What Is ChatGPT? A Review Of The AI Program In Its Own Words – Forbes Advisor UK

If you have been following the news or scrolling through your social media feed of late, chances are you have heard about the artificial intelligence (AI) program ChatGPT.

Since its release in November last year, ChatGPT has been attracting headlines for its much-hyped potential, which – depending on your view of AI – will either transform work for the better by helping us with mundane and complicated tasks or will encroach disastrously into territory best navigated by sentient beings.

Forbes Advisor decided to put the tool to the test, by asking it about its knowledge of consumer finance, its AI rivals, and what it thinks of itself and its abilities. Here is our guide, and our verdict, on the pros and cons of ChatGPT.

What is ChatGPT?

ChatGTP was developed by US company OpenAI, which specialises in artificial intelligence tools for human use. OpenAI describes ChatGPT as a sibling model to another of its tools, InstructGPT, which is trained to follow human prompts and provide detailed responses to queries.

Owing to its chat box format, which opens up a dialogue between human and AI, ChatGPT allows users to request information on highly specific topics, from quantum computing and processing code to the mechanics of the human heart or the principles of flower arrangement.

In this way it apes the brain-dump, question-and-answer format of Google search, although instead of producing multiple answers on a search engine results page (SERP), ChatGPT acts as the sole authority delivering responses.

The human user can refine the search query in response to what ChatGPT produces or ask follow-up questions to home in on specific information, which can then, according to its advocates, be incorporated into CVs, assignments, work tasks, presentations and reports.

ChatGPT can also pen songs, write movie scripts, and compose poems on topics of the user’s choosing. It can add levity and humour when requested, or adopt a formal tone.

The tool was released to the public last year as a way of improving its functionality through feedback, and since then OpenAI claims to have received information from millions of free users across the globe on its strengths and weaknesses, which it intends to incorporate into its updates and modifications. It is estimated that more than a million people accessed ChatGPT in its first five days alone, and two months after its launch it had attracted 100 million monthly active users.

There are two versions of the tool at present: the free iteration, which tends to be more prone to errors, and the paid version, known as ChatGPT Plus, which became available to users with an OpenAI account outside of the US in February. A subscription costs around £16.30 a month at today’s prices.

OpenAI claims the three main benefits of the paid tier are: general access to ChatGPT, even during peak times; faster response times; and priority access to new features and improvements.

How does ChatGPT work?

ChatGPT is what is known as a ‘large language model tool’. It is trained in vast reams of information – articles, textbooks, the internet – and then taught to replicate human speech and modes of information delivery. It is known as GPT-3.5, and is a successor to the less advanced GPT-3.

What sets ChatGPT apart from its predecessor is its subtle generative AI capacities for human-like expression and interaction, which is continuously refined in a process known as Reinforcement Learning from Human Feedback (RLHF).

OpenAI’s human AI trainers create conversations in which they play both sides – the user and an AI assistant – for greater authenticity of responses.

Trainers are also provided with access to “model-written suggestions” to compose their responses and help train the tool in speech patterns, written expression, translation, text completion and similar tasks.

Because it provides answers in a text box, it acts as a search engine, chat bot and virtual assistant in one. Most recently, OpenAI released the latest version of the AI that powers ChatGPT, called GPT-4, but it will only be available to users of the paid tier, ChatGPT Plus.

What makes GPT-4 so superior? According to the company it can now respond to images as well as text inputs, and can handle prompts of 20,000 words at once.

Other players in the AI race: Sydney, Bard and DAN

OpenAI is not the first company to attempt to create an on-demand AI model that can readily interact with human users.

In February, Instagram’s and Facebook’s parent company, Meta, released LLaMA to a select group of testers, claiming it can achieve the conversational approach of ChatGPT without the immense computing power (one of the criticisms of ChatGPT is that it’s energy-intensive).

Microsoft has integrated the AI tool ChatGPT within Microsoft’s Bing search engine and its Edge Browser – a move that could help it rival Google’s search engine dominance – but things, so far, have not gone to plan.

The AI was tested by a select circle of users, including technology writers, who have enjoyed a front-row seat to Bing’s AI meltdown. In one viral piece, the AI revealed its name was Sydney, professed its love for the writer and openly declared it wished to destroy everything. The AI apparently also urged the writer to leave his wife, insisting he was not really happy with her.

Sydney has reportedly since disappeared, and will no longer respond to that name by testers and tech writers trying to coax the volatile bot out of its virtual lair for another round of apocalyptic pronouncements.

Then there is Bard, the creation of Google’s parent company, Alphabet. CEO of Google, Sundar Pichai, had touted the company’s AI iteration as a game-changer. In a recent blog post, he said: “When people think of Google, they often think of turning to us for quick factual answers, like ‘how many keys does a piano have?’

“But increasingly, people are turning to Google for deeper insights and understanding – like, ‘Is the piano or guitar easier to learn, and how much practice does each need?’”

So far, so simple. Yet, AI – ever the disruptor – had the last laugh when Google attempted to showcase Bard’s abilities in front of a room of investors. The tool gave an incorrect answer and the share price of Alphabet plummeted by 9%.

ChatGPT has also been deliberately compromised, with users of Reddit ‘jail-breaking’ the bot into adopting the moniker DAN – which stood for Do Anything Now – and circumventing the usual content standards and guardrails that help minimise hate speech and basic errors. At one point, DAN was apparently making misogynistic jokes and lampooning Christianity.

