Forum

How are we protecting the Internet from the flood of fake news, powered by Transformers?

  • Andrei Petreanu profile
    Andrei Petreanu
    19 August 2019 - updated 1 year ago
  • 6 comments
    Total votes: 4

Natural Language Generation tasks have the power to produce amazing human-form text on any given topic, and were considered,up until recently ,to be very hard tasks to manage. Transformers and Attention models changed that, and the first ones to give us a taste of what can come were OpenAI : 

"Better language Models and their Implications" : https://openai.com/blog/better-language-models/

Immediately after, as a response, Google showed off BERT (that incredible "make-a-rezervation" demo from Google I/O), and BERT is now considered to be the state of the art in Natural Language Understanding.

Google Bert and OpenAI GPT-2 tech was open-source (in terms of research, stats, small-pretrained models, code) but the entire power of those NLG systems was kept hidden from the public, to protect against what could become a 24/7 flood of generated fake news on different topics.

Last week, NVidia released Megatron Language Model, the biggest trained language model of its kind, with incredible generative capabilities. Link here : https://nv-adlr.github.io/MegatronLM. "largest transformer based language model ever trained at 24x the size of BERT and 5.6x the size of GPT-2"

This was a full public release, including code and train-protocol

That means anyone who's got a lot of GPU power can now generate infinite human text on any given topic, resembling something of a news-bomb :)

Snippet from OpenAI's concerns on this matter :

"We are aware that some researchers have the technical capacity to reproduce and open source our results. We believe our release strategy limits the initial set of organizations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems.

We also think governments should consider expanding or commencing initiatives to more systematically monitor the societal impact and diffusion of AI technologies, and to measure the progression in the capabilities of such systems. If pursued, these efforts could yield a better evidence base for decisions by AI labs and governments regarding publication decisions and AI policy more broadly." May 2019

 

So what OpenAI has said in May is now public and available for everyone in April, by NVidia.

Are we not concerned? I'm training these models for work, as a Tech-Lead / Consultant.

I am very concerned.

Tags: 
#fakenews #bert #transformers #public

Comments

Log in to post comments
Maximilian J. Haarich's picture
Maximilian J. H...,
27/08/2019 23:25

Dear Andi,

I can only agree with what you say. Yes, it is fair to call my approach idealist. And yes, we also need rigid laws right now, maybe as a plan B or a transformation phase with. Your comparison with the straws is an enlighting example: right now it is without discussion the best thing to do. Hopefully, it is one step towards people realizing the consequences of their everyday cunsumption behaviour.

But maybe formulating laws could also be counterproductive? It might make people avoid straws because of the law and not because of the environment, eventually making them care less about environment where it is not enforced by law!? One document I have to think of all the time is the Universal Human Rights Declaration. It is not legally binding at all but is yet (or because of that) so powerful. An intereing question for me is whether it could be powerful on its own, i.e. without so many other countries "translating" it into regionals laws that are enforced. I am far away from seeing a complete answer and would be curious how your students approached a moral compass?

All the best,

Max

Total votes: 0
Maximilian J. Haarich's picture
Maximilian J. H...,
26/08/2019 21:12

Dear Andi,

thank you for your reply. I would rather compare the sun screen to each one's critical thinking. The attempt to counter AI deepfakes with AI deepfake detectors would in my opinion only fuel the arms race (like assumed in this article about the DARPA investments to spot deepfakes: "The adversary will always win, you will always be able to create a compelling fake image [...]".)

So who has the power to do something about deepfakes? Each and every one of us. Each one that produces or is confronted by deepfakes. I would not look for an organisation to be responsible for that. We all are responsible. In the end there are always individuals who create situations where deepfakes are produced or recieved. We need a change in beliefs and values. This is something that the EU or the UN can support like it was done once by means of the Universal Declaration of Human Rights. (Not because of what was said in this dcument, but because of its function to enable other organisations understand the implicit basic values and to adopt to these in their own work, in their own way.

All the best,

Max

Total votes: 2
Andrei Petreanu's picture
Andrei Petreanu,
27/08/2019 12:13

Thank you too, Max.

I find your answer to be accurate, but quite idealist, and i think passing the responsability to each individual will only burden the ones that live their life oblivious, and unknowingly irresponsible. In my mind *| time for another metaphorical parallel |* its like asking everybody to stop using straws, instead of regulating the production of plastic straws with the company that produces them in the first place. 

 

That being said, I really do hope such an education (and change in beliefs & values) will be part of every EU contry's target for the dawn of this deep-fake digital age.

 

I have also held a course in Machine Learning at the University of Bucharest, last year, and students are eager to get what moral compass they should obide by, but understanding the implications/potential for harm is closely related to understanding the cause and the underlying technology, and sadly, not everybody is a machine learning engineer.

 

Thank you again.

Andi

Total votes: 0
Maximilian J. Haarich's picture
Maximilian J. H...,
24/08/2019 15:59

Dear Mr. Petreanu,

 

thank you for your insightful thoughts! I deeply understand your concerns. May problems can arise if people are misinformed on large scale, espcially in connection with democratic elections. Yet, with the increasing availability to produce fakenews/deepfakes for everyone, humanitys awareness and ability to detect fakes rises as well. People could always be fooled. Imagine living in ancient times and your only source for information about past event is a giant oilpainting or some words carved into a stone. How could you resist that statement if it was untrue? Faking information is as old as mankind. Of course, we need to protect the most vulnerable. But the solution to me can only be to increase their critical thinking. We have to empower everyone to productively question the informations recieved instead of trying to eradicate the misuse of communication channels.

I see it like this: if I have a problem with the sun burning my skin, I put on sun screen. I won't try to to change the sun.

Warmest,

Max

Total votes: 0
Andrei Petreanu's picture
Andrei Petreanu,
26/08/2019 10:59

Hello Max,

The equivalent of sunscreen in this situation would be AI solutions that detect and correctly classify (with large accuracy rates) text fakes from unreliable sources. I would even go as far as to only mention the sources, because the text is now so, so good. So even with a 98%+ mAP (commonly found as s.o.t.a in the industry), the flood can be permanent and continuous, leaving 2% of a huge (potentially infinite) chunk of misinformation data to float around.

 

I'll leave you with this : if applying sunscreen (aka classifying unreliable sources) is as easy as it sounds, why has it now worked in the past, with giants like G or F having huge trouble in keeping fake news in check? If they can't do it (or there are political/business incentives not to do it), who else can? Is there an un-biased organism / institution / organization that deals with such control, and has the manpower / knowledge to understand the implications?

 

Thank you for your reply.

Andi

Total votes: 0
Farisch Hanoeman's picture
Farisch Hanoeman,
23/08/2019 18:53

Dear Andrei Petreanu, 

I think you're right to be concerned. I am too. I believe, as EU citizens, we should not rely on big tech firms to solve this problem. There is too much at stake here. Just think about the many ways people can be manipulated during elections by these algorithms. I believe there is just one problem: who decides what is true and what is not. I think here lies an opportunity for journalists to partake in this challenge. The EU should fund this, because the truth must be a common goal.

There are several ways to solve this technically. Here is one for a start: when a website publishes a lot of untruthful information the browser must annotate that the information on that specific website may not be true. Maybe even a warning page before you enter the website. As explained above; you can use an AI to detect fake news or journalists.

Total votes: 0