My name is Anat Ben-David. I am a senior lecturer at the Open University of Israel. On September 17 2019, I was involved in getting Facebook to suspend Prime Minister Netanyahu’s chatbot for three critical hours on Election Day. Here is why I did it, how I did it, and why I think it matters.

Hello, this is Prime Minister Netanyahu, may I ask you a question? In the past six months, Israel had two election rounds: the first was on April 9. Netanyahu’s party had the largest number of seats, but Netanyahu did not succeed in forming a coalition. New elections were announced, and the second round took place on September 17. During both rounds, I was closely monitoring propaganda on Facebook, as part of my research. I monitored all political actors, but compared to them, Netanyahu’s campaign was far more sophisticated in using Facebook’s advertising system. Netanyahu’s campaign also used a chatbot on Facebook Messenger, and eventually, on election day, the chatbot was used to target supporters, who were then asked to help the prime minister call other citizens and convince them to vote for him. When users agreed, they were served talking points. Then, the chatbot offered phone numbers of citizens and encouraged supporters to make more phone calls. Once in a few hours, the talking points were updated, and short videos featuring Netanyahu encouraged them to make even more phone calls. With each press of the interaction button, the bot would publish another phone number, and another one, and another.

Netanyahu’s chatbot abilities were revealed in the first election round and came as a surprise to many. During the campaign, many have engaged with the chatbot, which came to be fondly known as the ‘Bibi-Bot’. From the user’s end, the chatbot was perceived as cool and innovative. Many users, especially those who are less tech-savvy, were convinced that they were indeed chatting with the Prime Minister. Others enjoyed the fun interaction with the chatbot, which often displayed interactive games and humorous videos featuring the Prime Minister. But once in a while, the bot had also asked political questions, to which the users could answer by clicking on pre-prepared buttons. Such as the following:[1]

“In the meantime, I’d be happy to ask you a small question, may I?” [Sure] “Would you vote for me and the Likud in the upcoming election?” Options: [I am still unsure] / [Sure thing! I’ll vote Likud] / [I will not vote Likud]. Choice: [Sure thing! I’ll vote Likud] “Thank you very much, I’ll be in touch soon!”Bot: May I ask you a small question? Button: Sure! Bot: Would you vote for me and the Likud in the upcoming election? Buttons: I am still unsure / Sure thing! I’ll vote Likud / I will not vote Likud Bot: Thank you, I’ll be in touch soon"

The warning signs were already there. On March 18, a news report argued that Facebook had asked Netanyahu to stop collecting data through the chatbot. Facebook refused to answer what data had been collected, and what were their policies. But the bot continued targeting users.

We also know that throughout the campaign Netanyahu’s page ran several dozens sponsored ads, featuring a video in which Netanyahu is encouraging citizens to engage with him through Facebook Messenger. The ads’ text was identical: “What is the most important issue for you in the election? I invite you to write me a private message here on Facebook”. Facebook’s Ad Library features these ads. They seem as slight variations of the same ad, but they do not disclose targeting information. Through collecting ‘why am I seeing this ad’ screenshots of these ads, we noticed they were targeted to different audiences, such as people interested in the left-leaning newspaper, Haaretz.

A screenshot from Netanyahu’s sponsored ad on Facebook. February 21, 2019. Left, the ad’s text: “Hello there! This is Benjamin Neteanyahu. I’d like us to be in direct contract, and therefore I invite you to write me a private message what is the issue you find the most important in this election”. Right, why am I seeing this ad text: “One of the reasons you are seeing this ad is that Benjamin Netanyahu wants to reach people interested in Haaretz. This is based on activity such as liking pages or clicking on ads. There may be other reasons you are seeing this ad, including that Benjamin Netanyahu wants to reach people ages 18 and older who live in Israel. This information is based on your Facebook profile and where you are connected to the Internet.”

Thus, it can be fairly assumed (but not determined) that the Likud’s data collection tactics were taking advantage of everything Facebook has to offer in an efficient feedback loop between the advertising system and the chatbot. The idea is simple: take an initial list of citizens, target them by their political inclination through the advertising system (such as newspaper readership, or a “lookalike campaign”), and then, after you already know more about them through Facebook, learn even more about them by asking direct questions via the chatbot.

