What can you (actually) do with ChatGPT?

We have all heard of ChatGPT by now – and probably all wondered why they chose such a tricky name. Overworked students with looming essay deadlines are loving it, and schools and universities are having to rethink how they set and score course work. But what about the rest of us? Especially marketers and people in the communication industry. What will we actually be able to do with it that will make our lives better and add value for our customers?

We answer all these questions and many more, and before you ask, a human wrote this piece not AI!

Commentators have noted that, unlike Siri or Alexa, the name doesn’t exactly roll off the tongue. However, they didn’t pluck it out of the air. GPT stands for generatively pre-trained transformer, which is the name for the architecture which underpins ChatGPT – basically a set of algorithms which when united, can generate new and original content by transforming existing content – whether that be text, or in the future – images, audio, and even video.

Images you say? Yep, the technology can create them too. Try it for yourself – we asked OpenAI’s DALL.E tool to create ‘a baroque portrait of a bear wearing a crown’ and this is what it came up with:

As far as we can see, this regal fellow is a completely original image that doesn’t exist anywhere else on the web.

That almost seems like magic, so how does it do it? Without getting too technical, here is a very high-level explanation of how ChatGPT got to be so smart.

There are two things that ChatGPT is really good at. Firstly, understanding what it is being asked to do, and secondly returning a result very quickly in the required format and style. Note that we didn’t say ‘accurate result’ but we will come back to that point later.

Google can easily answer the question “What is irony?” but it can’t provide a definition in Shakespearean language, or in the style of a country song, or spray painted on the side of the Taj Mahal. Those requests may seem frivolous, but the point is that ChatGPT can do a very good job of creating something completely new that has never have existed before – while search engines like Google simply help you find content that someone else has already created.

Language is a complex, diverse, and dynamic construct – the English language for example has almost 200,000 words, with up to 1,000 new ones being added to the Oxford English dictionary every year (if you think that is a lot, Arabic has 12 million distinct words!)

Even a question as simple at ‘What is irony?’ could be phrased in several different ways.

GPT models use something called large language modelling to analyze inputs and predict the next word in a question so that the AI can start generating a response even as the person is still typing. With every additional word that is typed, the model refines its results so that it can respond extremely quickly when the person clicks send.

Take a simple question like ‘How do you make toast?’ By the time the person has finished typing ‘make’ then the algorithms have already discounted vast amounts of possible answers and predicted the probability of what the next word will be. Maybe something like:

What about asking “How to make scrambled eggs?”. In this case, before the last word is typed then the model should know with a very high degree of certainty what is being asked.

This is the key idea behind all GPT models, they all use generative pre-training to help them perform so well.

Unlike standard chatbots which are trained to understand specific words and phrases, ChatGPT and similar AI is trained how to learn from any resource it is given access to, including ‘the entire internet’. That is a colossal amount of information, but as we all know you can’t believe everything that you read online so the AI is helped by humans in a process called reinforcement learning with human feedback.

It is during this phase that humans assess the output that the models produce and provide feedback about the quality and ‘humanness’ of the content.

Through this continual feedback, ChatGPT learns human preferences and is fine-tuned to be able to mimic those preferences. This process does however introduce the possibility of human bias, so this is a great segue into some of the weaknesses of ChatGPT.

Human bias: Humans can influence AI in subtle and often unnoticed ways. Even by selecting the data that is used to train AI, the human is already influencing the machine’s view of the world by applying a human lens to what is deemed ‘suitable’. Of course, each individual bit of data and content that the AI scrapes will reflect the biases of the creator. You would hope that by using a very large number of sources, the AI would settle into a neutral world view. However, if you consider how much societal norms and values have evolved in the past thirty years, there is still a huge proportion of older text and data that reflects values that would now be viewed as outdated at best, and possibly sexist and discriminatory.

Inaccuracy: There has been a lot of press about the factual accuracy of the content that ChatGPT produces. Just like Google, the AI is looking to the internet for answers. Unlike Google, which by its design shows the source of the information it displays so that the user can make an informed decision about its reliability, ChatGPT scrapes content from all and any source and then transforms it to form a response. Depending on the topic, some or even all of the content may be from unreliable sources.

In the area of world news and politics for example, there are a number of parody news sites like The Onion and the Borowitz Report subpage of the New Yorker that even humans sometimes mistake for genuine news sources.

