Laura Carter

Today, a federal court blocked enforcement of a U.S. immigration policy that targets and censors non-citizen independent technology researchers. The Coalition for Independent Technology Research—of which I am a member—sued to challenge this policy, which was being used to call for visa revocation and even deportation of people who study online harms and societal impacts of technology. Today’s decision in CITR v. Rubio suspending the policy’s enforcement is temporary—not the end of our case against government censorship—but it’s a major step forward for independent technology research.

As an independent technology researcher and an immigrant to the US, I'm taking a moment to celebrate this important milestone: a ruling that helps counter the deep chilling effect this policy has had across the independent research community. Independent work by researchers, advocates, journalists, fact-checkers, and trust and safety workers is crucial in understanding the impact that technology continues to have on the societies we live in. We should be able to do this without fear of censorship, repression or reprisal.

SoftBank is a Japanese investment conglomerate most famous—or notorious—in the English-speaking world for their string of investments in startups you've never heard of* (and companies like WeWork, which you have heard of for being wildly overvalued). A presentation by their founder and CEO, Masayoshi Son, for the 2026 Shareholder AGM** has been doing the rounds this week for being, frankly, completely deranged about AI.*** Using a metaphorical goose.

As best I can tell, the presentation makes the argument that the value of the company can be expressed in golden eggs. Over 16 years, SoftBank has gone from a valuation of 3 billion yen to 74 billion yen. Son's argument is that this value increase doesn't come from the eggs (slide 48 states that 'eggs do not lay eggs') but from the goose that laid them. And according to slide 55, 'ASI'—almost certainly 'artificial superintelligence,' a term that AI superfans use for hypothetical systems that are orders of magnitude smarter than humans—will be a goose that produces even more golden eggs.

Obviously, this is nonsense. ASI isn't real. And the golden egg metaphor obviously doesn't work: goose eggs don't lay more eggs, but rich people's money absolutely does turn into more money, through interest, dividends, and returns from selling stocks and so forth.

But this slide is inadvertently a pretty good encapsulation of how a lot of people—and unfortunately, many people in decision-making roles—think about 'artificial intelligence' in general. That's not how geese work! A goose isn't a factory with feathers, containing an assembly line producing eggs (golden or otherwise). It's a bird: it will lay only when it's old enough, when it's properly fed and cared for, and at the right time of year.**** Laying eggs isn't risk free: geese, like other birds, can get egg-bound if the egg gets stuck in the cloaca. And eventually the goose will get too old to lay.

It's also not how AI works: at least, not the AI we have today. Large language models aren't intelligent: they are enormous computer programs that use data and machine learning to model what human language looks like and produce samples of it based on the model. This is impressive! But it's not intelligence. There's a lot of money invested in making the argument that it is, and that soon LLMs will be as smart as humans, and soon after that exponentially smarter than us (that's the 'artificial superintelligence'). But LLMs aren't human, and they don't learn like humans, and they don't interact with others the way humans do, and there's no real evidence that they will any time soon (or indeed ever).

Factories also require maintenance and supplies, of course. But they're very different from those required by geese! Treating a goose like a factory ignores all the reasons why geese lay eggs and why factories produce products: just as treating a large language model like an intelligent human ignores all the reasons why LLMs produce generated text and humans produce...language, communication, creativity, relationships, and all the other signs of intelligence you can think of. Trying to get more eggs out of a goose with more electricity, better machinery and increasingly automated processes is not going to work. Nor is trying to get 'more intelligence' out of an LLM with more electricity, data and chips. However much tech companies and investment conglomerates would like to think they are the same thing, geese aren't factories, and humans aren't LLMs.

The original fable, of course, ends with the owners, driven by greed, killing the goose that laid the golden eggs: only to find that it's just the same as all their other geese on the inside. Treating LLM-generated text as intelligence, and therefore extrapolating a misunderstanding of where it comes from, what it can do, and what its limits are, is likely to end badly. Masayoshi Son and SoftBank might lose billions of yen, but they will still have plenty. The rest of us might not be so lucky.

