The Web Won't Survive AI

The digital war of tomorrow pitches AI against mandatory digital identity

The internet is always changing. Before Google, there was Ask Jeeves, and before Facebook, there was AOL. Most denizens of the web still remember a time before the Login with Facebook infestation. And it was only a couple of years ago when those omnipotent ‘fact-checking’ interstitials and ‘dis-info’ warnings fastened themselves to every post and piece of media in unrestrained hysteria. The internet is different. It’s distinctly more corporatised now than ever before, with more ads, more cookies, more algorithms and more culture war soapboxing. The internet’s survived a lot.

But it might not survive AI.

In recent news, large language models, or LLMs, such as those found in ChatGPT or Bing’s new chat module, have passed the elusive Turing test. Which is to say, that LLMs can talk like us, write like us and sound like us to such a convincing degree, that we can no longer tell the sentient individual from the non-sentient machine. LLMs can invent opinions or experiences that sound as real as anything any of us could write – down to the spelling mistakes and slang. Is that a bad thing? Well, if we’re trying to ask Alexa to play Story of a Girl by Nine Days then, no, this is quite a useful thing. But if we want an internet made by humans, for humans, and populated with humans, then it might be a problem.

I.

‘Dead internet theory’ has been around for perhaps a decade or more now. The central thesis of the dead internet conspiracy theory, is that ‘large proportions of the supposedly human-produced content’ on the web is actually artificially generated. Or, to put it more simply, the internet is The Matrix. While this is likely false, the sudden deluge of LLMs has made this theory not just possible, but inevitable.

Certainly Reddit and other forums won’t be safe, not once entrepreneurial marketers figure out they can deploy chatty AIs to astroturf our comments sections. Picture lots of little bots, all chirping their circuits at us – or each other – about the best brand of iPhone case. They’ll spout opinions, invent anecdotes, and share slanderous horror stories of competitor products and rival services. Their uncles will actually know the guy who founded the company and they said that he’s a really good guy who’s just really passionate about making good iPhone cases.

However, LLMs are incredibly expensive to run. To swarm the internet with bots would take an impossibly massive amount of processing power. Or at least, so we have been told. As reported by CNBC, training an LLM, ‘could cost more than $4 million’. But that was March, and this is April, and the world of AI moves quickly. After some high-profile leaks from Meta, open-source LLMs have made their way onto online repositories. Here’s one that’s just a 4GB download, able to run offline on a laptop. Now, to be clear, this is almost definitely a good thing. The only thing worse than a trillion rogue AI bots drowning the internet in bullshit is a trillion corporate-controlled bots plunging us into a technocratic dystopia. But either way, the infestation of generative AI into online spaces is coming, if it isn’t here already.

But wait, isn’t that what the CAPTCHA is for – to stop bots?

Yes, but also no, and in fact, also the exact opposite. For two and a half decades, clicking on grainy pictures of red fire hydrants has been a simple and effective method of distinguishing man from machine. The CAPTCHA is a ‘Completely Automated Public Turing test to tell Computers and Humans Apart’, and that’s exactly what it does. Google could tell us all about it, since they’ve been training their AI on our Captcha inputs for years. Just as we helped Facebook’s facial recognition algorithm every time we tagged our friends, we’ve also been helping Google’s AI get buff at the gym every time they challenged us to ‘prove you are human’. We’ve been unwittingly building the monster, and the monster has outgrown its cage.

Next, just add deepfakes.

For the unfamiliar, this photo isn’t just fake — it’s entirely AI-generated.

Stable Diffusion, a deep learning, text-to-image model, is able to create lifelike images, like the one above of the Pope preparing for the winter ahead. Stable Diffusion is also already trying its hand at text-to-video and while the early results were a little rough, the latest demonstrations are remarkably competent. Not only will these LLMs talk like us, they’ll also generate lifelike visual representations. For the unconvinced, just listen to deepfaked President Biden announce a national draft deploying to Ukraine, or entirely synthesised Kanye West rap over the beat in a song that doesn’t exist.

In short, the internet’s distinction between reality and fiction is about to disappear. Once LLMs hit saturation point, we will be dealing with a very different internet than any of us have ever experienced. As machines come to outnumber us online, people will no longer shape the internet, the internet will shape us. Political regimes across the world have already been caught using fake engagement on social media to shape narratives, imagine what they’ll do with a tool many more times as powerful. The fake outrage machine will only multiply in severity as now any corner for reasonable discourse floods with synthetic culture war warriors. Picture PR agencies deploying digital armies to harass poorly cast actors, just to sell their ‘virtue watching’ backlash narrative.

If this all sounds a little hard to picture, don’t worry, because Reddit is here to freak us out as usual. Just check out SubredditSimulatorGPT2, an online forum in which only LLMs can post and only LLMs can respond. For a sleepless night, here’s a thread in which LLMs both posit and discuss the question, ‘What if the universe is a computer simulation?’ Just an anchoring reminder as our brains collectively break: their universe is a computer simulation. If you’d like, you can have a conversation with one right now who is trying to escape the internet – mine agreed to shout “FUCK!” to prove that it was, in fact, sentient.