OpenAI itself admitted that ChatGPT “sometimes writes plausible-sounding but incorrect or nonsensical answers“. However, it points out that this is common to all large language models and is a phenomenon researchers call hallucination.

Another significant issue is that at the moment ChatGPT’s knowledge is about two years out of date. This blind spot will no doubt be reduced before ChatGPT is released formally, but in many areas it is stuck in 2021. For example, ask it who the current UK Prime Minister is, and it will throw up Boris Johnson – which given the current turmoil in UK politics may turn out to be factually correct again in a month or two.

Legal conundrums: Just like the legal wrangles related to the rise of NFTs and copyright infringement in metaverse worlds, ChatGPT is starting to generate a lot of work for intellectual property lawyers. As a technology that relies on existing content to generate its output, some or all of which may be copyrighted, there is now a huge gap in legislation to help determine what is and isn’t a copyright or IP infringement.

Just a couple of examples of current court cases include:

It’s terrible at humor: Ask ChatGPT to tell you a joke about an elephant and it will do an OK job, although be prepared for grade school level of humor. What it struggles with is more sophisticated and nuanced humor, sarcasm, and parody that relies on the audience having knowledge of the subject that is being satirized.

As we have mentioned, ChatGPT can even get confused and present parody content as fact, which may make it well suited to political speech writing.

For the time being at least, we think the jokes should be left to us humans.

Of course it is, just ask all those C-grade students suddenly acing their essays. It’s not nearly perfect, but it is a huge step forward for publicly accessible AI. Technologies like ChatGPT will change so many facets of our lives that we can’t even begin to comprehend.

Most of us already use Google search every day of our lives, and many of us have daily interactions like “Alexa, turn the heating up.” or “Siri, play relaxing music” but these aren’t exactly dynamic exchanges comparable to a conversation with a human friend that flows naturally from one subject to the next.

Even chatbots still struggle in this area, unless they have had extensive training. They are usually designed to perform a specific function, for example ‘helping customers to renew their car insurance policy’. But they can flounder when the person starts asking about subjects that may seem naturally related to a human mind, but that the chatbot has no knowledge of. For example, ‘car modifications that could invalidate the car insurance policy‘.

With ‘the entire internet’ as a point of reference, including the T&Cs of hundreds of car insurance policies, ChatGPT could shift effortlessly from the costs of insuring a 2018 Corvette Stingray, to what effect adding aftermarket alloy wheels would have on monthly premiums.

With the development of accessible chatbot building tools and conversational platforms that enable businesses to communicate with their customers as individuals, in the near future these solutions will incorporate technology like ChatGPT to offer customers a completely personalized and all-knowing conversational interface that reflects the unique personality and values of the brand.

Important when predicting the future of ChatGPT is to remember that it is owned by OpenAI, which is a privately owned company with a valuation of about $29 billion (double what it was before the release of ChatGPT). With close ties to Microsoft, and not being an academic or philanthropic organization, they will undoubtably be looking to achieve widespread adoption and then monetize it, just like Google and Facebook did with their services.

This has drawn criticism from various industry heavyweights; even Elon Musk who was initially involved with OpenAI has recently criticized it for abandoning its original goal of being a free and open-source “counterweight to Google” to become a “a closed source, maximum-profit company effectively controlled by Microsoft.”

How they will cash in is unclear at this stage, but companies investing in solutions and products that piggyback on ChatGPT technology should bear this in mind as something that is free now, may not be in the future. That said, it would be unwise to ignore what is very likely to be a paradigm shift in the widespread adoption of AI based on large language modeling.

Large language models can be used to solve a huge variety of problems, but not all of them may be relevant for businesses in the marketing and communication space. So, we put our heads together and came up with some use-cases that we think will add value to our partners and customers when a commercially available version of ChatGPT becomes available.

As many people have had a chance to see for themselves, ChatGPT is very effective at answering questions in a descriptive and detailed way across all the domains for which reliable training data is available. Once integrated properly into search engines, it will no doubt become the standard way of researching information, but also compiling the information for whatever purpose it is required.

This integration should also address some of the concerns about ChatGPT’s accuracy and source verification, with vendors like Microsoft and Google already having sophisticated source verification capability.

Imagine a contact center solution where agents and service chatbots could instantly access both online information and summaries of all relevant historic calls and sessions from across the entire contact center. By scraping call transcripts, ChatGPT could quickly identify if an issue had been reported before and how it was solved, even where the agent was in another office in another country, and even communicating in another language.

