AI: Considerations for people who make decisions

Recently I gave two presentations in quick succession to the Dutch Network of Government Service Providers and the Dutch Advisory Council for Science, Technology and Innovation about AI. In these two different and brief presentations, I hoped to share some insights that would be useful to the people currently at the helm, the people who can or must make decisions about AI policy. Many thanks to the NPD and AWTI for the invitations and the stimulating discussions!

There is also a Dutch version of this post.

This article originated with these two presentations. The first, at the NPD, focused on the concrete challenges facing decision-makers. The second, at the AWTI, was about what is needed to build a trustworthy (European) AI infrastructure. And, for that matter, also touched on what that actually means. After these presentations I started writing, and eventually came up with this article.

While writing, I discovered just how challenging the subject really is. There are passionate advocates of AI, and there are people who believe it is the devil incarnate. These camps are so polarized that it is impossible to write something that everyone will be happy with.

It is, however, entirely possible to write an article that makes everyone angry about some part. So here goes.

I do hope that both supporters and opponents will come away from this piece a little wiser. Perhaps there is even one thing we can all agree on: what exactly is the rush? Does every municipality really need to deploy Copilot this year? Did Dutch parliament really have to roll it out for everyone already?

What follows is not a “both sides” argument. For people who utterly dislike AI, you’ll find below that AI genuinely can do some remarkable things. For AI enthusiasts, I have a long list of reasons to think carefully about whether it is really such a good idea though.

Overall, I am deeply concerned that we are moving way too fast, and I don’t think that is an unreasonable position given the facts outlined below.

Spoilers — click here for an extended summary

AI as a phenomenon is capable of truly remarkable things, such as winning a Nobel Prize. Whether large language models are actually beneficial for organizations as a whole is another matter. Using AI to produce larger documents more quickly, only for other departments to summarize them again with AI, is not progress. AI is also highly disruptive to organizational career structures: where will your future senior staff come from if AI takes over junior-level work? And will senior staff really enjoy spending their careers checking AI-generated output?

Organizations are experiencing a strong Fear Of Missing Out that pushes them to adopt AI as quickly as possible. Yet despite how impressive AI really is, there is no urgency for most companies and institutions. If AI ultimately proves to be a good idea, it will still be there next year. Sometimes it seems as though people believe that if they don’t deploy now, they’ll never be able to catch up later.

AI comes with enormous problems and risks that deserve careful consideration. It emits absurd amounts of CO2 and consumes so much electricity that additional power plants are being built just to support it, all while we’re already struggling with severe grid congestion and increasingly destructive wildfires. The intellectual property status of the data that goes into AI systems is a massive issue, and it also creates uncertainty about who owns the AI’s output.

People brush these concerns aside surprisingly easily, but make no mistake: the AI industry will make them our problem. Try asking Microsoft whether they’ll indemnify you against intellectual property claims arising from Copilot’s output (a fun question to ask). Meanwhile, AI is disastrous for digital sovereignty: it’s surveillance capitalism all over again, imported from the United States. AI providers leak personal data on a massive scale, sometimes even data belonging to their own employees.

It’s difficult to know whom to listen to for advice. After investments totaling three or four trillion dollars, the AI industry itself is now so deeply committed that it can no longer be considered an impartial source. Much of the media, many consultants, and the broader software ecosystem have become so enthusiastic that it’s increasingly difficult to distinguish fact from hype. On the other hand, plenty of people feel so threatened by AI that they can only produce bad news about it. Then there are the moderates, who say, “AI is just a tool; it all depends on how you use it.” That’s true enough, but they never explain how to use it well, or if this is likely to happen, which makes the advice of limited practical value.

The pressure to roll out AI is so great that most organizations never define their expectations in advance or establish how they’ll measure whether an AI experiment is successful. As a result, they end up with many vaguely defined “successes” that may not actually be successes at all. From the executive suite, everything quickly looks impressive. But are your employees actually becoming more productive or happier? Nobody measures this.

There’s also the fundamental question of what we actually want from AI. It may seem appealing that AI makes writing reports and documents easier, but wasn’t writing those documents itself part of the thinking process that led to good plans? “Writing is thinking.” If you let AI do your writing, who actually learns anything from the exercise? Not you.

The companies providing AI services have also become financially precarious. There is considerable debate about whether the AI bubble is about to burst, or whether it is already beginning to do so. AI itself isn’t going away, but if today’s funding dries up, it could suddenly become much more expensive. That becomes a serious problem if your organization has, perhaps through staff turnover as well, become dependent on it.

