AI on the brink

A radical problem for academics thinking about the technology

How can higher education best respond to new developments in AI? How do we respond if the current AI wave collapses?

Exploring the first question has been the theme of this Substack since I launched it. (And I want to apologize for the lack of recent posts here; I explained the reason on my blog.). Today I wanted to step back to look at the second one.

Will generative AI be a major force to contend with in the near and medium term futures, or will it rapidly dwindle into marginal insignificance? Will it becomes something vast, on the scale of the World Wide Web in the 1990s, or fall off a cliff like NFTs did over the past year?

Academics need to consider these very divergent possibilities, and to think through them now, in order to best prepare for how best to respond to the technology. That means considering the possibility of AI collapse.

Large language models (LLMs) are at a huge inflection point now, and this makes it very difficult for academics to decide what to do about them. That’s true for individual professors, staff, and students, and even more so for organizations and institutions. A series of forces are increasingly pressuring AI, perhaps to the point of drastic decline, marginalization, or reset. (I first wrote about these issues back in May. Things are getting more intense.)

Let’s break these down by categories.

(NB: Today I’m going to focus on proprietary, corporate AI, with a coda on open source. I’m also going to emphasize textbots, rather than image creators, because that’s where a lot of the information is now.)

LEGAL THREATS We can start with resistance to AIs scraping the web for training content. Generally speaking, the AI giants have done this to textual and visual content without seeking permission, which flies smack in the face of how copyright works. I doubt their use of Creative Commons-permitted materials was gone in full agreement, either. Since those businesses are obviously doing so for profit, they can’t claim fair use in American copyright law.

This leaves copyright owners with some choices. We can ignore the scraping and move on. We could also try to block our materials from serving as AI training fodder in the future. Last month I noted how some creators were aiming to do so, then asked about academics making such a choice, and have seen no responses (yet! and please let me know if you know of some.) Recently the New York Times announced it was blocking its high-profile and paywalled content from AI scraping. None of these actions individually threatens AI training. If enough intellectual property owners due so, it might, but I think that would take a truly enormous cultural movement.

“Legal threat to AI”: one of Craiyon’s visualizations.

Another choice, and to today’s point, is for a content owner to send Google et al cease and desist notices for copyright infringement, or just to sue them directly. In my post last month I mentioned some early lawsuits. More have been filed since, like this ambitious one. Now there are reports the New York Times is mulling suing OpenAI. To issue the customary denial, I am not a lawyer, but I can research what lawyers and legal scholars have been saying. We could imagine suits ending with AI firms agreeing not to scrape the plaintiffs’ materials. But they could go further. Ars Technica mentions two extra outcomes:

1. “fines [of] up to $150,000 per infringing piece of content.”

2. “the destruction of ChatGPT's dataset.”

For the former, we can imagine the classic case of suing someone out of business, as the fees accumulate beyond a mere cost of doing business and end up ruining the enterprise. For the latter, that’s the end of the software, and the same threat might face any attempt to recreate it. Professor Pam Samuelson has also raised this specter. So that’s two ways we could envision generative AI being hit very hard. Such cases could wend their way through courts for a while, but content owners might also persuade courts that alleged damages are imminent and require fast relief, which could lead to a swift injunction.

What happens to classes assuming AI services will be available this fall and spring semesters if courts shut them down? Similarly, how should researchers plan on upcoming usage?

BUSINESS CASE FAILS Besides the legal threats to AI viability we should consider economic ones. Business based around LLMs could fail, and businesses using generative AI could stop doing so for economic reasons.

This might sound perverse, given how much excitement and hype have been flowing around AI for the past year, not to mention the tidal waves of capital investment, but it’s possible we don’t yet have a business case for the technology. Consider reports from mid-2023 that Open AI lost half a billion dollars in 2022. Just running ChatGPT costs OpenAI $700,000 per day, according to Analytics India.

Right now OpenAI takes in some money through subscriptions and licensing. If that income doesn’t reach healthy levels, and Microsoft tired of pumping billions of dollars into it, the non-profit could fail.

Similarly, if Microsoft’s leadership finds that its various AI offerings - Bing, Office Copilot - aren’t bringing in enough revenue to justify the cost, then it could modify or end the offerings. The same applies to Google with Bard. Recall that Google entered the LLM arena belatedly. And both companies are notorious for killing off offerings they deem unsatisfactory.

It’s possible that all companies will actually find AI to be financially rewarding. Consumer demand might exceed business costs. Yet we haven’t seen a lot of convincing evidence about this so far, and an economic collapse is one option to consider.

Craiyon imagines “failed business model”

Now think of faculty, staff, and students relying on commercial AI for fall research, classes, and operations. What happens to assignments based on ChatGPT or AI assitantance in PowerPoint?

GOVERNMENTAL REGULATION Next up is possible state action against AI businesses.

Why would a government do this? We’ve previously discussed privacy reasons, and in this post fears of intellectual property misuse. Today I’d like to add the large environmental impact of AI training and query operation. The carbon footprint is significant, which leads to (among other things) the spectacle of a technology we can use to understand and fight climate change making the climate crisis worse. Further, the energy and water demands of so much computation can squeeze a world already struggling on both counts. As Shaolei Ren, associate professor of computer science at University of California-Riverside, has put it:

Even putting aside the environmental toll of chip manufacturing and the noise pollution of running AI servers, training a single large language model like GPT-3 can easily consume hundreds of megawatt-hours of electricity, generate many tons of carbon emissions, and evaporate hundreds of thousands of liters of clean freshwater for cooling. As the AI industry booms, concerns with AI’s environmental costs are also growing, especially in marginalized communities that often rely on coal-based energy sources and/or are vulnerable to extended droughts.

