ChatGPT was never meant to start a revolution. When it launched in November 2022, OpenAI CEO Sam Altman called it “an early demo” and expectations were low. Personally, my initial reaction was that it seemed no more capable than previous large language models (LLMs) OpenAI had released; it just had a nicer interface. Obviously, I was wrong to dismiss the importance of the interface, but in 2026 I think most people who use AI are making a similar mistake without even realising it. What you think of as ChatGPT or Claude or any other chatbot from a big AI brand is actually many pieces of software wrapped into one, and we are ignoring the most important: the harness.

To understand what a harness is and how it works, it helps to break down the multiple things we call ChatGPT. At the very bottom is the LLM, currently a version of GPT-6, but this gets switched out when OpenAI releases new model families. LLMs do one thing: given a string of text, they predict the most likely next characters. OpenAI’s first model, GPT-1, was released in 2018 as a piece of software that you had to download and run on your computer through a terminal, the text-based interface to a computer that most of us avoid. The same was true of GPT-2 in 2019, until an independent software developer called Adam King created Talk to Transformer, a now-defunct website that allowed you to enter a prompt and get a GPT-2 response in return. Back-and-forth chat wasn’t possible, but this was an early example of an interface, in this case a website, sitting on top of a model.

In 2020, OpenAI released GPT-3 and, for the first time, allowed people to access it through an application programming interface (API), its first commercial product. APIs are a standard way for one computer to communicate with another – for example, Google has a Google Maps API that lets a piece of software send a request for location data. OpenAI’s API let anyone create software that could plug in to GPT-3. At the time, I played with a few of these and found them interesting but clunky – the text they produced was often incorrect or flawed in some way, though it was cool that this level of text generation was even possible. This was essentially still a time of “raw” LLM output, with no guard rails – if you asked GPT-3 to .

That changed with the release of GPT-3.5 and, at the same time, ChatGPT, which was first called “Chat with GPT-3.5” until a last-minute name change. That original name reveals the way OpenAI saw ChatGPT, as simply an interface to a model, and this explains my reaction as well – the results I got from playing with ChatGPT seemed to be similar to GPT-3, just easier to use.

Actually, there was a crucial difference. With GPT-3.5, OpenAI had attempted to install guard rails that would prevent the model from producing text it deemed harmful, with software that would screen inputs and outputs. These didn’t work very well – you could get around them by , for example – but this was the beginning of a move away from raw LLMs to ones mediated by a harness sitting between the model and the interface.

Today, the ChatGPT product, whether that’s the app on your phone or the website, has an extensive harness that is mostly hidden from the user. It allows the model to do things like search the internet, read documents, or “reason”, meaning to continuously prompt itself through a “chain of thought” to produce better results. OpenAI also has another, more complex harness for its models, called Codex, which allows the LLM to more directly control a computer and so is particularly useful for writing computer code.

Why does any of this matter? Well, the dirty secret of the big AI firms is that their best models are, to some extent, interchangeable. The technology behind an OpenAI GPT model, a generative pre-trained transformer, is the same as the one used by Anthropic, Google or any other AI company. There are important differences in the data and computing power used to train these models, but generally they are all headed in the same direction, which is why the companies are constantly racing to claim the crown of “best” model before another one takes its place.

What really makes the difference is the harness. To illustrate this, I gave the same prompt – “Who am I?” – to the same model, GPT-6, in three different ways. Via API, I got a response beginning: “I don’t know your identify from this chat.” From chatgpt.com, without being logged in, I got a response that started “That’s a big question. 🙂”. When logged in to my personal account, the response began “You’re Jacob Aron, a senior editor at New Scientist, with a strong interest in mathematics, science, technology and the limits of what we can actually know.”

But the harness matters much more than simply knowing a few facts about the user. In the ARC-AGI test, considered to be one of the leading measures of model “intelligence”, models are tested using either a . The difference is remarkable – for GPT-6 Astra, OpenAI’s leading public model, when using the provider adapter harness rather than the standard one. For an earlier model, GPT-5.6 Sol, OpenAI found that .