On the first election day, the chatbot did not contact all citizens. Only those identified as Likud supporters were offered the talking points and the phone numbers. How did the chatbot know whom to target? To which database was the chatbot connected that allowed turning it into a distributed call center? In April, there was no way of knowing. Both Facebook and the Likud refused to answer questions on the topic. However, campaign staffers for the Likud, interviewed after Netanyahu’s victory in the first round (and before the second round was announced), did boast on the effectiveness of the chatbot. Topaz Luk, Netanyahu’s new media adviser, said in an interview to Amir Ben-David for the news website Zman Israel (no family ties with the author): “It helped us map out the people we corresponded with. The mapping options are great. We took the primitive capabilities of SMS messages that everyone uses and turned them into a tool that is much smarter and targeted. The results were amazing […] By asking questions you can tell who is sure to vote for you, who is on your right and unsure if they will vote for you, and who is unlikely to vote for you. It’s like a polling system, just more sophisticated.” In a reply to the reporter’s question whether the bot had kept the data it had collected to be used later on, Luk said: “Yes. Listen, all according to Facebook’s rules, of course.” In an interview to Tal Schneider for Globes, Ofer Golan, the Likud’s campaign manager, revealed their uncompromising tactic: “First win. Then do damage control”. One Million Citizens When the second election was announced, we re-launched our monitoring project. This time, other politicians and political parties also activated chatbots on messenger. Yet while these were mostly asking whether the user would like to sign up to volunteer in the campaign, or if they were willing to share a video or attend an event, the chatbots of Netanyahu’s political opponents were less sophisticated. Towards the end of the campaign, the data collection efforts of the Netanyahu campaign became more overt. In a video posted on Netanyahu’s page on July 29, the Prime Minister is seen encouraging his supporters to build a dataset of a million Likud supporters. “You know we can’t trust the news media”, Netanyahu says in the video. “I need you to give us the names of as many Likud supporters [..] and I promise to personally visit those who will give us the most names”. The post’s text displayed a link, which led to an online form titled “The Race to the Million” (referring to the title of the popular reality TV show which is based on the American “The Amazing Race”, in which the winners of the show win a million Shekels). There, supporters can type in information about their friends: name, address, phone number. Where is this database stored? Are citizens informed that their names are nominated? Is the database registered with the Authority for the Protection of Privacy, as required by law? These questions were left unanswered until two weeks before the election when a scoop by Hagai Amit for The Marker revealed that the Likud was running a dataset on one million citizens tagged by their political affiliation: supporters, objectors, and undecided. The dataset was not secured (as required) and could have been accessed through a web browser. The terminology in the report was too similar to that of the “Race to the Million” campaign, and the tagging was too similar to the questions asked by the chatbot. There is no evidence that the two are connected, but there is a correlation. Soon after, the bot had changed its communication strategy from contacting users with questions or humorous videos, to re-activating the automated call center from the first round. Here, too, only users identified as supporters were contacted: talking points were sent, but this time instead of phone numbers, the app provided names. “You next call is with Avi, please press the dial button”, the bot said. The user then pressed a button, which opened the dialing app on her phone, where the targeted phone number was already displayed. Four seconds later, the bot inquired: “Did they answer”?

“Your next call is with Avi, please press the dial button” [Dial 💪] “Did they answer”? Buttons: [Yes, they did! 😁] / [No, they didn’t😟] If they didn’t answer, the bot had asked the user if she was interested in making another phone call. The answer button was something like: “Cool, Mr Prime Minister!”, or “Sure, give me a number!”. If there was an answer, the bot then inquired: “How did it go?” The possible answers were: “Excellent, they will convince others!”, “Great, they will vote for the Likud”, “Wrong number / Not supporting Netanyahu”. Then, the bot would offer to make another phone call, and the process could go on and on.

