Who will win the generative AI race and what does it mean for client listening? Paul Roberts of MyCustomerLens looks for answers.
The marketing world is littered with examples of corporate battles being waged between rival providers, seeking market dominance with new technologies.
In such cases there can usually only be one winner and it isn’t always the challenger expected to triumph. Sometimes there is no winner, as the heat of battle clouds judgment, leading to corporate hubris and poor decision-making.
Who remembers VHS v Betamax, BBC Breakfast Time v TV AM and SKY v British Satellite Broadcasting (BSB)?
In the fight-to-the-death for satellite TV dominance, Rupert Murdoch’s Sky bought and subsumed Robert Maxwell’s BSB to become BskyB. Anyone who kept their little-used Betamax video recorder as a dusty museum piece in their attic will know who prevailed in that particular battle of technologies.
The world of artificial intelligence (AI) is currently facing a pivotal moment in its development, with rival behemoths seeking market supremacy in the brave new world of the Super Chatbot.
Fastest out of the blocks for this new generation of generative AI has been ChatGPT, used by one million people in its launch week last December, and by 100 million users in January.
The early success of ChatGPT has sounded the starting gun on a corporate battle for dominance of the new technology involving two of the world’s biggest digital beasts, Google, and Microsoft.
What is currently unclear – and will only become apparent through user experience – is which, if any, of the three will emerge as the equivalent of VHS and which will be Betamax?
Perhaps all will be successful, with a proportionate share of the market. Perhaps the setbacks of the technology will outnumber the advantages and they will all be superseded by a more advanced version of AI. The reality is that, at present, we don’t know.
Marketers rushing to take advantage of the new technology, amid fear of being left behind, may well be putting their own positions, and the positions of their firms, at risk by adopting unreliable and comparatively untried models.
There are more established versions of AI, including natural language processing, which may not command as many column inches among the tech commentariat, but which are more robust and proven to work.
AI and client feedback
When seeking to analyse customer feedback and create client intelligence, for example, natural language processing is currently more reliable.
Drawing on mainly closed data sources, it can be used to reliably identify themes and trends and to create heat maps and mood charts which record and map how clients and their services are viewed.
Generative AI, in contrast, works on open data sources and is programmed to generate new content that may well be inaccurate or outdated and which may create issues around copyright, data privacy and algorithmic bias.
ChatGPT is powered by GPT-3.5, its parent company OpenAI’s upgraded version of its GPT-3 language model, which draws most of its source material from a common crawl of web content since 2008, including from web pages from outbound Reddit links, internet-based books and from Wikipedia.
Google’s LaMDA (Language Model for Dialogue Applications) has been in development for several years and was trained on billions of records of documents, dialogues, and utterances, with 137 billion parameters, 2.81 trillion sentences, and 1.56 trillion words.
It also draws from external sources, allowing it to read a sentence or paragraph, understand how the words relate to each other, and then to predict what words come next with logic and specificity.
Meta’s Blender Bot 3 is designed to make chatbots feel and sound more human, by improving the quality of conversational AI. It searches for answers live online, along with its database and chat histories, containing 175 billion parameters.
Businesses using, or considering using, generative AI for marketing purposes – where issues of accuracy, tone and relevance are critical – should be aware of its restrictions and vulnerabilities.
It is important to stress that, while these generative Super Chat Bots demonstrate an appearance of reality, none is completely reliable.
So much so that Sam Altman, CEO of OpenAI, has warned businesses not to rely on ChatGPT for anything at this stage, as further development is needed for it to become more trustworthy.
Because models have been trained on very large data sources, they are not industry-specific and therefore, when they look for themes around brand and what clients are talking about, they might be making sense of information using a lens that is irrelevant or contradictory to the needs of a business marketer.
Most open source data, relevant to businesses, are B2C customer comments on social media and Amazon reviews which would not be relevant, for example, to professional services marketers, seeking insights that would be better acquired through a B2B lens.
One of the most frequent criticisms of generative AI is that it ‘makes things up’. The Super Chatbots are prone to ‘hallucinations’ and may misrepresent a product, service or company in marketing copy.
Because it is generative, it is predisposed to create, even if it lacks sufficient source material to provide an answer that is accurate and informed. It will just generate something that sounds plausible, even if it has little or no basis in reality.
Marketers seeking to analyse their brand, analyse feedback, or perform a competitor analysis risk being presented with information that isn’t true.
In the context of customer feedback, a generative AI model may perceive that a lot of people are talking about personalisation, and so it will produce what it perceives to be a typical quote from someone talking about personalisation. It may appear to be real, but it can be an invention, made up to approximate what sounds plausible.
Currently, the best way to apply AI to feedback analysis is to use more established techniques such as natural language processing (NLP). This robust methodology is used to consistently classify unstructured text comments.
NLP can quickly classify what topics the feedback is covering, the sentiment of each comment and how well the feedback links to broader brand and service themes. Crucially, these machine learning tasks apply the same logic and process to every piece of analysis. So, it can be trusted to identify the relative significance of different topics and to highlight when certain themes or attributes aren’t being mentioned.
Established AI tools like NLP and machine learning may not have the same media attraction as LLMs and generative AI, but that doesn’t make them any less robust, or fit-for-purpose. On the contrary, they remain the best way to get predictable and relevant summaries of customer feedback.
Anyone who remembers taking their rented VHS films back to Blockbusters will appreciate that all technologies are, by their nature, transitory and that rushing to join the bandwagon is not always the smartest approach.
Paul Roberts is CEO of MyCustomerLens, an AI-driven, always-on client listening platform for professional services firms