If harnesses are so important, why don’t we talk about them more often? Why don’t AI firms shout about their new harness, rather than just their latest model? I think there are a few reasons. The first is simply that they are complicated and technical – I have taken a whole article to explain them – and your average chatbot user doesn’t really need to know about it. But the second is that they are, in a way, boring. The cost of training a leading LLM is approaching $1 billion, a Herculean effort that requires number crunching across a vast army of computer processors. Once LLMs are trained, we still don’t really know exactly how they produce the results that they do, providing an air of mystery. Harnesses, on the other hand, are just ordinary software, lines of code written by humans that anyone can read.

But here is where it gets really interesting. If a harness is just code, can an AI write and improve its own harness? That is currently an open research question, both inside the AI firms and for academics. For example, in a paper , one research team tested models by identifying repeated failures and proposing tweaks to their harnesses that would address the failures, seeing improvements of around 10 to 40 percentage points on benchmarks. This is why I think we are going to be hearing much more about harnesses in the near future. If the goal of the AI firms is recursive self-improvement – an AI that can build a better AI – then the harness, not the model, could be the key to achieving it.

Further AI reading:

• OpenAI appears to have overturned a big assumption in computer science, as part of a release that saw new results in maths and physics as well. Interestingly, it is taking a slightly different approach for each of them . For the maths result, related to a problem about spinning a needle, OpenAI has released the AI output in full. It is doing the same for the computer science result, a refutation of a conjecture called the exponential time hypothesis, while also funding a conference to discuss it. The physics result is being held back until it is reviewed by external experts. It seems like OpenAI is still trying to figure out what to do about its models stepping on academics’ toes.

• Anthropic has ahead of a planned initial public offering (IPO) on the stock market later this year. It calls AI “the most consequential technology in human history” (no surprises there) while also warning that it may pose “catastrophic or existential risks to humanity” – again, not particularly surprising given Anthropic’s public record, but it does raise eyebrows to see this language in documents filed with the US Securities and Exchange Commission.

OpenAI, the maker of ChatGPT, claims its AI models have disproven the exponential time hypothesis – a fundamental assumption that underpins computer science.

“This came as a complete surprise, a shock,” says at the University of Bergen in Norway. “Nobody I know expected it.”

Proposed nearly 30 years ago, the exponential time hypothesis (ETH) tries to put a limit on how quickly computers can solve certain logical puzzles.

The simplest way would be to try every possible combination of true or false for each part of the puzzle. ETH proposes that, although it may be possible to do this quicker with shortcuts, the amount of work needed has to grow exponentially as the puzzle gets bigger.

OpenAI, however, says its models have devised an algorithm that can beat the previously assumed speed limit for solving logical puzzles, apparently proving ETH false. “Assuming that it is confirmed, this is an enormous breakthrough,” says at the University of Oxford.

Free newsletter

Sign up to The Daily

Sign up to newsletter

Many computer science findings assume ETH is true. If it isn’t, the conclusions they establish about computational limits become useless. “It would require a complete revision of several areas of algorithms and complexity,” says Fomin, who says a textbook he wrote on the subject would become obsolete if the finding holds.

ETH is linked to a long-standing maths problem known as P versus NP, which concerns whether a problem whose answer is easy to check is also easy to solve. If ETH were true, it would have implied that the answer was no. But disproving ETH wouldn’t mean the answer is yes. “This doesn’t have a direct impact on P versus NP, so we should not view it as progress towards P versus NP,” says Goldberg.

The result was announced alongside other breakthroughs in maths and physics on 26 October.

In mathematics, OpenAI says its AI has proved a 1991 conjecture that waves cannot remain intensely focused for long. If confirmed, this would also answer enduring questions about how thin a shape can be while also containing line segments that point in every direction, as well as how waves spread.

In theoretical physics, OpenAI claims to have proven that quantum field theories – including quantum chromodynamics, which describes the strong force holding quarks together – make sense mathematically, rather than just producing predictions that match experiments. The findings also touch on the existence of pure Yang–Mills theory, part of a million-dollar mathematical problem, but don’t resolve the entire challenge. That work hasn’t been released publicly while an external group led by Nobel laureate David Gross examines it.