In summary, the future of AI remains uncertain in many different ways. Its enormous potential (Nobel Prize!) is obvious, but that doesn’t mean everyone in an office should rush to have AI produce and read all of their documents. For most organizations, there is no train they’re in danger of missing.

A hasty decision to embrace AI, however, can cause enormous damage, to the climate, to your legal position, and to society through increased strain on the electricity grid. It can also drive away part of your workforce and leave you with an unsustainable organizational structure, because where will the next generation of senior staff come from?

In short: wise people don’t jump into the ditch just because everyone else is doing it. Take some time, wait and see, and you may spare your organization a great deal of disruption and unnecessary turmoil. A year from now, the picture will probably be much clearer.

AI: We Simply Don’t Know

The AI of today is not the AI of next year, and its societal impact is changing nonstop as well. It’s a genuine roller coaster. Still, there are ways to make sense of the confusion and arrive at a reasonable understanding of what is happening, and of what is, and is not, wise to do. This article hopes to contribute to that.

Anyone who claims to know exactly how things will turn out is, at the very least, confused. Take this PwC advertisement, for example:

Source - It says in Dutch “The value of AI is not distributed evenly. 74% of all AI value is captured by just 20% of companies”

Which reminds me that approximately 87.539% of all statistics appear to be made up.

A more honest graph comes from the Federal Reserve Bank of Dallas:

Source

According to the “Dallas Fed,” AI will make the economy four times larger, or sixteen times smaller. The truth is probably somewhere in between!

So don’t take anyone who claims to know exactly what’s going to happen too seriously. That doesn’t mean, however, that it’s pointless to think carefully about how we should approach AI as a phenomenon.

The Nobel Prize in Chemistry

Some people like to dismiss AI as nonsense. “Fancy autocomplete.” “A stochastic parrot.” Here’s a counterexample from biology, a field I happen to know quite well:

AI has, in fact, won a genuine Nobel Prize in Chemistry. And as Leiden University explains here, this was entirely deserved. Using an AI model, we can now predict the three-dimensional structure of proteins directly from DNA. This is something humans cannot do, and something computers had previously been unable to help us with in any meaningful way. That Nobel Prize was therefore more than justified. Biology and pharmaceutical research are already making grateful use of these new capabilities. The architecture of this AI model is remarkably similar to that of Large Language Models.

Some people argue that this AI is not truly intelligent, and “doesn’t know what it’s doing.” That’s largely true. But this AI application is already enabling discoveries that benefit all of us.

AI can also be remarkably good at hacking (although it doesn’t always hack what you intended). It can translate quite well and produce audio transcriptions that are better than human-level in many cases (although, once again, you really need to verify what it produces, and almost nobody ultimately does). You can also vibe your way into some very impressive demos.

So no, AI cannot simply be dismissed as “all nonsense.” But that is not a blanket endorsement either. In fact:

I Have to Mention It: The Lawsuits, the Casino, the Climate, Digital Sovereignty, and Grid Congestion

People almost roll their eyes when you bring this up, but every major provider of Large Language Model services is currently embroiled in enormous lawsuits involving publishers, academics, media organizations, newspapers, and others. Under normal circumstances we would hesitate to do business with companies that apparently engage in such controversial practices. But when it comes to AI, that suddenly seems perfectly acceptable.

It’s genuinely amusing to ask AI vendors whether they’ll indemnify you against intellectual property claims arising from their output. Ask your legal department to look into it. The answer, or the lack of one, may surprise you.

By Catboy69 – own work, CC0

The financial condition of AI providers is spectacularly worrying. They’re all investing in one another in increasingly circular ways, and one of the largest players, Oracle, has already seen its credit rating fall to just one notch above “junk”. It’s quite something to build your future on top of a questionable casino. Debt issued by Elon Musk’s AI company is now also being traded as a “junk bond.”

The Financial Times wrote this week:

“Whichever way, AI scepticism now feels like the consensus. As for what to do in preparation for a crash, you’re basically on your own. But if the music does stop playing, an awful lot of people are positioned and ready to say they told you so.”

We had also agreed, quite seriously, to reduce CO2 emissions. A great many countries signed up to that. Within a few years AI is expected to consume more electricity than France (or so it’s claimed), and old nuclear power plants are now being brought back online to satisfy demand. Apparently that’s fine too.

Closer to home, you may find that you can’t get an electricity connection for your new house in the Netherlands, because the grid is already overloaded. Yet we’re eager to build enormous AI data centers here, and preferably as quickly as possible. The implicit message seems to be: you can always live somewhere else.

We had also

[truncated for AI cost control]