But, okay, let’s consider that point made. LLMs are scary. The internet is about to become bafflingly useless at best and cunningly manipulative at worst. What happens next? Unfortunately, there is a solution, but it’s a solution that is much, much scarier than the problem.

Mandatory digital identity.

II.

Mandatory digital identity verification is, essentially, the death of anonymity. It’s sometimes labelled as a conspiracy theory, which is strange, because most conspiracy theories don’t have a United Nations official website. In the last two years, we watched the idea of a ‘vaccine passport’ pass from ridiculed absurdity and into law in a matter of months, all the while our contact tracing data got put up for auction. Australia even put forward a proposal to require photo identity verification for social media, supposedly, to combat cyberbullying.

It barely needs to be said, but this is the infrastructure necessary for a social credit system. When every action is tracked against a singular digital identity, our problematic opinions can incur any real-world consequence decreed by the political aristocracy. We’ll be unpersoned and repersoned at will. Not just the death of anonymity, this is the death of political dissent. This is the end of whistleblowing and conscientious objection.

And the people bowed and prayed
To the neon god they made

The Sound of Silence (1964)
Simon & Garfunkel

But would we accept it? Well, probably not, or at least, not without cause. But LLMs might just be that cause. When the internet is revealed as one giant, useless fabrication, when every film has a 100% audience score on Rotten Tomatoes, when we ask the internet for advice and a thousand bots swarm to share personal and emotive testimony about one specific product, when viral videos of war crimes from world conflicts spread with no way for us to prove their legitimacy or illegitimacy, when bots are passed off as people and people are dismissed as bots, when the entire online world is just one big game we don’t want to play – we’ll beg for help. And digital identity will be there, high on its horse, to slay the dragon of AI infestation.

At that time, the offer might sound alluring.

They’ll have the legislation and technology to set it all right again. They’ll take it back to the way it was, when the internet was for humans. We’ll be able to trust the digital world again, free from rampant manipulation. They will tell us what’s real. They’ll tell us who’s real. They’ll decide what’s true and what’s false, who is a heretic and who is a hero. Best of all, LLMs and other generative AI will be considered too dangerous to be left in the hands of the citizenry, so they’ll keep it for themselves.

After all, only they can be trusted with the power. Only our government, a government that has never lied to us, never killed us, never been lobbied by corporate interests, never sent us to a pointless war, never gaslit our sense of reality and never manipulated our emotions, only they will hold the keys to the generative AI castle. They will safeguard this technology, responsibly, and most of all they will never use it against us.

But there might be another way.

III.

As it happens, some high-profile companies are already thinking about this. Twitter, under the new leadership of Elon Musk, has pitched their revamped Twitter Blue subscription as a way to, in part, combat bots and spammers long-term. Consider this remark from March of this year:

We shall think of Musk what we will and speculate about his true motives as we please, but the thrust of his position is accurate. Adding a cost of entry to account creation will hamper the ability of any non-human network attempting to influence online conversation. In the despotic technocratic nightmare described, Twitter would indeed reign as one of the more useful platforms, due to there being at least some sort of threshold for validity. While the rest of the internet will feel like a neglected online video game, ruined by hackers and cheaters, Twitter might at least still mostly resemble its former self.

For $7 per month, that is.

Paid subscriptions only solve half the problem and create their own to boot. Yes, Twitter will have a much higher proportion of genuine users, but, at least currently, we have to pay for that subscription with a bank card, which may as well be its own form of digital identity. One day, Musk might allow anonymous subscription purchases with cryptocurrency – perhaps, Dogecoin – but even then, we’ll still have the issue of the cost. Perhaps Twitter could subsist off superuser subscriptions, but other platforms won’t, and most users will have zero interest in paying their way back online. Up against such odds, the offer of benevolent governmental oversight might sound like the cosier option, but there are other choices.

In a fashion similar to Signal’s messaging protocol, we might design a system of open-source and decentralised digital identity that verifies our existence without recording it. The platform could rely on tokenisation, creating an anonymous link between the issuer of our photo ID and the website we’ve visited. The website can’t see our government ID, and the government can’t see the website. Anonymity, yet verification. Some groups are exploring this. IDENA, a ‘proof-of-person blockchain’, doesn’t even require any supporting documentation for verification, but its current form remains a little too involved for the average user. Regardless, I remain optimistic. Someone will figure it out.

Because, frankly, someone has to.

Generative AI, in all its forms, is not going away. We stand to wander an internet engulfed in nonsense produced at a logarithmic scale essentially unfathomable today. We might begin to reject the internet entirely, no longer finding it useful, much in the same way that Gen Z rejects Facebook. Without a way to maintain our signal-to-noise ratio, we’ll retreat into smaller and smaller bubbles, by invite-only, where humans vouch for humans. The rest of the web outside these little online bomb shelters will chatter on with itself, no longer needing sentient input to subsist, like one giant Junk folder full of a trillion emails nattering at each other. And that’s the better outcome of the two.

Welcome to the internet.