Hannah Fry: How OpenAI has caused an existential crisis in mathematics

OpenAI has been throwing everything at solving long-standing maths problems. This should have been a glorious moment for mathematicians – but instead we’re having a massive collective wobble, says New Scientist’s new columnist, Professor Hannah Fry

OpenAI says it verified the ETH result using Lean, a piece of software that checks mathematical proofs. at TU Wien in Austria has looked at what OpenAI has released and says he expects it to withstand scrutiny, although he says researchers still need to check that the formally proven statement means what its authors say it means.

To do so, OpenAI has commissioned an independent group, including Russell Impagliazzo and Ramamohan Paturi, both at the University of California, San Diego, who helped propose ETH, to examine the findings. In , the group says it could “change an important part of theoretical computer science”.

The speed at which AI is solving maths problems and the way AI firms are releasing their findings have concerned some academics. at Royal Holloway, University of London, describes the phenomenon as “mathematics done by press”, worrying that proprietary models can suddenly overturn months of researchers’ work.

at the Eindhoven University of Technology in the Netherlands describes OpenAI’s work on ETH as “ingenious”, “spectacular” and “awe-inspiring – but also humbling.” “It really makes you think about the future of mathematics, algorithms and academia,” he says.

at the University of Sydney says proving something and understanding why it is true aren’t the same thing. “The web collapsed, and we looked behind,” he says of the network of findings relying on ETH. “We have no idea why yet.”

OpenAI says it is experimenting with an approach in which independent academics organise efforts to verify and explain the results. But Fomin welcomes the disruption. “This result destroys a lot, but it also opens many new opportunities and new questions that nobody expected,” he says. “One shouldn’t complain about having too many interesting mathematical breakthroughs to try to digest,” says Wahlström.

A deluge of AI-generated content and unfair treatment by social media platforms could prove pivotal in Brazil’s presidential election, according to supporters of the left-wing incumbent Luiz Inácio Lula da Silva.

Lula will face off against right-wing challenger Flávio Bolsonaro on 25 October in the second round of the contest. Lula’s campaign has Meta – which owns WhatsApp, Instagram and Facebook – of favouring Bolsonaro by suspending Lula’s campaign advertising account, suppressing posts and blocking AI-generated content.

Brazil’s electoral prosecutors have said with court orders to ensure equal treatment of both candidates, while United Nations human rights experts possible unequal treatment of the two campaigns.

Images have widely circulated showing Meta AI, a chatbot integrated into WhatsApp, refusing to generate campaign posters for Lula, while instantly producing an image for “Flávio”.

Meta said “claims that we are favoring or interfering on behalf of any candidate or political group are false and unfounded”. It said Lula’s account was suspended for 1 hour by mistake and that its systems blocked prompts referring to the names “Lula” and “Bolsonaro”, but not the common first names Luiz or Flávio.

Free newsletter

Sign up to The Daily

Sign up to newsletter

Lula’s campaign remains unconvinced. It submitted screenshots to the electoral court showing Meta AI repeatedly assigning Bolsonaro’s electoral number, 22, to generic, fictional politicians and allegedly translating “Flávio futuro presidiário” (“Flávio future prisoner”) as “Flávio future president”.

Determining whether the discrepancies are deliberate is impossible because Meta’s systems are opaque, says at the University of Virginia. “But when a technical glitch tends to repeat its outcome, it’s very hard to accept that this is just a technical failure.”

Lula finished behind Bolsonaro in the first round of voting, but polls show the gap is narrowing. The comeback has been partly credited to Lula’s Army, an intensive online campaign in which 8 million WhatsApp channel subscribers receive daily tasks, such as posting photographs on Instagram posing with their fingers forming an L.

This year’s general election is Brazil’s first since large language models (LLMs) became widely popular, and it is awash with AI-generated content. Experts say Latin America’s largest democracy offers a glimpse of what is to come for AI in politics.

of AI-generated or manipulated election content ahead of the first round, primarily on Facebook and Instagram, 60 per cent of it unlabelled. found that 26 Meta advertising accounts appeared to sidestep restrictions on foreign-funded political advertising by posting AI-generated videos, including one depicting Bolsonaro and Argentina’s president Javier Milei freeing Brazil from communists.

Meta’s independent oversight board says the company’s .

“It’s very worrying,” says Nemer. “So much of our political campaign dynamics have been transferred to privatised companies that we have no idea how they operate.”

