Fake News: Can Artificial Intelligence Help?

Slide 1

Slide 1 text

Fake News: can Artificial Intelligence help? Marco Bonzanini Workshop: Studying Russian disinformation and its effects in Europe Ludwig-Maximilian University of Munich, Germany

Slide 2

Slide 2 text

— Betteridge’s Law of Headlines “Any headline   that ends in a question mark   can be answered   by the word no”

Slide 3

Slide 3 text

ARTIFICIAL INTELLIGENCE

Slide 4

Slide 4 text

Slide 5

Slide 5 text

AI Computer Science Mathematics

Slide 6

Slide 6 text

AI Computer Science Mathematics Linguistics

Slide 7

Slide 7 text

AI Computer Science Mathematics Philosophy Psychology Linguistics Neuroscience

Slide 8

Slide 8 text

AI Computer Science Mathematics Philosophy Psychology Linguistics Neuroscience

Slide 9

Slide 9 text

Intelligent Agents Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 10

Slide 10 text

Intelligent Agents Any device that: Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 11

Slide 11 text

Intelligent Agents Any device that: • perceives its environment Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 12

Slide 12 text

Intelligent Agents Any device that: • perceives its environment • takes action Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 13

Slide 13 text

Intelligent Agents Any device that: • perceives its environment • takes action • maximises the chances of success Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 14

Slide 14 text

Intelligent Agents Any device that: • perceives its environment • takes action • maximises the chances of success Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 15

Slide 15 text

Intelligent Agents Any device that: • perceives its environment • takes action • maximises the chances of success Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003 what’s the goal?

Slide 16

Slide 16 text

Intelligent Agents Systems that: Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 17

Slide 17 text

Intelligent Agents Systems that: • act like humans Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 18

Slide 18 text

Intelligent Agents Systems that: • act like humans • think like humans Russell and Norvig, Artiﬁcial Intelligence: A Modern Approach, 2003

Slide 19

Slide 19 text

Act Like a Human

Slide 20

Slide 20 text

Act Like a Human A. Turing (1950) Computing Machinery and Intelligence

Slide 21

Slide 21 text

Act Like a Human A. Turing (1950) Computing Machinery and Intelligence ??? Human AI

Slide 22

Slide 22 text

Act Like a Human A. Turing (1950) Computing Machinery and Intelligence ??? Human AI a.k.a. The Turing Test

Slide 23

Slide 23 text

Passing the Turing Test

Slide 24

Slide 24 text

Passing the Turing Test Ex Machina, 2015

Slide 25

Slide 25 text

AI: Main Ingredients

Slide 26

Slide 26 text

AI: Main Ingredients AI {

Slide 27

Slide 27 text

AI: Main Ingredients • knowledge AI {

Slide 28

Slide 28 text

AI: Main Ingredients • knowledge • reasoning AI {

Slide 29

Slide 29 text

AI: Main Ingredients • knowledge • reasoning • language understanding AI {

Slide 30

Slide 30 text

AI: Main Ingredients • knowledge • reasoning • language understanding • learning AI {

Slide 31

Slide 31 text

Machine Learning

Slide 32

Slide 32 text

Machine Learning ML gives computers the ability to learn  without being explicitly programmed

Slide 33

Slide 33 text

Types of Machine Learning

Slide 34

Slide 34 text

Types of Machine Learning • Supervised • Unsupervised • Semi-supervised • Reinforcement

Slide 35

Slide 35 text

Types of Machine Learning • Supervised • Unsupervised • Semi-supervised • Reinforcement Sample inputs + Desired outputs

Slide 36

Slide 36 text

Types of Machine Learning • Supervised • Unsupervised • Semi-supervised • Reinforcement Find structure in sample input

Slide 37

Slide 37 text

Types of Machine Learning • Supervised • Unsupervised • Semi-supervised • Reinforcement Incomplete training signal

Slide 38

Slide 38 text

Types of Machine Learning • Supervised • Unsupervised • Semi-supervised • Reinforcement Reward or punishment

Slide 39

Slide 39 text

Applications

Slide 40

Slide 40 text

Applications • Filter spam emails

Slide 41

Slide 41 text

Applications • Filter spam emails • Play poker/chess/go

Slide 42

Slide 42 text

Applications • Filter spam emails • Play poker/chess/go • Drive safely on a tortuous track

Slide 43

Slide 43 text

Applications • Filter spam emails • Play poker/chess/go • Drive safely on a tortuous track • Drive safely on a busy road

Slide 44

Slide 44 text

Applications • Filter spam emails • Play poker/chess/go • Drive safely on a tortuous track • Drive safely on a busy road • Answer simple factual questions

Slide 45

Slide 45 text

Applications • Filter spam emails • Play poker/chess/go • Drive safely on a tortuous track • Drive safely on a busy road • Answer simple factual questions • Give legal advice in a specialised field

Slide 46

Slide 46 text

Slide 47

Slide 47 text

Slide 48

Slide 48 text

http://www.independent.co.uk/life-style/gadgets-and-tech/news/tay-tweets-microsoft-creates-bizarre-twitter-robot-for-people-to-chat-to-a6947806.html

Slide 49

Slide 49 text

http://www.stuff.co.nz/technology/digital-living/78265631/Microsofts-AI-teen-turns-into-Hitler-loving-Trump-fan-thanks-to-the-internet

Slide 50

Slide 50 text

AI in Summary

Slide 51

Slide 51 text

AI in Summary • What are you trying to solve?