That’s at the global level. Ren points out that such draws occur unevenly around the world, and sometimes unjustly:

AI’s environmental costs are not evenly distributed and can be disproportionately higher in certain (sometimes already-disadvantaged) regions than in others. This has raised serious concerns about AI’s environmental inequity, which is stealthily emerging but not yet as widely known as AI’s algorithmic inequity…

The AI Now Institute compared the uneven regional distribution of AI’s environmental costs to “historical practices of settler colonialism and racial capitalism” in its 2023 Landscape report.

We could imagine a government concerned about the climate crisis deciding to use regulations to pressure AI companies to reduce their carbon emissions.

Beyond water and climate is a broader sense of regulatory urgency. The excitement and hype about AI has elicited governmental interest. For example, this recent Foreign Affairs article calls for a quickly rolled out, powerful, and global policy system to ward off a wide range of downsides.

How might such state moves represent an existential threat to AI compaies? A scared or hostile government could regulate firms out of existence, or at least out of operating in their regions. A state could impose costs high enough for an AI business to withdraw. Or regulations might fall on users, making their access to LLMs difficult enough to stop them from using them.

TECHNOLOGICAL FAILURE While generative AI has improved over the past few years, and we futurists can envision ways it might continue to grow, there is also the possibility that the tech could plateau or degrade.

Case in point: this paper finding ChatGPT performance actually getting worse over time and between versions on some tasks. To be fair, researchers also detected improvements. But a decline could drive users away and sap business models.

I can add the “copy of a copy” problem. That is, if we assume some web content appears created by AIs which publish errors through hallucinations, and then future AIs train themselves on (among others) that flawed content. We could see a forthcoming ChatGPT 6.0 being faster, but also more likely to emit bad responses.

CULTURAL MOVEMENTS Beyond legal and policy actions, demand for large language models could plummet. That’s if enough people are appalled, bored by, or enraged at generative AI to reduce usage numbers low enough to convince Microsoft et al to alter or suspend operations.

There is also the recent development of ChatGPT recently losing users.

Web analytics firm SimilarWeb reported this week that worldwide traffic to OpenAI’s ChatGPT webpage, including both desktop and mobile traffic, dipped 9.7 percent in June. In the U.S. alone, the traffic drop is estimated to be 10.3 percent. at the same time, worldwide unique visitors to the ChatGPT website fell 5.7 percent. SimilarWeb also says that the time people spent on the webpage also dropped 8.5 percent. 

In other words, we may have passed peak ChatGPT. It’s unclear if similar user numbers have shown up with Bing, Bard, and users of Microsoft’s Copiloted applications.

My view is that this is due to two factors: users dropping away when they lost interest and many students being in summer session. Analytics India thinks people using the tech secondhand, through an API, rather than from ChatGPT directly, can also explain the drop.

I’ve written previously about this. We could see a wave of cultural disengagement or opposition sweep across nations. That could empower the challenges mentioned above: spurring new lawsuits, encouraging governmental policy, depressing business cases.

---

Wrapping this up, I think it’s clear that there are many ways for generative AI business to get cramped, squeezed, or generally clobbered in the near future. These are possibilities, not probabilities: options we might take.

I raise them because as higher education explores the many uses and implications of generative AI, we should bear in mind that it might, in a sense, collapse. And that can have ripple effects across academia. There are faculty (like me) building assignments using AI. There are students, faculty, and staff counting on AI to assist with their writing, image creation, number crunching, etc. Some faculty are researching LLMs in various departments: computer science, media studies, business, technology and society, art, etc. IT departments have created practices and policies around these tools, or are working on them now. And there’s surely a galaxy of innovative practices and experiments under way. All of this falls into question if, say, OpenAI closes its doors, the European Union blocks Microsoft from running Copilot, or a judge orders Google to delete Bard and all its works.

And yet.

Some think that AI’s benefits outweigh its costs. If I may ventriloquize and do so analogically, fire did a lot of bad things, but overall it was better for humanity’s development that we domesticated it. Many people are bullish on LLMs, and some are in academia. For example, Donald Clark argued clearly for this side in last week’s Future Trends Forum.

There is another aspect to this inflection point which I don’t think has gotten enough attention. So far I’ve been speaking of AI as it appears in the hands of giant companies, like Google, Microsoft (owning OpenAI), Baidu, Amazon, etc. Yet these pressures on AI appear differently in the open source world.

For more on open source AI and education, here’s the excellent Ruben Puentedura on the Future Trends Forum:

What is generative AI becomes an open source success story, as people run or use open tools instead of proprietary ones? Again, I refer to that Analytics India piece:

instead of going for what OpenAI offers, which is a paid, proprietary, and restricted version, why would people not go for an easily modifiable Llama 2? Arguably, it is also a lot better in certain use cases when compared to GPT. That is already happening. According to Santiago, an AI developer, “I’ve talked to two startups migrating from proprietary models into Llama 2.”

Cory Doctorow recently argued that there are major problems with all open source AI tools, often consituting openwashing. I’m still assessing this claim, but wanted to cite it to complete today’s post, in case public interest in open source AI falls. We could see LLMs fail across the board.

For now, back to AI firms: perhaps we’ll see the businesses fall away and open source AI become the norm. How would that change academic engagement with the technology?

That might be a good subject for another post.

(I’m hoping to dive into open source tools with practical experiments as soon as I can. Time is an issue right now, between the personal reasons I mentioned here and the fact that there’s a giant stack of other professional work on fire now. Hopefully I can fire up a laptop with some apps and report back soon. You Substack readers are a great support in this effort!)

(many thanks to a range of friends for conversations on the topic, including Chris Rice)