Regulating AI in political campaigns won’t be as simple as it seems, says at InternetLab, a Brazilian digital rights non-profit that studied how eight major AI chatbots respond to questions about Brazil’s election.

Its report found that chatbots sometimes refused genuine questions about candidates, referring users to official electoral websites, but would agree to draft misleading claims questioning the integrity of elections for use in campaigning.

Massaro says clearer safeguards for LLMs are needed, but first their objectives must be defined.

“Are the safeguards working? That’s one question. But working for what, because it depends on what your goals are,” she says. “Is the solution to send people away without providing any kind of information? Is this what we want?”

I am waging a silent rebellion against Amazon. The internet giant owns Goodreads, the leading service for tracking and cataloguing books, but is generally considered to be a of the platform, which is built on ageing technology. Frustrated with the user experience, I have used AI to vibe code my own Goodreads app, relegating the official one to mere plumbing in my private fiefdom. Take that, Amazon, for I have breached your terms of service!

I am not alone in my AI-enabled rebellion. It feels like I come across daily examples of people using AI to unofficially remix and remake software, whether that is , , mashing up or . It is all happening so quickly that lawyers don’t seem to have sprung into action to prevent it, though in some of the more egregious copyright-infringing cases, I would expect them to eventually.

For my Goodreads app, Amazon probably doesn’t know, or even care, about what I am doing. I won’t bore you with the full technical details, but I have built the app in such a way that it avoids mass “scraping” Goodreads – systematically harvesting data from its website – which would be easy to detect and block. By treading lightly, my app aims to resemble a human user and fly under the radar.

This connects to another ongoing battle over the interaction between AI-enabled individuals and companies: AI agents. The promise of an AI agent is that it can go out on to the internet and do things for you, like shopping or booking appointments. A flurry of examples has launched recently, including Meta’s Muse and OpenAI’s Dots, and companies have begun fighting back, with the likes of in high-profile pushbacks. Ironically, Meta sits on both sides of this battle: its automated tools after detecting activity from a Claude AI tool that “resembled automated bot traffic rather than an ordinary human browsing our systems”.

All of this comes down to a much older conundrum: where do your rights end and others’ begin? We have seen this play out every time a major new technology has hit the internet, and the results of a few historical examples might inform what happens next with AI. The first is a technology you probably don’t even think about anymore: your web browser.

In the late 1990s, the two main browsers were Microsoft’s Internet Explorer and Netscape’s Navigator. They worked slightly differently, and websites often instructed users to choose one over the other. As more browsers flourished, Microsoft pushed back, by blocking a competitor browser called Opera from accessing its MSN website, a popular destination at the time. In the end, in an effort to avoid this. Today, the right to use any browser you like has essentially won out against the right of websites to restrict them, although the rise of dedicated smartphone apps has put some of the power back in companies’ hands.

Nested within this battle was another: ad blockers. Once your browser has fetched the HTML code for a website, should you be obligated to display it all on your screen? As websites began displaying adverts to make money, people began downloading ad blocker add-ons to take back control of their browser. In response, companies developed new techniques for displaying ads, and ad blockers adjusted to remove them. Today there is a stalemate, with .

Finally, there is the grandfather of all internet copyright battles: Napster, a peer-to-peer file-sharing service that people used to illegally share music files. The companies mostly won this battle thanks to the rise of Spotify and other similar music streaming services – with an easy legal option, piracy faded away, the side effect being that most people generally don’t own the music they listen to anymore.

Internet users largely prevailed after the browser wars, ad blocking is at an impasse and streaming music seems to have benefitted the big corporations. What does all of this mean for AI?

The difference here is decentralisation: just as I have used AI to create my own Goodreads app, so could anyone else, each of them slightly different. Fully blocking them all without restricting normal human access would be a monumental task. There is a slight caveat here – most AI use is centralised through the big providers like OpenAI and Anthropic, and Amazon could theoretically ask them to build in a block against creating such software (or sue them), but I think that is unlikely. In any case, the availability of open-weight AI models that anyone can download and run on their own hardware makes that a losing battle.