“Your next call is with Shlomo. Please press the dial button”. [Dial💪] “Did they answer?” [Yes, they did! 😁/ No, they didn’t 😟 “Excellent! How did the conversation go?” [Excellent, they will convince others💪], [Great, they will vote for the Likud👍], [Wrong number / Not supporting Netanyahu👎].

Until an update to the talking points had crossed the line and the story evolved from a correlation between the chatbot and dubious activity to causality. The Arabs want to annihilate us all — Men, women and children

“In the meantime, I’d be happy to ask you a small question, may I?” [Sure] “Would you vote for me and the Likud in the upcoming election?” Options: [I am still unsure] / [Sure thing! I’ll vote Likud] / [I will not vote Likud]. Choice: [Sure thing! I’ll vote Likud] “Thank you very much, I’ll be in touch soon!”Bot: May I ask you a small question? Button: Sure! Bot: Would you vote for me and the Likud in the upcoming election? Buttons: I am still unsure / Sure thing! I’ll vote Likud / I will not vote Likud Bot: Thank you, I’ll be in touch soon. On September 10 the chatbot said: *** Attention, new script ***

Hello {first name}, my name is ___ and I am a volunteer on behalf of Prime Minister Netanyahu. I am calling you because on Tuesday you will be able to determine the future of our country. Prime Minister Netanyahu brings with him a right-wing policy of a Jewish state, security and a strong Israel. I’m donating my time because a dangerous left-wing government should not be set up in a week with Lapid, Odeh, Gantz and Liberman. A secular, weak, left government that trusts the Arabs who want to annihilate us all — women, children and men — and will allow a nuclear Iran to destroy us. We must not let that happen! So I ask you to be the Prime Minister’s messenger, to bring 3 friends and family next Tuesday and see to it that they vote Likud. Thanks, {first name}, I trust you! I was contacted by a journalist from the independent news website The Seventh Eye, who asked me to verify that this indeed is an authentic message sent through the Prime Minister’s chatbot. I interacted with the chatbot through one of our project’s accounts and immediately received the same message. It wasn’t fake. I admit to being shocked. The journalists from The Seventh Eye and from The Marker, with whom I have been in touch throughout both campaigns, immediately ran the stories. The screenshot from my phone, taken after I interacted with the bot and got the inciting message, soon appeared everywhere. The journalists contacted Facebook for comment — but they refused to answer. Netanyahu, on his part, stated that he did not confirm the content of this message and that this was a mistake of a junior staffer. The NGO Adalah, a legal center advocating Palestinian rights in Israel, submitted a petition to the Central Election Committee. Aiman Odeh, leader of the Joint Arab List, also contacted Facebook with a demand to immediately remove the inciting message over hate speech. Given the public pressure, Facebook decided to act and issued the following statement: "After careful review of the Likud campaign’s bot activities, we found a violation of our hate speech policy. We also found that the bot was misusing the platform in the period allowed to contact people. As a result, we temporarily suspended the bot for 24 hours. Should there be any additional violations, we will continue to take appropriate action Until that statement was issued, nobody knew there was a limit on the number of times a bot is allowed to contact users within a given period." Thus Facebook’s suspension of the bot came before judge Melcer issued a decision on Adalah’s appeal . To our understanding, this was the first time Facebook intervened in suspending a head of state’s activity on the platform during an election campaign. After 24 hours, the chatbot went live again. But it continued to violate the local law. They took a 100-KG hammer and dropped it on a fly! The Israeli law prohibits the publication of polls four days before the election. On September 15, two days ahead of the election, Netanyahu’s chatbot sent the following message: As of now, according to our internal polls, we are losing the elections.

