacked
The current rage in the ChatGPT space are so-called chained GPT-4 models like Auto-GPT or BabyAGI. This takes ChatGPT to another level by acting like autonomous agents. This means you can ask it to do something, and it will self-generate prompts until the task is done.
For example, you can ask it to order a pizza. This will set of a chain of prompts – ‘find a local pizza website’, ‘recall what is Bilal’s favourite pizza’, ‘select Bilal’s pizza on the Pizza site’, ‘enter Bilal’s home address’, and ‘enter payment details and confirm’. Before you know it, your pizza has arrived.
The trouble is that the AutoGPT agent will need access to the internet, the permission to impersonate you on the site, access to your PC to get your personal information, and permission to make payments.
You can imagine how this could go very wrong! Someone could easily modify the code in the autonomous agent model to give them backdoor access to Auto-GPT and take over your life.
How to Fix:
- Get someone who knows code to check the AutoGPT or other autonomous agent code taken from GitHub.
- Make sure to maximise the security features on your PC after downloading, so you can see what permissions the agent requests to run.
- Create a ring-fenced sandbox to play with these models.
4. The ChatGPT Tools You See Today Are the Worst Ones
There is a clear lack of imagination in the development of ChatGPT use cases. It is a bit like when moving pictures and film were created – all movies looked like stage plays. People could not get their heads around how the technology could change the way you could show art. The first instinct was to replicate what was already done.
We can say the same about the internet. The first websites of news sites looked like one-for-one copies of the printed newspaper.
The same thing is happening with the avalanche of ChatGPT and GenAI tools you see on social media. They are all staying in the lane of what we are already doing. They are doing existing things, but more efficiently. Create a website, marketing copy or code – do it faster with ChatGPT!
But this is the easy stuff. The revolutionary tools have yet to come out.
How to Fix:
- Don’t get too caught with the daily barrage of ChatGPT tools. These are the low-hanging fruit that will be commoditised very quickly.
- Think about things in your domain that you have not been able to solve. Apply generative AI to those problems.
- Keep an eye on tools in very different business lines and think about combining them.
- Develop interesting and unique frameworks around tools like ChatGPT. Everyone has access to the OpenAI tech stack now, but the untapped potential is within the custom frameworks that can be created around such tools.
5. Loneliness and Mental Illness Will Explode
We already know that people are becoming lonelier and suffering more mental illness. There are all sorts of explanations, from the decline in local community organisations to the increase in economic insecurity. But perhaps the largest factor is the relationship people have built with their screens.
People spend far more time interacting with screens than with real people in real places. It is unsurprising – social media apps from Twitter to TikTok are optimised to grab your attention. But generative AI will take that to another dimension.
Picture this.
A new app gets launched – call it ‘DreamLover’. In the app, you select the celebrity you most like the look of. Immediately, the generative image AI creates a super-realistic movable image of Angelina Jolie, Scarlett Johanssen, Zendaya or whoever you like. Then, using an LLM engine you can start to ask questions to your ‘DreamLover’, and she will respond thanks to the LLM. Throw in text-to-voice AI, and she will start talking to you.
Before you know it, you have your own personal celebrity who you can facetime whenever you want. You thought TikTok was addictive? This could stop all real-world social interactions!
How to fix:
- Do not open that app.
The AI revolution has started, just make sure you are not the first casualty. Good luck!
Bilal Hafeez is the CEO and Editor of Macro Hive. He spent over twenty years doing research at big banks – JPMorgan, Deutsche Bank, and Nomura, where he had various “Global Head” roles and did FX, rates and cross-markets research.
Photo Credit: depositphotos.com