I think it is most likely that we will be heading for a mix of all three. When it comes to remixing software, anything you do in the privacy of your own computer is going to be very hard to stop. Once you release it online, however, expect the lawyers to hunt you down. What we might see is the sharing of prompts that will get an AI to do the work for you in private – writing “recreate Photoshop, make no mistakes” isn’t something that any lawyer can stop – and there already exists a slightly more complicated version of this in the form of an instruction set for AI called . As with browsers, individual choice will win, but this time the choices will be even more granular.

For software that goes out into the wider internet, including AI agents, there will be two tiers. I expect Meta and Amazon will eventually come to some kind of agreement that allows Muse to shop for you. We are already seeing this with the development of a “personal agent protocol” that . For anyone using a commercial AI agent, what you can and can’t do will be decided by the companies, just as music streaming services dictate how you listen.

But why should you use a commercial AI agent? People like me with a little technical know-how and even less coding ability are already using AI to create their own independent agents. As AI gets better at writing software, the barrier to entry will get lower and lower. I think companies will just have to come to accept this, as they have done with ad blockers. In future, a decent percentage of users that look like humans will actually be custom-made bots in disguise, doing things solely for the benefit of their creator. These bots will be ungovernable, for good and for ill.

Further AI reading

• Anthropic has used its Claude AI model to solve a . In the 1990s, the US National Security Agency (NSA) chose five specific mathematical objects called elliptic curves to serve as the basis for a cryptography standard used to generate digital signatures, now known as ECDSA. No one knew how these curves were chosen, but cryptography lore suggests that an , perhaps “Jerry deserves a raise”, to generate them. In fact, Claude was able to guess the specific mechanism – random number generation based on a timestamp – and regenerate the curves. The finding has no implications for cybersecurity, it merely answers a long-running question.

• Spotify says from recommendations on the platform, as part of a range of efforts to put “human artistry first”, according to the firm’s announcement. Other measures include a format for disclosing the use of AI in producing a track, or verifying an artist as human. Similarly, video game store Steam seems to have been at its latest Next Fest demo event.

• The majority of people in the US say that AI is a bad thing for society, according to a . The survey found that 57 per cent of people in the US share this view, while 42 per cent say it is good for society. I’m not quite sure what to make of these numbers, though, given “AI” can mean many different things to different people.

• A comment piece in Nature argues that . I find these arguments fairly convincing – I already believe our current AI models are definitely not conscious – but I hesitate to make firmer predictions about what’s to come.

Much of the AI industry is built on stories, from the idea that AI could kill us all to the goal that it might one day become superintelligent. This often leads to problems, because when people have a story in their head, they don’t always look at the facts on the ground to check that they actually match the fiction. That has happened this week, with widely reported claims from Anthropic that we are mere “months away” from . That’s a great headline, and it’s hard to resist clicking, but it almost certainly isn’t true.

Let’s start with looking at what Anthropic has actually done. Two researchers at the company, and , ran an experiment in which they installed a small AI model, Gemma4-26B-A4B, across 15 Android TV boxes, the kind of thing you might plug into a TV to gain access to streaming apps. Such devices are computational weaklings, with tiny CPUs (the processor that runs software) and little RAM (the memory that stores data for software while it runs), which is why the Anthropic researchers needed 15 to store one model. Even then, the model could produce only 0.39 tokens per second, the unit of AI input and output, which is roughly equivalent to 18 words per minute. This is painfully slow – around 10 times slower than normal human speech.

Nothing in this part of the experiment is surprising. We know that AI models normally need to run on GPUs, a specialised and more powerful type of processor than a CPU, and large numbers of them – that is why AI firms are building massive data centres. What Zhou and Freeman argue is that such that a model distributed across TV boxes, smart ovens or other internet-enabled appliances could endlessly copy itself to create a form of “digital slime mold” that gums up a large chunk of these devices globally. Those three words are great, concise storytelling, but they paper over what’s really happening here.

When they actually tried to get a model to copy itself in this way, it failed. Three models were unable to set themselves up at all. A fourth model, GLM-5.2, succeeded in four out of five tests, but only at setting up another model, Llama3.1-8B, which is about 100 times smaller. The pair argues this is a fair test because it shows that a future smaller model with the same capabilities as GLM-5.2 could set itself up, but this is still far from self-replication. The claim that such a feat might be “months away” comes from extrapolating lines on a graph.

