Data Ethics Club: Ask Me Anything! How ChatGPT Got Hyped Into Being#
Whatâs this?
This is summary of Wednesday 18th December Data Ethics Club discussion, where we spoke and wrote about the article Ask Me Anything! How ChatGPT Got Hyped Into Being by Jascha Bareis. We were fortunate enough to have Jascha join our session to introduce his work and participate in some of our discussions! The summary was written by Jessica Woodgate, who tried to synthesise everyoneâs contributions to this document and the discussion. âWeâ = âsomeone at Data Ethics Clubâ. Huw Day helped with the final edit.
Article Summary#
ChatGPTâs release triggered a social frenzy of transcendental predictions fuelled by sensationalist narratives and hype. Hype is a societal dynamic triggered and driven by strategic agents using figurative language to trigger emotions and craft a community. Actors mechanise hype to create attention, followership, and investments. Hypers can be understood as problematic appropriators of futures, exploiting epistemic uncertainty of the future for opportunist purposes without care for the difference between fact and belief or truth and lie.
Hype of ChatGPT is propagated through several forms including its interface, behaviour of industry leaders, media coverage, and receptive audiences. The design of ChatGPTâs interface anthropomorphises the chatbot in a way that implies self-reflection and empathy, making it difficult for users to manage their expectations. Potential to mislead is further exacerbated by providers of systems failing to make clear that Large Language Models (LLMs) have problems with truth. Powerful actors in the industry cultivate hype by disseminating an apocalyptic narrative that silences political and structural regulatory questions about infrastructure control. Media coverage further feeds hype by depicting industry competition as an exciting spectacle to observe, thereby distancing the audience. The audience themselves are complicit and necessary for the phenomenon of hype to take place. An effective way to defeat hypers is to ignore them or not give them the stage to avoid feeding the attention machine. Public advocates and policy makers must tackle the authority and credibility of knowledge production.
Discussion Summary#
Are hypes around technology driven by single actors or must they be understood as a societal dynamic? If yes, who/what drives it?#
At the outset, we werenât sure how weâd define hype and found that the normative dimension of hype discussed in the paper wasnât clear. We conceive hype as a phenomenon that warps narratives and has a spectrum of credibility but goes further than what it is calibrated for. Success stories are told, and failures are not. Hype encompasses both promotion and counter-narratives, such as government lawsuits and policy consultations.
Looking towards Silicon Valley specifically, hype cycles have very precise triggers which the article doesnât really touch on. However, these triggers are helpful for defining hype and understanding how people get on board. The hype cycle has existed in Gartner for a long time, and technologists use it in their day to day work. There are criticisms of the Garner hype cycle, however. Hype is thus not a new thing, but the scale and rapid increase in the phenomenon is something to be wary of.
Hype has a mutually reinforcing snowball effect, as when other parties are onboard things have the appearance of being derisked. General purpose technologies are driven by capitalism; itâs almost as if the more people get on board, the more hype is manifested. From a sales perspective, the principles of persuasion denote that scarcity and conformity make things more appealing to people. Initially, it was difficult to log into ChatGPT as it was trying to operate so far beyond its capacity; this only enhanced an air of exclusivity. Just like the queue for at the queenâs funeral, the more people that talk about it, the more people get involved - people donât want to miss out. There is an element of game theory here, reflecting a dichotomy between wisdom and foolishness of the crowd. Exaggerating capabilities with hype makes people willing to pay more than what the market rate would reward.
We wouldnât be surprised if it turns out that generative artifical intelligence (AI) doesnât work for many use cases. The situation may be similar to blockchain, where the problem wasnât a lack of use cases but too much complexity. Hypers themselves have a range of competencies in their knowledge about technology. ChatGPT wasnât a huge technological breakthrough, it just presented well with a very effective and engaging user interface. Depending on how you define intelligence (e.g. capacity for self-awareness, creativity, and understanding), LLMs arenât necessarily intelligent or really AI at all; it is the human that is inputting the intelligence. The misinformation saying that LLMs are AI is another way of showing how LLMs have been hyped.
ChatGPT is used by a lot of people, but not used well; many people just throw everything into it. We feel that without studying or understanding how LLMs work (e.g. how hallucinations occur) itâs difficult to understand how to use them well. Rather than explaining how the technology works, hype further mystifies it. Many people believe that AI is ChatGPT. We have seen LLMs integrated into the teaching and assessments of certain university courses and wonder why the reaction to LLMs (i.e. the eagerness to include them) is different to past tools like Google Scholar. We also wonder why ChatGPT is the LLM that everyone talks about, as people were using LLMs before. The answer to both of these questions is hype.
When ChatGPT came out, it was a shock that the public started using it. There is a reason it is called ChatGPT and not ChatAI; the company really did see it as a demo and not something that would kick off. Silicon Valley has a track record of releasing what seems to be a âtoyâ without proper consideration of unintended consequences. Product-market fit involves finding an appropriate market where the product will fit, and is a good indicator of when the public will pick up on something.
Humans are highly driven by convenience and laziness, and the innovations that win are generally those that make tasks easier for people. ChatGPT is very accessible to people, which is perhaps why the hype is so strong. It embodies the promise of general purpose technologies such as the internet and smartphones, which further drives the hype. Perhaps it would be better to focus on appropriate use cases. The right use cases for AI are probably jobs that are tedious and boring. In healthcare, for example, convenience would be useful for data gathering, cleaning, and access. However, for summarising papers we would prefer to just read papers ourselves instead of getting a LLM to do it because of issues like hallucinations.
