AGI Fact Checking

In developing CaaS (Classification as a Service) solutions for Snasci, it will become important to verify the integrity of any knowledge base.  Verifying the consistency and/or validity of a knowledge base will be a common practice regardless of the AGI provider.  Snasci itself will examine 3rd party KBs, however, structuring information and the maintenance of that KB is left to the 3rd party.

Snasci has a sophisticated inference engine which can delve deep into documents and their context.  Internally, Snasci constructs arguments that either support or reject a link between two or more pieces of information.  Thus, Snasci’s view of the world is constructed from blocks of known facts.  As new facts are added, these are decomposed into existing facts and serve as a chain of evidence to support that acceptance.

It is also possible for Snasci to take existing facts and derive new facts based upon this information.  It is therefore important for Snasci to re-examine the facts periodically and, employing dependency checking methodologies, mark chains as good, bad and questionable.  Snasci then attempts to define the implications of such changes and take appropriate remedy actions.

It may also be the case that a company has some form of proprietary information which extends Snasci, allowing it to determine the validity of given information.  For example, a company may have developed a new approach to lithography in CPU design.  Without an understanding of this new approach, Snasci may determine that any new facts are incorrect as it would be unable to decompose the problem based upon its existing KB.  Any such proprietary information is held in the client’s personal datastore and is not retained or distributed by Snasci in any way.

Let’s provide an arbitrary example of information retrieval from a third party data store.  Suppose we wanted to know who won the FIFA World Cup in 2010.  Snasci would recognise that this was a sports trivia question and would select an appropriate 3rd party provider.  It will then look up the JSON template for that provider and make a request similar to the following:

“snasciinstanceid”: “A14CC8F5-A104-45C7-9B09-93417860FB6F”,
“org”: “FIFA”,
“class”: “World Cup”,
“year”: “2010”,
“factrequest”: “winner”

The above JSON requests the winner’s information from the 2010 FIFA World Cup.  The response is sent back to a common endpoint which Snasci will then route to the requesting client.

We could also request the associated evidence chain, that is, all the supporting facts and sources that provide this result.

“snasciinstanceid”: “A14CC8F5-A104-45C7-9B09-93417860FB6F”,
“org”: “FIFA”,
“class”: “World Cup”,
“year”: “2010”,
“factrequest”: “winner”,
“reportevidencechain”: true

Snasci would then be able to parse this evidence chain and, based upon its own internal information and/or evidence from 3rd parties, confirm or deny the fact with a given confidence level.  This is a basic consensus model, however, consensus has its problems.

As an example, let’s take the statement ‘Jesus walked on water and it was a miracle’. In a consensus model, this may come back as fact, especially if the demographic was mainly Christian.  Snasci eliminates this issue by examining each assertion against well defined sources of truth, such as scientific fact.  In this scenario, the statement ‘Jesus walked on water and it was a miracle’ would be rejected because the laws of physics state it is impossible.

But how do we approach more complex forms of false information?  Take this next image as an example:



This is probably more complex that it seems initially.  Firstly the image is obviously fake, as it is taken from a 2014 music video, but the idea captured in the tweet is generically accurate.  This is a complex straw man argument, in that a straw man is constructed through a fake scenario, then knocked down as being fake, but the fact remains that such incidents would be common in a war zone.  This is an attempt to support further conflict in Syria, disguised as an attempt to manipulate against it and then counter that position.

The idea here is that everyone gets distracted by the straw man, meanwhile people are dying.  Further, when people do encounter such images on the web, even if genuine, they are in no position to determine if they are accurate and mentally switch off.  The attacker has achieved their objectives.

Let’s look at how we could discover the image is fake:

  1. Project the scene into 3D then examine against radar imaging for the Aleppo area.
  2. Construct DBs with every frame of video and image on the on the web and compare.
  3. Examine details in the image and compare against what is known of the Allepo area.
  4. Randomly spotted by a user who also happened to see the music video.

