Two years after the ChatGPT explosion
Two years ago OpenAI released ChatGPT 3.5 and kicked off a global generative AI frenzy. Today I’d like to share some reflections on that two year period.
(My first blog post was a couple of weeks after the launch. We hosted several Future Trends Forum sessions at that time, too: 1, 2, 3.)
As of this writing, ChatGPT still seems to dominate the field, despite major competition from giant American technology companies (Meta, Google, Microsoft), intensive Chinese research and development, a raft of start-ups around the world, and a bunch of open source initiatives. Readers will recall that OpenAI keeps iterating ChatGPT, which now has a series of models.
How do people use ChatGPT and other generative AI tools now? It’s easy to determine potential uses and to identify available functions. From everything I’ve seen the leading purpose is making stuff: text in particular, along with images (see above), computer code, and audio (music, voice), with video coming online, albeit mostly not for consumers yet. A second purpose is conversation of all sorts, from informal chat to therapy and romance. AI companions continue to serve demand. A third function, one seriously underappreciated, is as simulation host (one of my posts on AI for simulations) (and a blog post on it). A fourth purpose is rising: AI-backed Web search. Now, to generate this output we now have some degree of multimodal input options beyond text: voice and images now available.
In all of that usage we see a major yet subtle shift in how we can use information. The pre-ChatGPT world largely turned on document production, publication, and search. We Googled to find stuff, and the stuff was largely publications: web pages, pdfs, Word documents, videos, and so on. That echoed decades or centuries of information practice, depending on literacy and culture, where we did something similar with analog technologies: producing books, scrolling a radio to find programs, watching stage plays. Generative AI changes the game, as search is at best a minor function. Instead, tools like ChatGPT produce content on demand. This is why prompt engineering is a vital practice, as the experience of interacting with information is now different in this way. (more on this point)
These two years of rapid development and interest occurred with a massive wave of hype, and saying so is now apparently mandatory if not a cliche. Mainstream media went wild for ChatGPT and its competitors in a way perhaps explained in part by the automation threat now touching media jobs. Something similar happened across social media. Beyond media discourse came a torrent of investment as financial and technology firms poured billions of dollars into building up generative AI applications and companies. Businesses have launched project after project (remember Google Bard?) while also adding AI to preexisting services: Google Apps, Microsoft Office, LinkedIn posts and messaging, MailChimp campaign creation, etc. So far these two years of frantic investment has not yielded a clear business model, although we’ve seen some companies that provide AI in addition to many other functions doing well financially, despite repeated claims of a bubble (for example): Microsoft, Google, NVIDIA.
While this wave of hype poured forth a countervailing movement also pulsed for the past two years. I’ve called it counter-hype, although the British expression “techlash” can serve as well in a general sense. Critic after critic has described many issues with generative AI to various degrees of influence and prominence. Perhaps the most visible during the first year of generative AI’s boom time has been a TIME magazine article calling for militaries to bomb AI servers. Criticism has now settled into identifying a series of problems: poor quality output (hallucinations or slop; which term you use flags a political stance for some); results biased across multiple axes (gender, race, nationality); unfair labor practices, notably for those paid to survey output for problematic content; threats to the creative economy; deepfakes; environmental impacts (CO2 emissions, fresh water usage). Some criticism aligns with contemporary progressive politics with the latter’s emphases on identity, economic exploitation, and climate change. They appear in various forms, such as academic public scholarship. (This morning I received the latest from Brian Merchant, who starts by calling out the lack of two year anniversary celebrations.)
Beyond these criticisms we’ve seen a series of lawsuits, usually alleging generative AI training with massive copyright infringement. As I’ve said, any judge involved in these cases could suspend a company or project’s AI operations. In addition to courts we’ve seen various governments attempting laws and policies to regulate this new technology, from trying to control deepfakes to warding off labor problems. Outside of governments and courtrooms we’ve seen pushback in public attitudes, which polling reveals to contain a great deal of skepticism and anxiety. Part of those responses involve a major, bipartisan fear of professional work degradation and loss, to some degree rooted in generations of industrialism’s creation and destruction of jobs. Those views connect with pop culture, which has been increasingly featuring the techbro as villain. Several battles over AI waged in pop culture haven’t led to triumphs for the technology, such as the Hollywood writers’ strike or the feud between OpenAI and the actress Scarlett Johansson over the apparent use of her voice. Overall, I think AI counter-hype is well established. Indeed, it might outlast the hype, especially when we experience AI economic of cultural downturns. (Some notes on recent counter-hype)
Let me turn from this macro picture to several smaller domains.
Hardware Generative AI is a thing of software, but robotics have been busily evolving these past two years. Connecting the two makes intuitive sense and we’ve seen several efforts to implant generative AI into hardware. None have really taken off yet, although AI in cars seems to be continuing.