Are hypers always opportunists and âbullshittersâ or arenât there also firm believers in their cause?#
This paper made us angry because it systematically laid out the many powerful people and institutions that repeatedly and brazenly lied in order to pump hype and thereby their profits. Venture Capital Hype cycles are different to how the media and public pick things up and hype them, as they have different motivations behind them. Products become of value because people invest money. If you want funding, you jump on the bandwagon and say what youâre doing is AI whether or not it actually is; even if itâs just algorithmic.
There are lots of different stakeholders in the AI hype cycle, but we never really define them. The names highlighted as key players in the paper were not necessarily the ones we were expecting, such as Sam Bankman-Fried who may be on the list because of his connection to effective altruism. Hypers could be a mixture of people involved for âgoodâ and people that actually want to do good. Some people at the top - we wondered how many - do think that AI is a solution to the worldâs problems. A lot of industry leaders would say anything (no matter how obviously false) if it would increase profits, but perhaps it is the true believers that are the most dangerous.
The paper ChatGPT is bullsh*t, previously discussed at Data Ethics Club, also fuelled hype with its sensationalist title. Bullshitting is a social action which implies awareness and intentionality so in a sense the paper is anthropomorphising the tool. Although, the paper does distinguish between the tool itself and the toolâs creators or users.
At Data Ethics Club, we are not innocent of participating in the generative AI hype cycle. By discussing and dissecting it, we are also giving oxygen to the hype. However, someone has to push back against hype and simply not talking about AI might not be the answer. Whilst hype has some impact on subduing regulatory questions about infrastructure control, it may also have the opposite effect of drawing attention and putting additional pressure on regulatory bodies. We need more stories of forgotten technology which didnât turn out as expected.
Sometimes, issues of misunderstanding arise from using different language between marketing, ethics, and technology teams. There needs to be clarity within teams regarding what AI is capable of doing now, in the near future, and in the long term. Trust and transparency need to be clear, but âtransparencyâ and âtrustâ scare away money. However, if youâre investors, being clear about where youâre putting your money is actually advantageous.
Who decides that society is experiencing a hype or rather a promising future trajectory?#
Validation of whether a tool is real or actually works tends to lag behind the hype bubble. Limitations of many LLMs were intentionally hidden by liars interested in inflating the hype bubble rather than releasing a working product. This is the way that startups operate: the leaders are salesmen (usually men), and developers have to try and produce what was sold. To sell a product, the technology does not have to exist in the real world and if you donât have the right data, you can create synthetic data.
To understand the actual capabilities of a tool, it is important to listen to stories from the coalface, the place from which the real work is being done. Founders may have boundless enthusiasm and belief, but developers often soon realise that the emperor has no clothes. Success comes in when developers are able to convince the client that what they were sold is the same as what has been delivered.
Hype can be identified as it happens rather than retroactively by examining the sorts of words used in overheated discourse. For example, the use of apocalyptic vocabulary would indicate the existence of a hype bubble. It is unclear what will happen over the next year with the AI hype bubble and global politics generally. In the discussion, it was good to hear the authorâs perspective about the future and the news regarding Elon Musk attempting to give the Reform UK party a massive donation.
What change would you like to see on the basis of this piece? Who has the power to make that change?#
We wonder how LLMs and hype will affect education: how we will identify cheating, how we can monitor the use of ChatGPT, and the implications ChatGPT has for trust. Skills atrophy may occur, making things like coding too easy and dumbing us all down. Online exams and tests will also be affected; we took a test during which we had to lockdown our browser to one tab, which made the experience uncomfortable and added a level of stress.
Opacity in LLMs is problematic as they are black boxes with limited transparency regarding how training data is controlled and what has been added to it. If you put a dataset into ChatGPT, it is not clear who sees that data or if it could become public by the company having access to it and regurgitating it in a later output. We werenât sure if people even realise it is a possibility that their inputs can become outputs. Training LLMs from data siphoned off the internet makes LLMs cannibalistic as outputs go onto the internet, to become training data, to become outputs again. Things that arenât true become fact. Corruption of the internet is happening rapidly; a project analysing the usage of human language had to shut down because so much of the data has been polluted by generative AI. Eventually, the internet may be full of AI generated content.
Attendees#
Huw Day, Data Scientist, Jean Golding Institute, https://www.linkedin.com/in/huw-day/
Euan Bennet, Lecturer, University of Glasgow
Amy Joint, ISRCTN Clinical Study Registry, https://www.linkedin.com/in/amyjoint/
Jessica Woodgate, Ph.D. Student, University of Bristol
Ola Michalec - Lecturer @ Bristol Digital Futures Inst
Paul Lee, investor
Kamilla Wells, Citizen Developer, Australian Public Service, Brisbane
Michelle Venetucci, PhD Candidate, Yale University
Hessam Hessami Data scientist https://www.linkedin.com/in/hessam-hessami-ph-d-6723a2b9/
Philippa Davies, Research Fellow in Evidence Synthesis, University of Bristol. Also part-time Data Science MSc student. First time joining today
Zoë Turner, Senior Data Scientist, NHS @Lextuga007
Dan Lawson, Director of Jean Golding Institute, https://people.maths.bris.ac.uk/~madjl/
Nevena Nikolajevic, Data Projects & Data Consultation at Civic Data Lab https://civic-data.de & CorrelAid https://www.correlaid.org/en/