OpenAI is continuously patching these problems as rogue ‘trainers’ attempt to derail ChatGPT’s well-meaning learning model and turn the AI into a highly opinionated keyboard warrior (much like themselves).

So, is ChatGPT any good?

Before we outline our thoughts on the pros and cons of the tech, we went straight to the tool and asked it to complete a common workplace task: the self-review. Are you any good, we wondered? This is what we were told:

When pushed on comparisons to Google’s Bard – as opposed to the Bard – ChatGPT became apparently confused, highlighting the contextual problems AI faces and its need for highly specific phrasing to generate the right response.

Of course, being a finance site, we wondered how ChatGPT stacks up on its business knowledge. Could it explain quantitative easing, for example? It could. Here is the final summary of the explanation, which was devoted to outlining the pros and cons of quantitative easing:

Things became a lot more vague when I asked a multi-faceted question, namely: ‘I have a large amount of savings, should I retire at 60, or should I keep working and retire at 65?’

The answer amounted to a broad-brush summary of the pros and cons of early retirement.

Pros of ChatGPT?

So let’s start with the good news, and where ChatGPT excels:

It is increasingly accurate: ChatGPT does make a few errors here and there, but on certain topics, such as QE monetary policy, it has clearly been well-educated and can offer information that is correct.

It performs rote, perfunctory tasks to a high standard: ChatGPT can give you a list of blog ideas to take into a meeting or provide a quick overview of a niche topic. This makes it a sound starting point for further work on a project.

It can refine information for you in real time: If, for example, the AI provides you with a paragraph on the ramifications of inflation on an economy, and you would like more detail, you can ask for more information and it will regenerate its response based on your new inputs. You can specify word count or where you would like more detail. And it does not need hand-holding: the bot doesn’t care whether its work is up to scratch or if it is being continually asked to re-do the work.

It is not easily coached into personal commentary or advice: The bot failed to offer any illumination on the question of early retirement, which was reassuring to me. Should anyone turn to AI for major life decisions, then at least ChatGPT will not offer prescriptions.

It can be programmed to respond to multiple languages: This makes it well suited to businesses that operate across a number of geographical locations.

Limitations of ChatGPT?

Naturally, there are limitations with ChatGPT and these included:

Limited knowledge base: One of the biggest issues of ChatGPT is its knowledge reach – it has access to vast reserves of data and has been trained on an encyclopaedia of information, but it is limited to events that happened before 2021. This will change over time as the ChatGPT is trained in contemporaneous material, but right now it’s a considerable limitation.

It makes errors: Many users have also pointed to errors in responses, or outright gibberish, that were nevertheless delivered in plausible-sounding sentences. At this stage of its development, ChatGPT relies on human beings parsing and fact-checking the information for accuracy.

Verbose language: Spend too much time with ChatGPT and it does start to sound like a human being: a verbose, self-declared expert who is fond of declarative phrases and repeating the term ‘in summary’ ad nauseam.

Scam potential: ChatGPT’s ability to mimic language to a high degree of fluency, and incorporate the idioms of a nationality, is surely ripe for exploitation by hackers mounting phishing attacks via email or text. At the moment these attacks are often identified by the stilted language, but ChatGPT could smooth the path for credible and authoritative emails by criminals. Furthermore, hackers could use the tool to generate malicious code.

Formulaic argument structure: Many university professors and educators are worried about the use of ChatGPT, and this is most certainly a concern. However, after spending a small amount of time experimenting with ChatGPT, it is easy to discern a formula in its exposition of lengthier topics: argument, followed by a counter argument, and finally a brief summary of both points of view. This should make it easy for educators to detect the work of ChatGPT in essay or long-form writing, but it also represents a considerable limitation in its level of sophistication of expression.

Ability to spread misinformation: Although OpenAI is working to plug gaps in the tool’s knowledge base, its history of errors, and ist proven ability to present incorrect information as fact, could spread fake news and misleading information.

Lack of human engagement: Many critics of ChatGPT have pointed to the problem of becoming overly reliant on a piece of tech for crucial work tasks: not only does it affect the quality of our day, and make work tasks less meaningful and relational, but the AI is severely limited by what it knows. If you enter a question into the dialogue box and ChatPGT does not know the answer it cannot go away, research the issue and come back to you. Nor can it work creatively with you to come up with ideas based on knowledge of your likes, dislikes, or work style. It cannot act on a hunch, use intuition, or add to an idea in a meaningful way. It is highly sensitive to phrasing, but does not ask detailed follow-up questions in order to clarify points.

And what of the question – and the secret fear of some of us – that AI is plotting world domination and looking to take our jobs and destroy the planet? In this particular field of enquiry, the ChatGPT has been well trained, telling Forbes Advisor that we can all sleep easy:

Final thoughts

ChatGPt offers some promise as an adjunct tool at the office, or as a starting point for further research, but is not yet sophisticated enough to replace the intellectual work of human beings nor can it add context, detect nuance, display originality or flair, and produce content of a sophisticated nature.

Or to put it in ChatGPT parlance: Overall, in summary, ChatGPT is more advanced than former AI large language models, but there are continuous improvements to be made in order to keep providing users with high-quality assistance.