Slide 52

Slide 52 text

AI in Summary • What are you trying to solve? • Current state of the art: narrow problems

Slide 53

Slide 53 text

AI in Summary • What are you trying to solve? • Current state of the art: narrow problems • Generalising is difficult

Slide 54

Slide 54 text

AI in Summary • What are you trying to solve? • Current state of the art: narrow problems • Generalising is difficult • Artificial General Intelligence is not here

Slide 55

Slide 55 text

AI in Summary • What are you trying to solve? • Current state of the art: narrow problems • Generalising is difficult • Artificial General Intelligence is not here (yet )

Slide 56

Slide 56 text

FAKE NEWS

Slide 57

Slide 57 text

Google Trends for “fake news”

Slide 58

Slide 58 text

Google Trends for “fake news” No fake news before November 2016?

Slide 59

Slide 59 text

http://www.bbc.co.uk/news/uk-40098804

Slide 60

Slide 60 text

https://www.theguardian.com/media/2017/jun/13/fake-news-manipulate-elections-paid-propaganda

Slide 61

Slide 61 text

https://www.theguardian.com/media/2017/jun/13/fake-news-manipulate-elections-paid-propaganda ~$55k to discredit a journalist ~$200k to instigate a street protest

Slide 62

Slide 62 text

https://twitter.com/carolecadwalla/status/872767628712005632

Slide 63

Slide 63 text

https://twitter.com/carolecadwalla/status/872767628712005632 Democracy  to the highest bid?

Slide 64

Slide 64 text

News (Fake or Not)

Slide 65

Slide 65 text

News (Fake or Not) • Easy to manufacture

Slide 66

Slide 66 text

News (Fake or Not) • Easy to manufacture • Difficult to verify

Slide 67

Slide 67 text

News (Fake or Not) • Easy to manufacture • Difficult to verify • 83.1M posts/month on WordPress.com

Slide 68

Slide 68 text

News (Fake or Not) • Easy to manufacture • Difficult to verify • 83.1M posts/month on WordPress.com • 500M tweets/day on Twitter

Slide 69

Slide 69 text

Information Overload

Slide 70

Slide 70 text

AI vs FAKE NEWS

Slide 71

Slide 71 text

Information Overload

Slide 72

Slide 72 text

Information Overload Recommender systems can choose for us!

Slide 73

Slide 73 text

No content

Slide 74

Slide 74 text

No content

Slide 75

Slide 75 text

No content

Slide 76

Slide 76 text

Filter Bubble

Slide 77

Slide 77 text

Filter Bubble stream of news recommender  system environmentalists football nuts

Slide 78

Slide 78 text

Filter Bubble news on  climate change

Slide 79

Slide 79 text

Filter Bubble are you going  to “like” it?

Slide 80

Slide 80 text

Filter Bubble

Slide 81

Slide 81 text

Filter Bubble transfer market news

Slide 82

Slide 82 text

Filter Bubble are you going  to “like” it?

Slide 83

Slide 83 text

Filter Bubble

Slide 84

Slide 84 text

https://www.theguardian.com/technology/2017/may/22/social-media-election-facebook-filter-bubbles

Slide 85

Slide 85 text

Fact Checking

Slide 86

Slide 86 text

Fact Checking Full Fact, The State of Automatic Factchecking, 2016. 1. Monitor

Slide 87

Slide 87 text

Fact Checking Full Fact, The State of Automatic Factchecking, 2016. 1. Monitor 2. Spot claims

Slide 88

Slide 88 text

Fact Checking Full Fact, The State of Automatic Factchecking, 2016. 1. Monitor 2. Spot claims 3. Check claims

Slide 89

Slide 89 text

Fact Checking Full Fact, The State of Automatic Factchecking, 2016. 1. Monitor 2. Spot claims 3. Check claims 4. Create and publish

Slide 90

Slide 90 text

Monitor

Slide 91

Slide 91 text

Monitor • What to monitor?

Slide 92

Slide 92 text

Monitor • What to monitor? • How to access?

Slide 93

Slide 93 text

Monitor • What to monitor? • How to access? • Scalability & robustness

Slide 94

Slide 94 text

Monitor • What to monitor? • How to access? • Scalability & robustness • Long-term availability

Slide 95

Slide 95 text

Monitor • What to monitor? • How to access? • Scalability & robustness • Long-term availability • Engineering problems, not really AI

Slide 96

Slide 96 text

Spot Claims

Slide 97

Slide 97 text

Spot Claims • Human judges don’t always agree      

Slide 98

Slide 98 text

Spot Claims • Human judges don’t always agree • Text analysis to the rescue    

Slide 99

Slide 99 text

Spot Claims • Human judges don’t always agree • Text analysis to the rescue • Natural language is “difficult”  (paraphrasing, common sense, idiomatic expressions, …)

Slide 100

Slide 100 text

Check Claims

Slide 101

Slide 101 text

Check Claims • Look-up high-quality knowledge base

Slide 102

Slide 102 text

Check Claims • Look-up high-quality knowledge base • … how do you model the universe?

Slide 103

Slide 103 text

Check Claims • Look-up high-quality knowledge base • … how do you model the universe? • Context is important

Slide 104

Slide 104 text

Create and Publish

Slide 105

Slide 105 text

Create and Publish • Communicating the results is crucial

Slide 106

Slide 106 text

Create and Publish • Communicating the results is crucial • Automated journalism?

Slide 107

Slide 107 text

Create and Publish • Communicating the results is crucial • Automated journalism? • How do you build trust?

Slide 108

Slide 108 text

— DJ Patil “The hardest part of data science is getting good, clean data. Cleaning data is often 80% of the work” “We didn’t have better algorithms, we just had better data.” — Peter Norvig

Slide 109

Slide 109 text

Can AI help? Is this the right question? Automation vs AI Data Quality Summary