* Someone I know worked at a startup which received SoftBank money. They told me the funding went “straight up the founder’s nose.”

** I have linked there to an archive.org copy of the presentation. There is currently a version of the presentation on the SoftBank website, though I haven't been able to find a direct link to it from the page about the 2026 AGM. The AGM was conducted in Japanese, and the slides are in English.

*** Even compared to some of the nonsense doing the rounds currently, this is ludicrous. Which is saying something.

**** Unlike chickens, which do lay all year round. And also unlike chickens, there aren't many goose egg factory farms (geese are factory farmed, but for their meat and for their livers in foie gras).

This year, I’m working through Nathalie Tasler’s prompts. I'll keep updating this post as the month goes on.

Day 9

Cheating a little here, with from the Scriberia illustrations commissioned during a Turing Way book dash, back when I was in the very early stage of thinking about this idea.

Day 8

Mastodon: I'm writing a book about how different people are—and are not—represented in different data sets. In particular, it's about how the ways that different actors collect data exclude or misrepresent people who are in some way out of the mainstream: because of their #gender, their way of life, their identity/ies.

LinkedIn: My current book project focuses on exclusion and misrepresentation in data sets. When we collect data about people, the information that we choose to collect is a choice: and not everyone fits neatly into predefined categories.

Day 7

I have a tendency to ramble in my writing, and to use a lot of run-on sentences. I'd like to get better at writing more concisely—and at editing my own writing.

Day 6

I don't take as much care of my readers as I should: at least, not in academic writing. Sometimes this is because I'm racing to meet a deadline and 'done' is better than 'perfect,' sometimes it's because I'm so fed up of a paper that I just want shot of it. And sometimes it's because I feel pressure to 'perform' academic inscrutability. I'm trying to move away from that—my book project idea is for a general reader—but it does creep in, especially when I've been reading a lot of academic literature that tends towards the pompous or the jargon-y.

Day 5

A possible abstract for a paper or conference submission:

Missing drafts Machine learning—including generative AI—relies on training data, whether it is in the form of structured data, unstructured text, or images. But this data didn't spring into existence fully-formed as a .html, .txt or .csv file. This paper will make visible the information and thought that went into creating the training data that is used to build machine learning, large language and image generation models. From court case submissions that are used in crafting the final judgement produced by a judge, to tried-and-failed experiments which never produce academic papers, to the previous drafts of novels which exist only on the writer's computer: this paper will show how this hidden information is vital to understanding how training data was created, and therefore to understanding the performance of machine learning models.

Day 4

To write, I need quiet, a door that closes, to be left alone. I need to spend time settling in to the work, to pick up where I left off before, and to quiet my brain from the day-to-day.

This isn't always feasible! But it's easier to come by in the afternoons.

Day 3

Weakness: sometimes it is useful to spend time looking for a reference! Future Laura isn't always grateful when Past Laura leaves her with all the work to do. Sometimes looking for the reference helps me realise that the point I was trying to make isn't actually that useful or relevant.

Strength: when I am writing well, I am able to put a lot of words down—and it's always easier to edit existing text than come up with it in the first place.

Day 2

When I am writing at my best, I am like a well-oiled machine: the words flow easily from my brain through my fingers and on to the screen. I don't get bogged down in looking for the perfect reference or exploring a footnote – I simply note where these are needed and keep going.

Day 1

I'm hoping to use these prompts to help me think through my current book project!

The public doesn’t want AI ‘mainlined into the veins’ of the UK: at least not when it comes to the public sector.

How do we know? Well, my former colleagues at the Ada Lovelace Institute have been asking them. Over the last six years, Ada has done a lot of research into what the public want from data and AI. Earlier this year, they asked me to help pull together findings from their work with 16,000 people in four nationwide attitudinal surveys and 400 people in deeper qualitative studies, to help answer the question of how the general public