Superhuman Bullshit

In the age of Superhuman AI, the ticking time-bomb at the centre of the Internet is Superhuman Deception.

Images of family events that never happened, fabricated scientific articles, fake people, messages and video conferences can all now be generated on demand. In a world where over 90% of content is generated by AI, a second biblical flood has begun; the flooding of the internet with spectacular bullshit.

The Dead Internet

Bullshit is a greater enemy of truth than lies. Disinformation is a mere rounding error on the internet compared to the vast swathes of bullshit. While respect for the truth is amongst the foundations of civilisation, the accelerating flood of spectacular bullshit online has created The Dead Internet.

The Dead Internet is a wasteland, killing all meaning and social value across a growing number of social platforms, from global platforms like Facebook to niche online tech communities. The remaining survivors lament the proliferation of click-bait and bullshit content. The Internet doesn't deserve our attention anymore.

While Search was meant to help everyone discover signal in the noise, it instead undermined itself. By prioritising engagement farming to keep us online, search engines neglected their core mission to discern quality from quantity. The Bullshit virus quickly infected Google and other prominent search engines, where finding an answer became a hostile, losing battle. Things have gotten so bad that Elon Musk banned engagement farming on Twitter last month.

One authority figure cannot cancel bullshit while preserving free speech online. The power to detect and opt-out of online bullshit belongs with the end users of the Internet.

Model Collapse and Peak Internet

Human attention has become a scarce resource online, as have humans themselves. Peak Internet may be here. As Superhuman AI models run out of human data to train on, AI researchers are feeding them more and more synthetic data, causing some models to collapse or go “MAD”, or Model Autophagy Disorder, a term coined by AI researchers to describe how the quality of large AI models trained on synthetic data can collapse as later generations of the model are trained on synthetic output.

But MADness is not confined to AI models. It is spreading across the Internet as a whole. Our ability to discover the information we need when we need it, and to distinguish fact from fiction, is under direct threat.

AI models may need us more than we need them, but our need to understand the world around us is permanent. Search engines have failed to help us avoid bullshit and deception. In the Age of Superhuman Deception, we need defensive AI that can identify deception and the people behind it.

Big Stick Diplomacy

Online Safety regulations and AI Acts are necessary, but the UK and its allies must also invest in Big Stick Diplomacy. “Speak softly and carry a big stick” is the only means of active defence. For a prize as big as the Internet, active regulation may be required too. Citizens and taxpayers reasonably expect public spaces to be kept safe and well maintained. The Internet is the largest public space of all, and the infrastructure for all modern economic activity. Building national capabilities to identify and counter superhuman deception is now a matter of public trust.

As part of his resignation note last week, departing OpenAI alignment lead Jan Leike posted on Twitter that “We are long overdue in getting incredibly serious about the implications of AGI”. His words follow a string of high-profile resignations from top AI institutions around the world in the past year, including the so-called Godfather of AI Geoffrey Hinton from Google. These researchers shouldered huge reputational risk and left some of the highest paying jobs in industry to raise the alarm.

They did. We should listen.

Mutually Assured Destruction: MAD Again

“If it gets to be much smarter than us, it will be very good at manipulation because it would have learned that from us. And there are very few examples of a more intelligent thing being controlled by a less intelligent thing”

Geoffrey Hinton

MAD in the context of Superhuman AI means one thing, but in the context of Defence and National Security its meaning is even more significant; mutually assured destruction. MAD describes how maintaining a globally active, defensive nuclear arsenal ensures our safety from adversaries. The same can be said of many weapons. The UK builds and procures these as a fundamental part of its storied tradition in Big Stick Diplomacy. That tradition of deterrence now needs to cover the internet.

MAD applies to superhuman deception because whoever builds the most deceptive model first, wins. And in the absence of superhuman defensive AI, the internet will cease to exist as a source of social value and collective wisdom. Netizens will lose trust in governments to protect public goods and services and to hold bad actors accountable.

To protect it from Superhuman AI deception the internet needs a defensive AI arsenal of its own. Hostile states are building theirs already. Many of our most valuable public spaces have become digital, and our ability to discover information and distinguish it from bullshit is fundamental to keeping them worth using.