Falsehood flies, and the Truth comes limping after it

https://quoteinvestigator.com/2014/07/13/truth/

This is an early draft of an idea that’s been swirling around for too long and needs some air. It’s a mess right now but having thrown the clay of the idea on the wheel, I can start to work it.

There’s a bit of it here but it needs fully working out in the open.

The Problem

Lies, I presume, emerged very soon after the speech innovation. What makes Fake News any different? I’ll argue it is a toxic cocktail of factors:

Viral Media: content spreading with perfect fidelity with a single click but little accountability. Engagement Engines: content that stokes the emotions wins. Complex, sophisticated content.

Viral Media

Social Media that is engineered specifically to spread content exponentially is best described as Viral Media. Facebook, Twitter and WhatsApp lead here. Content comes with a personal endorsement but quickly loses its provenance.

Even with mass media, it was possible to assess the source. A known organisation was required to publish news and was publicly responsible for every word in every copy. It mattered whether the headline appeared in the Sunday Times or the Sun.

Social media has massively obscured that attribution by becoming the platform where news in consumed and propagated. Faking a screenshot or video — hijacking the credibility of the original source — is trivial whereas verifying the truth of a claim is relatively hard. Organisations and technologies which are slowly and carefully distilling complex truths are no match for simplistic, emotive lies with the rocket of recommendation systems under them.

Youtube and Facebook generally don’t assess truthfulness: engagement is the thing. As every engaging storyteller since the birth of language has known, base emotions are the way to go. Fear, rage, lust and kittens rank highly. Truth is a niche interest, mostly covered by Wikipedia (spoiler: giving a clue to a solution).

Given the dominance and self-sufficiency of the social media platforms, any solution which doesn’t directly intervene in the viral lifecycle is doomed to failure.

Twitter rebutting Trump https://www.theverge.com/2020/10/11/21511682/twitter-disables-sharing-trump-tweet-coronavirus-misinformation

Engagement trumps truth in social media where it counts.

Moreover, their power and their dependency of the powerful on them makes them resistant to change.

Evaluation by Inspection increasingly failing

In a world of increasing complexity and shrinking attention spans, truth is becoming harder and harder to determine. Screenshots are the root of all evil. Images and video cannot be trusted.

Even before the latest generation of AI systems which produce convincing images and videos to order, an image could be captioned, photoshopped and taken out of context to any end. Generative AI makes this this not only trivially easy to produce but also threatens to make it ever more compelling. With recent developments in generative AI, the faith in being able to evaluate the validity of some content merely by close inspection is becoming rapidly untenable.

Misplaced Trust

When reality is shaped by the social media channels you subscribe to with little regard to source then you end up with a deeply fractured social reality that is largely pandering to people’s ready biases and hot buttons.

Does this mean that there is no hope? No, since a little trust can go a long way.

Truth is a very nuanced and qualified thing which cannot usually be simply reduced to a Boolean. The classic process of quality debate is still the only path to sketching a best approximation to the truth. An essential dimension to such a debate is trust. Is this source trustworthy when they assert new information into the debate?

We are looking for a trust minimisation system. One where the amount of trust we have to take is reduced to sources proven most trustworthy. Truth is your best guess when you minimise trust. It only works when you subject your personal hypothesis to the same trial — otherwise you venture into paranoia.

You trust the coffee barista not to poison you because the hypothesis that she might just popped into your head without evidence. Everything you have seen indicates that this is not going to happen. The same for faked moon landings or a flat earth hypothesis: you’d have to trust small community of rebels with scant evidence against the overwhelming consensus of professionals. Not to say the popular opinion cannot be wrong – medical opinion for some time was that the brain was an organ for cooling the blood. But that held because there was an abundance of trust in authority and scant evidence.

General form of Occam’s razor?

A Proposed Solution

Lay summary:

• Everything that is published has a proven author or is anonymous and untrusted. Proven authorship.
• Everything that is said about a publication can be seen by anyone. No bubbles.
• References are filtered. Trust is earned.

Some terms I’ll be using:

• Post – an image, tweet, blog post. An immutable block of content with a definite publisher. It may reference a number of other posts.
• Publisher – a person or organisation that is responsible for producing a post.
• Platform – where the Post is made available. Responsible for linking Post to Publisher based on URL, blue checkmark or whatever.

Attribution: proven authorship

The first pillar is in verifying authorship. This is an established technology with public key encryption. This can prove that someone said something, a news organisation recorded some media. Sign everything.

Currently, I have to verify at source: did this person really tweet that? Has this image been tampered with?