Generative AI in politics The past two years saw a great deal of concern about AI in politics. There were fears that political campaigns would use generative AI to produce deepfakes which would mislead voters. So far this has largely not occurred. The United States 2024 presidential election proceeded without compromise, although some of us abhor the results. I have been wondering about militaries and intelligence agencies using generative AI to develop plans; so far, and unsurprisingly, there haven’t been many stories of this.
Medical uses There have been many efforts to bring generative AI to the health care world. The reasons should be evident, as medicine and public health each involve vast amounts of complex data to analyze. In America we have the great problem of extreme costs and, ah, uneven access. For the technology world it’s a reflex to bring tech to bear on these problems, and this has been happening before ChatGPT 3.5 appeared. It’s important to note that most efforts have been very, very careful, given the lives at stake as well as very strong laws, such as the United States Health Insurance Portability and Accountability Act (1996) (HIPAA). I’ve heard from medical professionals that patients already speak of using ChatGPT to get health care advice, which means a competitor for or addition to “Dr. Google.” At the research side we’ve seen a growing body of studies trying to figure out where generative AI might be as good as or better than human professionals.
Education I’ve been writing, making videos, and speaking about this every week for the past two years, so here I’ll try to summarize what’s been happening. Rarely does a college or university actually impose AI on its workers and students. Instead, most campuses lack AI strategy or even basic policies, deferring decisions to academic units (think IT, educational technology, library) and academic departments. One side effect of this is that adjunct faculty - the worst paid, lacking in support and status, and also constituting the largest population of the American professoriate - might be the AI deciders, at least in the United States.
I’ve also seen counter-hype in the academy, especially in the US. I read and hear criticisms from faculty, especially in the humanities, the writing and composition field in particular. There are calls for instructors to refuse AI use, as well as recommendations for teaching criticism of AI. I’ve heard some say that their administration is imposing AI on campuses, or at least not blocking it.
How are academics teaching with AI? Answering this decently would require several posts, but for now I can quickly hit the top lines. Overall, there’s a huge amount of invention and creative pedagogy as professors try out different approaches and experiments. Some include generative AI work in class in various ways, such as having students collaborate in having a bit create something so they can workshop and learn from it. Some faculty teach AI more critically, including the aforementioned challenges within the curriculum, depending on the class subject. Others basically teach AI 101: what it is, how to use it, prompt engineering. I’ve also seen people teaching more hands-on stuff, like GPT creation, as well as my favorite, simulation exercises.
How many students use generative AI and how, two years in? Here I confess some frustration, as there seems to be little reliable data on this question. Most studies and reports I’ve seen rest on self-reporting, which is obviously problematic. Students have incentives to understate their use (if the tech looks sus, maybe I shouldn’t say I rely on it) and also to overstate things (they think I’m a young tech whiz, and I want to look cool/employable, so sure, I’ve definitely become a power user). (Self-reporting is also an issue for faculty and staff use, albeit for somewhat different reasons.) The best approximation I can make now is that some significant proportion of the student population is using AI, and more still have access to it.
Several problems loom over educational use. Perhaps the most salient is the cheating issue, which is still unresolved as of this writing. Generative AI detectors are at best badly unreliable. Reports indicate that OpenAI and Google have watermark capabilities which they could activate, but haven’t done so yet, and might not help much in any case. I’ve heard stories of faculty trying all kinds of alternative assessments, like using airgapped machines. Some turn to older assessments, perhaps with a degree of retro pride: oral exams, even written tests. The only solution I can see now, if generative AI persists, is to rethink assessment as a whole. But that’s an enormous ask for an academy often suffering from overwork and for faculty and staff already overburdened, not to mention the adjuncts aforementioned.
Related to this problem is the argument that students won’t do enough cognitive work when they learn with AI. Think of a pre-2022 learner developing their writing, paper after paper, revision after revision. ChatGPT et al can smooth that process, but the student won’t actually develop that internal skill. Or consider learners using bot summaries of difficult readings, like Cliff Notes on steroids.
Another looming education and AI issue is strategic, if less photogenic than academic integrity. A great deal of AI use is personal or consumer-based: logging into ChatGPT with personal credentials, paying for a personal Perplexity pro account. Yet academic institutions often prefer to have enterprise plans for technology, and here there are all kinds of problems. Chief information officers (CIOs) tell me they’re getting price quotes which would blast open their budgets. And there’s the other issue of relying on third parties who can change service offerings or go out of business. I actually posted two videos on this point:
An open question for me: are academic degrees declining in value? That is, will graduate schools, employers, and the general public view bachelors’ and associates’ with skepticism, to the extent they think a number of students either cheat or use AI in a way which reduces their learning?
So, two years have passed since ChatGPT 3.5 changed the world. It’s a major movement in the technology sphere, having significant impacts in the rest of the world. I hope these historical notes give us something to consider when we look ahead.
Are you seeing these developments, or others we missed today?
I think the developments in AI cannot be evaluated independently from other developments such as democratisation of knowledge, transactions, trust, communication, instruction. I think that the educational solution or way forward lies in the integration of all these affordances as a system.