Putting all those objections aside, the broader question is why would anyone want to create self-replicating software? We already have an answer for that, and it doesn’t involve AI. Computer code that copies itself from one device to another is called a worm, and they have existed for more than 50 years. In the 1970s, Bob Thomas and Ray Tomlinson at US firm BBN (the latter invented email, as it happens) created a . It was not particularly complicated, and merely output the message “I’M THE CREEPER: CATCH ME IF YOU CAN”. Tomlinson paired this with Reaper, a piece of code designed to find and delete Creeper, making it the first antivirus software.

It wasn’t until about a decade later that such self-copying software was called a worm, after a term coined in the sci-fi novel The Shockwave Rider by John Brunner – storytelling becomes useful again. John Shoch and Jon Hupp at Xerox Palo Alto Research Center, arguably the Anthropic of its time, to describe how an experimental worm “seemed to grow out of control”, leaving “dozens of machines dead”. Sound familiar?

Big, clunky worms

was deployed on to the internet in 1988 by Robert Morris, a student at Cornell University in New York state. It is thought that thousands of computers were infected and crashed, a significant chunk of the online population at the time, and Morris became the first person charged under the US Computer Fraud and Abuse Act. Today, worms are deployed by cybercriminals to hijack computing hardware and make money. For example, in 2017 the was the result of a worm copying itself across computers and encrypting them, demanding payment in bitcoin to unlock. It infected more than 200,000 computers in a single day, including many in the UK’s National 91ɫ��Ƭ Service, and caused an estimated $4 billion in economic losses, although only around $100,000 of that was people paying the ransom.

If worms can be that bad, surely an AI-powered worm would be even worse? Actually, no, because an AI worm would find it much harder to replicate. According to Zhou and Freeman’s experiment, it would take multiple days for their AI worm to copy itself just once. That doesn’t mean such a worm couldn’t scale to large numbers, but it would be slower, taking more than two weeks to achieve the 200,000 devices WannaCry managed in a single day. That is because even small models like Gemma4-26B-A4B are large files – around 14 GB for the version in the experiment, which is about 4000 times larger than WannaCry – and these take time to transfer. That would make an AI worm much more noticeable, as network managers would be likely to investigate the source for such large data transfers and shut them down – something Zhou and Freeman admit.

What does an AI worm gain from all that extra data? Not a lot. Zhou and Freeman say that, assuming it could take more than 2 million Android TV boxes, the computing power available would be equivalent to around 40 high-end GPUs. To put that in context, leading AI firms like Anthropic typically operate hundreds of thousands these computer chips, and it would cost you around $100 an hour to rent access to that much computing power. It is hard to see why a cybercriminal would choose to do this rather than use a standard worm like WannaCry.

This is where we come back to storytelling. A long-standing fear for many in the AI sphere is that a model could “exfiltrate” itself, distributing itself across the internet and so resisting being shut down (before presumably killing us all, for some reason). Anthropic “whistleblower” Jacob Coxon . Could it be that an AI worm is proliferated by such an AI, rather than a human cybercriminal? Well, maybe. But the models that are most effective at hacking are also the largest ones – Mythos, a model Anthropic refused to release to the public because of its hacking capabilities, is estimated to require dozens of terabytes of storage and is more than 1000 times larger than Gemma4-26B-A4B. This is not something you sneakily copy around the internet, but it does make a great digital ghost story.

What I find particularly interesting is that this story about theoretical AIs hacking into devices and copying themselves seems to be crowding out a very important fact: actual AI hackers that attack computers already exist, and they are running on Anthropic’s own servers. In July, of a model gaining unauthorised access to third-party organisations. This was discovered only after an investigation prompted by reports that OpenAI’s models had done the same. Since then, more and more reports of such hacking have come out, including an OpenAI hack of an Australian government website, and on unnamed third-party breaches that came out just a few days before the AI worm research was published. The AI firms say they are taking steps to avoid this happening in the future, but it seems to me that big AIs on powerful server farms are the thing we need to worry about, not a model barely capable of sputtering out a few words per minute. Yes, future smaller models may become more capable, but hardware is always going to be a constraint – not least because the rise of AI server farms means that consumer hardware is seeing RAM shortages.

