AI Underwhelm and Convenience

I am underwhlemed by Mustafa Suleyman. First there was his manifesto for Humanist Superintelligence, which was, let's be frank, sophomoric in its technosolutionism and conflation of human flourishing with economic prosperity.

Then came The Tweet¹:

Jeez there so many cynics! It cracks me up when I hear people call AI underwhelming. I grew up playing Snake on a Nokia phone! The fact that people are unimpressed that we can have a fluent conversation with a super smart AI that can generate any image/video is mindblowing to me. (Twitter, 19 Nov 2025)

We should call this an "Inverse Ratner". Gerald Ratner famously gave a speech to the Institute of Directors in which he called the products his company sold "total crap". The company failed within a year. The problem here was that the premise of the products - largely cheap jewellery - was that they were affordable luxury for people on low incomes so the business relied on the customers being impressed by the products. Ratner effectively called them fools for that.

Suleyman does the opposite: he berates his customers for not being impressed by his products. Like Ratner, he has made the mistake of airing his true opinion. Like Ratner he has shown no respect for his customers. But in this case, instead of accusing them of stupidity for being impressed by his products, he is accusing them of stupidity for being unimpressed! This Inverse Ratner is even worse business sense. If the customers don't like your products (and you are not in a Soviet-style managed economy) then there is something wrong with your product, not the customers.

Setting aside Suleyman's (and presumably the people who could fire him) incompetence as business leaders, let's think a bit harder about why Suleyman is impressed with AI and whether he is right that it is impressive.

Compared to playing Snake, having a fluent conversation with a computer is an amazing technical achievement. No doubt about that. The question is whether we should care about this.

Touchscreens were an impressive technical achievement and there were few if any cynics underwhelmed by them. They gave us a new way of interacting with computers which was easy to learn, intuitive (by and large), and allowed us to get more out of those computers. Perhaps Suleyman wants us to see our fluent conversation with AI in a similar way. However, it is not quite there, is it?

First, the people who are underwhelmed don't want a conversation with a computer. They see it as a tool. So they want to give it instructions and to be able to understand its output easily. Perhaps AI does help with this, but not massively. It is at best an incremental improvement which is probably not worth the learning curve for most people.

Second, because AI lacks all the necessary context that a human assistant has, you have to prompt it really quite carefully to get the result you want, and often send it back for revision. Persona-Context-Task-Format is not a natural way of giving instructions to a human, so there is a non-technical learning curve as well. Most people find giving clear instructions quite hard, but most people are equally good at interpreting poor instructions. If you don't give AI clear instructions, it guesses and sometimes guesses wrong in amusing or annoying ways.²

Third, sometimes it just fails at the task in ways that we don't normally expect a computer to fail. Or a human. I had an amusing exchange with Gemini 2.5Pro in which I asked it to create an image for the title slide of a presentation on Digital Sovereignty. As expected, it created a very boring and only tangentially relevant image, but unexpectedly, misspelled 'Sovereignty'. So I asked it to correct the spelling and reminded it of the correct spelling. It was nauseously apologetic and produced the same image with the same misspelling. This carried on for several minutes until I gave up. The point is not just that no human would do that, it is that no computer should do that. It is meant to be following instructions and these were very clear and simple, on the level of 'Find and Replace'. It only takes one experience like this to become rightly cynical about the technology. Of course, the technical achievement of getting it to produce the image is impressive, but no one cares if it still doesn't do what you tell it to.

Fourth, truly impressive technologies are ones that do things we couldn't otherwise do. Now it is true that some AI does that and in Suleyman's previous role at Deepmind he oversaw some impressive examples, like AlphaFold. The cynics are not denying that such AI is impressive. The problem is that Microsoft is forcing Copilot and related technologies on every Windows user and that only does things that humans can do perfectly well, such as put together some slides for a presentation, find a dubious image to illustrate a point, write an email, summarise a document etc. Yes, some people don't like doing these things and AI can usually do them to a middling standard more quickly - (though only thanks to the vast hidden resources being used up - but that doesn't make it mind-blowing. At best it makes it convenient, like a washing machine or calculator - at worst it allows the partial replacement of human workers with machines, a pattern that is good for the owners of the machines and bad for the human beings.

Perhaps, then, what Suleyman's cynical customers are really being instructed to be impressed by is the convenience that AI brings to our digital lives, and we see a lot of emphasis on this in marketing and enthusiastic accounts of how AI can improve our lives.³ Maybe it is convenient, despite points one to three, but we shouldn't assume that convenience is always and necessarily a good thing, an amelioration of the human condition. Etymology suggests that convenience is the opposite, the contradictory, of inconvenience, and we can agree that inconvenience is always a bad thing. But actually, 'convenient' and 'inconvenient' are contraries not contradictories: nothing can be both, but some things may be neither. For example, not having toilets on a train is inconvenient, but not having toilets on a mountain in the wilderness is neither convenient nor inconvenient. Your trek up Kilimanjaro would not be improved by finding toilets at the summit.⁴

So part of what the cynics who are underwhelmed by AI may be saying is that, while it may be convenient, it is not removing something they find an inconvenience. Whether convenience is a good thing or not depends on whether the alternative is inconvenience - a bad thing - or something which is not convenient but is otherwise worthwhile. Someone who chooses to wash their dishes by hand or add up columns of numbers in their head is not choosing inconvenience but a different, slower, more humanly engaged approach, so the convenience of a dishwasher is not of value to them.

The same is probably true of many AI tools from the perspective of the people who are underwhelmed. Take AI search. That is certainly convenient, but a traditional search with a list of links, where one can read the sources and snippets before choosing to click through and read some of the pages, is not inconvenient. It involves the exercise of curiosity and judgement; it involves taking control of your information sources, being critical, identifying the irrelevant or unreliable, and sometimes serendipitous discovery. Those are all good things, so even if they take more time than an AI summary, they are not an inconvenience and the convenience of the AI summary is actually a loss of value. No wonder it feels underwhelming.