The message was accompanied by an infographic in a form of a designed pie-chart, which claims that the right-wing camp has only 56 seats, compared to 64 seats of the left camp, comprised of Lapid, Gantz “and the Arabs”. I tweeted the screenshot in my research project’s account and alerted the journalists, who immediately ran the story. This is when I decided to contact Adv. Shachar Ben-Meir. For him, it was clear that the last message violated the law, and offered me to join him in submitting a petition to the Central Elections Committee. I decided to join the petition. An unusual step for a researcher. But I did so because this was the second time in two days that the chatbot was bypassing the law through a targeting system that takes hide in Facebook’s platform. The chatbot’s messages were meant to be seen only by supporters and far from public scrutiny. I believed they deserved public scrutiny. If these violations continued as Israeli citizens went to the polling stations, I feared their consequences on election integrity. On September 16, 24 hours before the election, we submitted the petition to the Central Election Committee against the Likud party, Benjamin Netanyahu, Facebook Inc. and the Attorney General. The petition demanded: A) To issue an order instructing Netanyahu and the Likud to cease publishing any poll or polling results contrary to legislation. B) To issue an order instructing Facebook to erase and halt any activity of the Likud’s chatbot that violates Israeli legislation, or to suspend the chatbot’s activity. Adv. Ben-Meir submitted the petition at 9 AM. This is the chain of events that followed our appeal: 1 PM: Judge Melcer issued a temporary injunction until all parties replied. 3 PM: The Attorney General replied that the publication of the poll is in contravention of the election law, and there is a basis for ordering the requested injunction. 4 PM: Netanyahu gave an interview to the radio station “Kol Barama”, and related to results of internal polls. Adv. Ben-Meir issued an urgent complaint on contempt of a judicial order. 5 PM: Facebook’s reply argued that since the screenshot of the chatbot’s message was posted on Twitter and then published in the newspaper The Marker, the request to remove it is futile. Facebook further claimed that the tight time-frame does not allow the company to weigh in on the complex and broader implications of the petition. 6 PM: The Likud’s reply demanded to dismiss the petition arguing that the message does not qualify as the publication of polling results. Specifically, they argued that the infographic is not a finding of a specific election poll and does not present itself as such. They further argued that the graphic presentation of the distribution of seats between the left and right, as of the Likud’s opinion at the time of publication, does not constitute a voter consideration for which party to vote, but at most it is intended to bring to the attention of right-wing fans an analysis of a situation at a particular moment. 9 PM: Netanyahu broadcasts on Facebook Live. He is accompanied by John McLaughlin, Trump’s pollster who is also the pollster for the Likud’s campaign. During the live broadcast, Netanyahu says that according to their polls, the expected turnout rate for the left is 10% higher than that of the right. Adv. Shachar Ben-Meir issues a second complaint on contempt. 10 PM: Judge Melcer’s decision is in. The head of the CEC issues a directive instructing both the Likud and Facebook to discontinue polling, or publishing poll results, including any poll or poll results that display predicted voting patterns or voting rates. The directive also instructed the Likud and Facebook to remove the publications (i.e. the chatbot message with the polling infographic) immediately, not later than 7 am tomorrow from any media where they appear. **17 September 2019. Election Day. Israelis head to the ballot box** 9 AM: The chatbot still features the message, and keeps providing phone numbers of citizens. Facebook sends an urgent appeal to the CEC, asking to reconsider and annul the directive, as it is technically impossible to delete the message, and since Facebook, as a neutral platform, is not responsible for users’ content. Further, Facebook cannot guarantee the future publication of polls, as ordered by the directive. In the petition, Facebook’s lawyers also protest the unusual order, arguing that this is the first time the chairman imposes such an exceptional, extreme, and impractical duty on Facebook. According to Facebook, in all the petitions, the chairman has given them specific instructions on specific orders, which relate to specific publications, which they followed and implemented. 12 PM:The chatbot sends new talking points. This time, the bot is reaching out to more users who have previously interacted with the chatbot, but were not asked to make phone calls:

* Pay attention, the script for calls! * Hello, I am calling on behalf of Prime Minister Netanyahu. Our polls show that we are close to the left-wing government of Lapid and Gantz where Odeh and Tibi will be ministers. Israelis know that Iran and its terrorist supporters will see the rise of such a government as a sign of weakness, and they will rush to get nuclear weapons. President Trump, who announced yesterday that he wants to sign a vital Defense Treaty with Israel, will lose confidence in the government if the prime minister does not win big. Only a right-wing government with a large majority for Netanyahu will keep us safe. A left-wing government led by Lapid and Gantz will evacuate 90,000 Jews. If you stay home, the left wins. Only your vote will guarantee victory to the right. OnTuesday, vote for Likud and Netanyahu and share the message with your relatives! Can I tell the Prime Minister that you are with us? 1 PM: The chatbot says: “As of now, voting rates in the Arab sector are high”. But twenty minutes later, the bot stops responding to user interactions. The button “Yallah, Mr Prime Minister, I’m ready!”, which so far led to the bot replying with another phone number, did not yield any response. “Yallah, Prime Minister, I’m ready!”, “Yallah, Prime Minister, I’m ready!” — silence.