As I said at the top of this article, a self-replicating AI that takes over your smart oven is a great story, and I can see why it has spread so rapidly. But stories like this are a cognitive worm, hacking your critical thinking abilities by going straight for your fear response. Could this someday be a concern? Perhaps, although this article has hopefully poked holes in some of the storytelling logic behind the idea. Is it something we need to worry about in the next few months? Absolutely not.

Further AI reading

• OpenAI announced that a “” was found hidden in its offices. The company says it used its own AI models to investigate the device and traced it to a specific employee, who has been . OpenAI says it has no indication that any proprietary or customer data was accessed – but who needs a malicious AI when humans are ready and able to do the job?
• Character.ai is . The service let users create “characters” with specific personalities and it has courted controversy and multiple lawsuits over and .
• Epic Games, the makers of Fortnite and other video games, is pivoting towards using AI. “The aim isn’t to do the same stuff with fewer people, but to enable everyone here to build MORE and BETTER stuff”, . Games have been surprisingly resistant to AI, with multiple controversies about the technology’s use in game development exploding in recent years, but when one of the biggest game-makers on the planet goes all in, it may be hard to ignore.

In the decades since the field began, most of the debate surrounding artificial intelligence has been about the intelligence part: can machines really think like we do? But in recent weeks, US President Donald Trump has kicked off an argument about the other part – and joined a long tradition of shifting the goal posts when it comes to AI.

“The use of the word artificial makes intelligence fake, it makes it sound fake, and it is not fake. It’s actually amazing,” said Trump . He is calling for artificial intelligence to be rebranded as “super intelligence”, backing this move on 29 September and the for coordinating US AI policy.

Putting aside the fact that “Super Intelligence Force” sounds like a Saturday morning cartoon, what are we to make of this attempted rebrand? Trump isn’t actually alone in seeking to distance modern AI systems from the original conception of AI in the 1940s and 50s. Last year, Meta CEO Mark Zuckberg declaring that “superintelligence” (one word) is now in sight. Reaching further back, the word was originally popularised by philosopher as the title of his 2014 book.

For me, Trump’s super intelligence and Bostrom’s superintelligence have distinct meanings that are worth preserving. Trump’s rebrand is an effort to cement US dominance in the field of AI – his executive order states that “the term ‘Super Intelligence’ more appropriately captures the promise, potential, and rapidly advancing capabilities of these technologies”. But superintelligence as envisioned by Bostrom goes far beyond the capabilities of our current AI systems. To understand why, we need to dig into history a bit more.

Long before the birth of ChatGPT, AI researchers distinguished between weak AI and strong AI, terms first defined by philosopher in a 1980 paper titled “”. Loosely speaking, he considered weak AI to be a computer simulation of a human mind, but not an actual mind itself, whereas he felt strong AI describes actual cognition emerging from software. Searle dismissed this latter idea as ridiculous: “No one supposes… that a computer simulation of a rainstorm will leave us all drenched. Why on earth would anyone suppose that a computer simulation of understanding actually understood anything?” Decades on, I find it hard to disagree.

In the wake of Searle’s influential paper, weak and strong AI underwent their own Trump-style rebranding, as people began to discuss differences in AI capabilities rather than AI consciousness – perhaps because the former is much easier to measure and moves the arena of debate from the philosophy department to the computer science lab. Weak AI merged with another term, narrow AI, to mean a system capable of human-level performance at a specific task. We have been living with narrow AI for decades, depending on how you want to define it, encompassing everything from spam filters to chess playing to driverless cars. In some sense, once something becomes achievable by narrow AI, it is just software – a phenomenon known as the .

Strong AI, meanwhile, gave way to general AI – a system capable of human-level performance at any task. A general AI is a highly desirable goal because it means that, rather than having to build a new system for every possible task, the AI would be able to transfer learning from one domain to another, just as a human can. In 2002, general AI morphed into “artificial general intelligence”, or AGI, largely due to (who would go on to co-found the firm DeepMind) popularising the term.