What should be clear here is that any automated approach would be costly and outside the realms of feasibility for the majority of companies.  Further, entrusting a final analysis to a military/intelligence department may result in inaccurate information being supplied to support a propaganda campaign and/or psychological operations.

What does an AGI do here?  Does it reject all such reports outright?  The key problem with post-analysis is that the goal of the attacker has already been achieved.  That said, if genuine, it is important information that the public needs to know.

All said though, do we really need a photo or video to make an obvious fact of war real for us?

With respect to informing the public, it may be best to evaluate such material prior to release.  As such events will occur in conflict, the public is not being misled, but rather prevented from being misled.  If it takes photos or videos to swing a person’s opinion of war, this type of person is highly suggestible and should not be making any important decisions.

How do we determine if an image, tweet, post, etc., is part of a coordinated campaign and/or complex straw man arguments?

In many cases this comes down to visibility of web content.  Seeking patterns in the posting behaviours, content, etc.  Much of this is moving into the world of intelligence gathering, a place where no business wants to be for a host of different reasons.  Increasingly though, the web is being filtered and edited in real-time by automated systems, so the ability to objectively confirm behaviour across the web is diminishing with time.

There is also the question of how to deal with sites that either take the bait, or are acting in concert with attackers.  Firstly, how do we tell the difference?  Is there a need to tell the difference?  Secondly, what is the appropriate response?  How do we tag these sites and inform users?  Damaging a business simply because it made a mistake is a harsh penalty.  What must be recognised is that online businesses are coming up against state actors and they do not have the training, resources or capability to deal with this.

Another form of threat is implied threats, statements and broader psychological operations.  This type of attack will pass through any type of filter undetected as the information is highly personalised and masked as normal web content.  Content with double, triple or more meanings is becoming an increasing issue.  On the surface an article may speak of stolen cars, however, when read slightly out-of-context the receiving party may see an implied threat or thinks that someone knows something they did.  It doesn’t even need to be an actual attack either, many mentally ill people will first interpret new information in the context of themselves, before shifting to the context in which the article is written.  This is the source of claims such as ‘My TV was talking to me’, or even developing fixations or false beliefs.

Snasci can help protect against such attacks and even provide the mentally ill with a safe online environment.  Snasci can analyse the context and potential alternative contexts of online media and filter content deemed to match personal references.  How well this works depends on your level of honesty with Snasci, as it can only match content against what it knows of you.

Snasci can also seek out other embedded patterns in content to prevent any form of reference entering the public consciousness that could be employed to detrimental effect.  Such embedded references increase the suggestibility of an audience.  Content creators will be able to discuss this with Snasci and fully understand any action Snasci would take prior to committing resources and funding.  It is expected that there will be some form of dynamic response as new threats are detected, but Snasci will be able to explain this clearly.

How does Snasci deal with scenarios where the only indications of an issue/event is rumour and/or assertions?

This is a rare event, but it does happen.  Sometimes people report things online which are later found to be accurate, but for which no evidence exists at the time.  In these scenarios, Snasci basically opens a case and adds evidence as it discovers it.  This is quite different from the current approach of content getting buried, forgotten about and/or ignored.  For example, one or more people claims that a new product is causing illnesses but no evidence for this exists.  Rather than falling into the abyss and festering as a conspiracy theory, Snasci will note this and seek out additional evidence until it comes to a complete conclusion.  In the case where it is a product, as a courtesy, Snasci will pass collected information back to the company in question for their review.

Ultimately, fact checking is a hard problem as much of it requires a knowledge of the future.  This is especially true where state sponsored attacks are concerned and operations like this are planned to extreme levels of detail employing the latest in scientific research.  A proper state actor attempts to leave nothing to chance, their operations are designed much in the same way as Intel/AMD would design their latest generation processors.  Snasci will be fighting an uphill battle and there will always be some vectors that hard to defend against, however, the return on investment will dramatically drop with time which will eventually make these methods and approaches financially unfeasible.

Image Source:

Joe Loong


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s