By hashing the content and publishing the record – perhaps on a blockchain – you can separate a dodgy screenshot with an authentic post. Of course, this doesn’t prove that the post is true. That’s where the next pillar comes in. Digital signing – hashing – blockchain. Unlike a webpage which can be edited and a tweet can be deleted/denied, once seen, it can never be denied or taken back.

• Truth ultimately boils down to trust. If you don’t have direct experience (trusting your own senses) of a situation then you are taking what you are told in the context of the teller. In an increasingly complex world where so many people can say anything and make it look good, it requires real insight and effort to evaluate stuff.
• When people get much of their information
• Social media has massively obscured that attribution by becoming the platform where news in consumed and propagated. Faking a screenshot or video — hijacking the credibility of the original source — is trivial whereas verifying the truth of a claim is relatively hard. Organisations and technologies which are slowly and carefully distilling complex truths are no match for simplistic, emotive lies with the rocket of recommendation systems under them.
• The trust granted to an image derives from the authorship. An anonymous image might provoke further investigation but it has zero reliability otherwise. Video is set to go the same way very soon. Screenshots too. The New York Times has introduced the News Provenance Project to this end.
• However, if this attribution is a central element of how the post is presented then it has little impact. Attribution enables global debate.

Universal Debate: no bubbles

• This was the factor I addressed in my earlier post, fantasizing of a social media re-engineering which optimised for debate rather than positive feedback.
• Generalised to Internet scale, you’d be able to see comments on any image posted to the trust chain. Since those comments would also be on the trust chain
• Everyone would be referring to the same image – whatever its filename or URL.
• Comments on the image that were posted to the trust graph would be retrievable, with authorship, to get the entire debate originating from that image.
• For any post with any broader interest, the scale of this debate becomes a problem as anyone who has slid down the thread of a widely seen Twitter post can testify. When wide interaction creates this problem it can also be turned to address it.

Universal addressing. If a post has a universal name – a hash – that is irrespective of where it is found then anyone can address it from anywhere. It could be an image or a tweet or anything else with a fixed format. Once it’s hashed – named – then it can be targeted for debate. It can be debunked or endorsed from any platform. Moreover, publishers can reference a thing in their post. A post would be unable to launder itself on any reputable platform. A universal commenting system. Merkle tree. That’s a lot of comments. Potentially unbounded. Fortunately, there’s an inspiration for distilling authority from a network of interconnections too. Google’s Pagerank.

It can be more personalised and explicit, of course. You get to choose who you trust but you also get to see how trustworthy they are.

Reputation. So we have posts and we have people talking about posts. There’s one more step: having publishers say something about other publishers (this is why everyone is a publisher), i.e. who they trust.

Maybe you trust a news organisation? If the news organisation says that a politician said a thing then you probably don’t feel the need to check with the official office of the politician. If your opinionated Uber driver said that a politician said a thing then you might want to check that.

The rough form for this rough sketch is a directed acyclic graph that doesn’t tell you directly what is true. But it can tell you who said what and what everyone thinks about that and each other. It’s a mechanism for grounding trust. So I have come to call it the Trust Graph – and I’ll be motivated to write properly about it once I hit Publish.

Block-level semantic web with reputation. The killer app for Web3?

An illustration?

• An item is published, say a photograph. We’ll call it i.
• Without attribution, the item has zero trust. So we’ll assume it’s signed by a publisher p which imbues it with a degree of trust to be determined later.
• If it’s a photo then it should have a caption to contextualise it. This is actually a new item that references i its hash and states the natures of the reference.
• Anyone can reference any item in the same way, creating a tree for that item. An item can include multiple reference, creating DAG. Acyclic since an item cannot know in advance the hashes of items that might reference it (or intermediate).
• A reader / evaluator would ideally be able to read and consider everything that has been said about i and the publishers of those to form their judgement. More likely, they will just consider a subset of those items. How to we support this selection? Ranking? Whitelist? Trust profile. Is it a UI issue? Partly.
• Aggregate trust profile. Like page rank.
• A publication can be as minimal as a like. Should be aggregatable and machine-readable.

Any intervention that does not take place in the viral sharing cycle is pointless.

• Probably the most significant challenge is the closed architecture of sharing platforms like Facebook, Twitter, etc.
• Create a Twitter example, maybe from led by Donkeys. Cannot currently prove you didn’t tweet something. Cannot have a mechanism for verifying you didn’t post something without enabling the converse. Privacy is not reduced if someone already has a copy of a tweet. Screenshots are the root of all evil.
• Behoves sharing platforms to maintain content integrity by hashing and fixing them. If a piece of media contains an attribution link then this can be reported on the platform. If not, then this fact can be clearly shown with the media.
• I’m hopeful that Facebook will adapt because the problems it is experiencing now are only going to get worse otherwise.

Reputation