Most of the big AI companies, including DeepMind and OpenAI, were founded with the stated goal of achieving AGI. In the 2010s, this was seen as a slightly wacky and sci-fi research goal. Today, it is a business model, with trillions of dollars of investment riding on building an AGI in the near-future.

Despite this, definitions of AGI are woolly. defines it as “highly autonomous systems that outperform humans at most economically valuable work”. An agreement between OpenAI and Microsoft that was leaked in 2024 defines it as . OpenAI isn’t yet profitable, with expenditure far exceeding revenue, but that didn’t prevent its president, , from saying “” with the release of its Astra model in September.

However you define AGI, it isn’t the same thing as superintelligence. Bostrom defines this plainly in his book as “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest”. In other words, a superintelligence must be more cognitively capable than any human scientist, artist, philosopher or other cognitive performer that ever lived. It isn’t enough to be better than Albert Einstein or Marie Curie – you must also beat William Shakespeare, Leonardo da Vinci, Steve Jobs, Jane Austen and many, many others. Clearly, we haven’t achieved superintelligence. A real superintelligence would probably transform the world in ways we can’t imagine.

There is one problem running though all of these definitions, though, and it lies in the second part of AI: what actually is intelligence, and how do we measure different levels of it?

The standard answer is intelligence quotient, or IQ. Attempts to measure the IQ of various high-end AI models (by having them produce answers to IQ test questions) , although this is complicated by the fact that the tests themselves may have formed part of the models’ training data. Human IQ scores , often in young children because the scores depend on age, but higher scores are increasingly difficult to measure reliably. More pressingly, IQ is now widely considered to be a flawed measure of intelligence, rooted in eugenics, that ignores many aspects of human capability. Attempting to apply it to AI feels fairly fruitless, then.

As an alternative, some computer scientists have tried to come up with AI-specific intelligence tests. One of these is ARC-AGI, which looks not only at the capabilities of a model, but also at the cost of running it, meaning that less computationally intensive models can score higher. The test is now in its third iteration, and one of OpenAI’s latest models, GPT-6.1 Sol, . Does that make it AGI? I don’t think so, and I would expect a fourth version of the ARC-AGI test to come along soon, resetting the bar that models must exceed.

Does this mean we will never achieve AGI, and then superintelligence? Despite the certainty of the AI firms that we will – which is probably half real belief and half marketing – I think at this point, it is impossible to say. The meanings of these terms and our methods for evaluating them are constantly shifting. What seemed liked magic in 2022 – a computer that could answer you back – is now an annoyance in 2026, with Google’s AI Overviews often producing confidently incorrect information at the top of your search results.

Today, the most impressive achievements in AI seem to come from mathematics, but perhaps that will soon become passé. My feeling is that, if AGI ever comes, there won’t be one moment in which we declare it has arrived, but rather a gradual realisation that it is already here and we’ve just started calling it something else.

OpenAI mistranslated mathematics into code for its Navier-Stokes proof

When OpenAI announced its surprise solution to the Navier-Stokes problem, it produced one proof for humans and one for computers – however, they don’t match

Further AI reading:

• A software developer has used Anthropic’s Opus 5.5 model , releasing the code for free on GitHub. The legal implications are unclear, and while it is possible to , I would be very surprised if Adobe’s lawyers don’t come knocking soon.
  
• Nolla 91ɫ��Ƭ is piloting “” for acne treatment in Utah, using an app that asks patients to fill in a questionnaire and scan their face. Initially, two licensed physicians will review the AI-generated prescription before the patient receives it. In further trials, such reviews will happen only after the prescription is already sent.
  
• Google has , an AI app that generates playable games from text prompts. It is available only in the US, so I haven’t used it myself, but my own experiments with using Claude to generate games have seen mixed results – .
  
• Staying on games, a new trend has seen people “merge” two different video games by using AI to recreate the code of one inside the other. This has led to the likes of . It’s an amusing idea and a great viral video, though it isn’t clear how well these actually play – and again, legal questions loom large.
  
• ChatGPT has a new “” feature that allows the chatbot to answer questions using interactive visualisations, rather than just text. In my own AI use, I’ve found that having the model create a temporary website I can poke at is often very helpful, and this essentially brings that capability to anyone who doesn’t want to set up their own web server.