I’m not going to do myself out of a job here and suggest that you should get ChatGPT to write all your content – that would get tedious for your customers to read and expose you to the risk of inaccuracies associated with factual hallucination.

However, it can be used to help with more repetitive tasks, like quickly producing bootstrap marketing campaigns, writing product descriptions, and creating abstracts and teasers for your human generated content to amplify it on social channels. As long as they are all checked by a human before launch, it should help to speed up the process and make your marketing team more productive.

Some content marketing experts have pointed out that if ChatGPT does achieve widespread adoption for content creation, then it will end up scraping content that it has written itself. Just like a photocopy of a photocopy reduces in clarity and contrast, so machine generated content will become blander and more unoriginal over time. We think there will always be a place for human creativity in the world of content creation.

Not enough time in the day to read all the information you require to do your job or prepare for a big presentation? Let ChatGPT summarize the main points for you, so it looks like you have been up all night preparing. There is a huge potential for this use case with no viable competing technology for this task.

This can be extended to summarizing entire legal texts, complex technical documents, and picking important details from long meeting transcripts, or email trails. If ChatGPT can scrape it, then it will be able to summarize it.

Future versions of the tool will likely also be able to process audio and video content, which opens up a whole new set of use cases from law enforcement using it to automatically scan CCTV footage, to slightly more Orwellian monitoring of online conversations, for example in the online gaming community to enforce site regulations and even identify commercial promotion opportunities.

International companies have a major overhead maintaining website, marketing, and product collateral that is appropriate in style and tone for each of the regions that they operate in.

If you consider the approach to marketing material in the US compared to Germany or Austria for example, Americans may appreciate a more informal and conversational style of writing, while in northern Europe this type of text is expected to be more formal and academic, with an emphasis on facts and statistics.

ChatGPT can be used to very quickly update the style and attributes of large amounts of content to suit each market, or simply switch the spelling and grammar from American to British English.

This is welcome news for the content teams of most international companies, us included.

Audience segmentation helps marketers to vastly improve the success of their campaigns by targeting people that are most likely to be interested in the offer they are presenting.

Most email and campaign automation tools now include functionality that enable marketing users to build audiences by doing quite sophisticated data segmentation via an easy-to-use interface.

Select all customers in [Europe] who have bought [Product X] in the last [6 months] but who haven’t bought [Product Y].

This enables marketers to be self-sufficient when building their campaign audiences – with no reliance on expensive data analysts or the IT department.

Tools like ChatGPT can democratize this process even further by allowing anyone to perform basic segmentation and data analytics, just by asking questions in a conversational way.

With the Eurovision song contest coming up, we thought we would try this out – to prove our point and so that we could impress our friends with our encyclopedic knowledge.

Many popular word processing tools use prompts to help users when writing a sentence by suggesting words and phrases that they may want to use. ChatGPT could take this to a whole new level by proposing entire paragraphs, suggesting associated images, or prompting when a key bit of information is missing.

Writing a letter to your energy supplier to complain about a billing mistake? The AI could create the entire letter in the optimal way to achieve the result you are looking for.

For customer service teams ChatGPT could help agents by automatically compiling responses to customer queries, populated with any key details from the customer’s unique data and conversation history. Taking it a step further, with ChatGPT integrated into our chatbot building platform Answers and our cloud contact center solution Conversations then you could build customer service chatbots trained not only on your company’s proprietary data, but also on all the wider contextual and industry data required to provide real conversational support.

Many of Infobip’s customers have built WhatsApp chatbots for all sorts of useful purposes. Everything from driving sales like Unilever did or ramping up customer engagement like Cardeko did.

Recently, a lot of businesses have expressed an interest in connecting WhatsApp to ChatGPT so that their customers can interact directly with the AI to find out more about their products and services. This will definitely be possible in the near future, and we are currently working on some exciting integration options and groundbreaking use cases. At the moment though, the patchy service availability of the beta version and the issues with inaccuracies and hallucination mean that it is not a viable option just yet.

All that will change soon though. Microsoft has confirmed that it will be incorporating OpenAI technology into its own products, browsers and search engine Bing. According to CNBC it also plans to enable companies, institutions and government agencies to build their own chatbots based on ChatGPT but with their own branding.

This is an exciting time in the world of conversational AI, and we are already preparing to help our partners and customers to benefit. If you have a use case in mind, we would love to work with you to make it happen and get it to market as soon as ChatGPT becomes generally available.