AI slop will eat itself

Introduction

This post is a rare excursion outside of the reassuring confines of spreadsheets and other office software, although there is still an important connection. AI is very quickly becoming an integral part of all office applications as well as a key part of everyday life. Attitudes to AI in general might well have an impact on the perceived benefits and reliability of the use of AI in specific applications.

The camera always lies

There’s a great deal of angst around the difficulty of knowing whether what we are looking at is true or a fake. Throughout recent history, states have been adept at controlling the messaging to their own citizens. Whether it really originated with Abraham Lincoln or not, the first two phrases of the quote: “You can fool some of the people all of the time and all of the people some of the time, but you cannot fool all the people all the time.” acknowledged the ability to do so in the mid-19th century. In the new AI age, almost anyone can generate content capable of deceiving us. Perhaps the effect of AI is to bring about the democratisation of the ability to mislead. AI extends to everyone the ability to convincingly mislead, though hopefully not to mislead all of the people all of the time.

Pushing water uphill

The obvious fear is that AI fakes will make it easier to fool some or all of the people for whatever reason. There is already considerable evidence of state and other actors trying to interfere with elections by spreading false information.

The immediate reaction to this seems generally to have been to try and persuade the tech gatekeepers to remove misinformation from their platforms. This has been conspicuously ineffective. A moral dedication to free speech has the coincidental benefit of also being good business. More posts generate more income, less moderation saves costs. Efforts to persuade platforms to make their content more reliable are at best ineffective and at worst counterproductive. Pretending that content can be regulated only reinforces inappropriate credibility.

Go with the flow

It could be argued that we risk creating the worst of all possible environments for the recognition of the truth. We suggest that content can be made to be reliable whilst being completely unable to make it so. Of course, the ideal would be to ensure that we are only exposed to verifiably accurate information. The current situation where there is still some expectation of accuracy might be the worst-case scenario: we might choose the wrong content to believe in. It might be better if we knew not to believe anything without being intensely critical and sceptical.

It would be interesting to know what would happen if we tried another approach. Much like a weedkiller that works by disrupting the way that a weed grows to make it develop in an uncontrolled way and kill itself, perhaps the better plan would be to acknowledge the impossibility of removing misinformation and seek to dramatically increase misinformation instead. Rather than encouraging people to report misinformation to organisations whose income is based on its prevalence, we could encourage people to counter it with equal and opposite misinformation. Instead of railing against fake content, we could neutralise it by drowning it in a sea of slop. Part of the problem of AI generated misinformation is the readiness of many people to believe what they want to believe, regardless of any evidence. By making content increasingly ridiculous, there is a chance we could make it harder to generate faith in the faked. It could also counteract the powerlessness we can feel in the face of big tech and malicious actors.

To take a practical and current example, as I was writing this article, there was news of multiple AI fake videos of leading politicians storming out of an interview with Laura Kuenssberg. Were we to want to weaken belief in the validity of those videos, is the better approach to try and remove all of them, in the almost certain knowledge that we will fail and thereby possibly add to their credibility, or would it be better to make all such videos seem ridiculous by generating more fake content showing politicians storming out of videos because of disagreements over jelly bean flavours or the likely victor of a battle between a giant lobster and a T Rex? Perhaps we are looking for the real-world equivalent of the Harry Potter Riddikulus spell.

Conclusion

New technologies often develop in unpredicted ways and have unforeseen consequences. In the face of this uncertainty, obvious ways to deal with threats might just exacerbate the risk and we might need to consider more radical approaches. 

Returning to spreadsheets, anything that generates healthy scepticism about AI output, whilst still recognising its enormous potential benefits, might not be a bad thing.

Additional resources

As well as the solution to many specific Excel and office software problems, you can also find plenty of articles on AI in the Power BI based ICAEW Excel archive portal:

The 'Open in full-screen mode' icon in the bottom right-hand corner of the embedded report should show the contents at a more readable size with the Escape key returning you to the post.