Problem Statement: People send messages whose content is unreliable which can lead to fake news being spread by miscreants.

Current situation: Anyone can draft a message and send it to n number of people who can in turn spread it to n number of people. There is no way to control spreading of any kind of fake news/message.

Approach:

We want to know which message is fake and stop it at the source and from being forwarded. Thus, there are two points at which a check for fake messages can be made – when a user sends a message and when a user forwards a message. If a message is found to be fake, then disable the send button and pop-up a message saying, “This message appears to be fake”.

How to figure out a message is fake? –

To begin with we will try to filter out fake messages related to politics.

This will need to be a ML solution as the application will need to be trained to identify fake messages.

1. Check for politically aligned word in the content.
2. If such a word is present the application should be able to do a web search (across reliable news channels) matching the content.
3. If the content is found to contain data not matching any of the data found on the web then this should be marked as fake.

What to do with users sending these fake messages? –

1. Check for user profiles of these users to see if they match any valid govt identity holders. If fake bar this user.
2. If a user sends fake messages more than five times, then they should be barred on the platform.

Metrics to test whether the feature is working:

1. % fake messages being filtered
2. % true/real messages being marked as fake
3. Reduction in speed due to additional check.

#1 should be at least 60% for the MVP and #2 should be no greater than 10%. #3 should be no greater than a few 5-6 msec otherwise it will hamper the user experience.

The problem can essentially be broken down into two parts: 1. The problem of identifying a fake message when it is sent, and 2. The problem of curbing it’s spread. I’ll try to tackle them separately.

Identifying a fake message

Let’s define what a fake message is. A fake message is a message which contains information which can be definitively proven to be false. While it is easy to determine that a message like “the sun rises in the west” is fake because it can be tested against common knowledge and/or scientific facts, other messages which make claims about events but have an agenda are harder to disprove since there may not be a single verifiable source of truth.

Our approach to defining a message as fake will thus be to assign a score to it and mark it as fake if it crosses a certain threshold. I can foresee 3 ways of doing this

1. Crowdsource the ability to mark a message as fake: Probably the least expensive way to build a fake message identifying system, we can rely on people to mark a message as fake. A user should be allowed to take this action on every message. If a message gets marked by enough people, that message can be then categorized as fake. Whatsapp will also have to keep a track of the forwarding of messages so as to assign the score to the entire chain of messages that have been forwarded. The downside of this approach however is that a majority of people believing something is fake doesn’t make something fake.

2. Use third party independent fact checkers: Facebook and Instagram are already using third party trusted independent fact checkers who will manually verify if a message is fake. This system is obviously expensive to implement and doesn’t scale well but it can be used in conjunction with the above system where if a message is being flagged by too many people, then the fact checkers can be involved. Also since this system is already implemented in Facebook and Instagram it should be easy to onboard onto Whatsapp.

3. Web crawling: For messages that can easily be tested over the internet agaist a single source of truth, a way to crawl the web to check the claim made in the message can be used. The goal of this feature is to catch the easy to identify fake messages. This can also be extended to images to check if the image being circulated is a morphed version of a popular image on the internet.

Curbing the spread of fake messages

When a message is identified as fake, we can easily trace the message back to the original sender and warn the sender for sending a fake message. We should allow the sender to have an appeals process like implemented on most social platforms. If a particular sender has a lot of reports against him or her, we can debar that user.

For users who forwarded the fake message unknowlingly, we can inform them separately to let them know that what they’ve forwarded is fake and to be more careful next time. A guide on how to check messages could be useful.

Once a message is marked as fake, it should show up as so on the UI of every recipient and no user should be allowed to forward the message further.

Challenges: The overall challenge will be that this won’t be a foolproof method. Also satire, politically incorrect opinions etc might get marked as fake. Keeping humans in the loop will ensure that we don’t toe the line of freedom of speech.