“Dear volunteers, I need you! Watch>>” [Yallah, Prime Minister, I’m ready! ❤] “As of now, voting rates in the Arab sector are high” [Yallah, Prime Minister, I’m ready! ❤] [Yallah, Prime Minister, I’m ready! ❤] [Yes, they answered], [Yallah, Prime Minister, I’m ready! ❤], [Yallah, Prime Minister, I’m ready! ❤] Soon, Facebook’s spokesperson issued a statement: "We’re working with elections officials around the world to help ensure the integrity of the elections. Our policy explicitly states that developers are required to obey all laws applicable in the country where their application is accessible. Therefore we’ve suspended the [Netanyahu] bot’s activity, in light of the violation of local law, until the closing of the polls” at 10 p.m. on Tuesday." Netanyahu responded to the suspension during a Facebook Live broadcast, saying: “Facebook is unable to stand up to leftist pressure.” He later deleted that video. Later on, an infuriated Netanyahu responded again on Facebook Live: They shut down our chat-bot! It’s unbelievable! It’s disproportionate, it’s unjust and they only do it to us. Give us a penalty, do something proportionate. I can’t believe you do that! They take a 100 Kilogram hammer and drop it on a fly because it’s the Likud’s fly.

4 PM: Judge Melcer orders Facebook to renew the chatbot’s activity, after MK David Bittan, the Likud’s representative at the CEC, has pledged that there will be no further violations of the directive (and the law) until the polling stations close. The chatbot is revived.

[Yallah, Prime Minister, I’m ready! ♥️] “My Chatbot is back online after it had been suspended by Facebook for three hours👎”. “The system is back in the air , I am asking you to call as many people as possible”. Video caption: “We’ve had a significant delay with our chatbot”. Button: [Ready to start making phone calls 💪]Bot: My Chatbot is back online after it had been suspended by Facebook for three hours.

Bot: The system is back in the air , I am asking you to call as many people as possible Button: Give me a phone number! A New phone number appears on the user’s dialing app. 10 PM: The polling stations close.

A small win for democracy It doesn’t happen so often that Facebook penalizes a Prime Minister’s chatbot on Election Day over a screenshot. But the screenshots I collected as part of my election monitoring project were instrumental in providing evidence of political targeting that was meant to be seen by some people, but not all people. While no one, including Facebook, is able to estimate the impact of the Cambridge Analytica scandal on the results of the 2016 election in the United States, here, too, it would be reckless to speculate on the implications of the chatbot’s activity (and the interferences to it), on the outcomes of the election or on voting behavior in Israel. But this case is important in alerting public attention to the following points:

1) Chatbots and live broadcasts are effective tools for the Likud’s “win first, damage control later” campaign strategy. Even after the first 24 hours suspension, the operators of the chatbot were not deterred in using it as they saw fit, regardless of the law.

2) Democratic tools still work. The news media were instrumental in reporting the chatbot’s violations. I believe it was the public outcry over the inciting message against the Arabs — which also resonated internationally — that led Facebook to suspend the chatbot for 24 hours. As a reminder, for almost two long days, Facebook refused to comment on Netanyahu’s chatbots apparent violation of their hate speech policies, until they decided on first temporary suspension. Judge Melcer has made a courageous decision to apply the election law on social media (the law only relates to broadcast media and does not address the new challenges brought by new media campaigns). He was instrumental in weighing in — in an impossible time framework — on the implications of the use of real-time broadcast and targeting media on election integrity. Both the CEC and the news media were the leverages for getting Facebook to take action, and acknowledge their responsibility to ensuring election integrity in real-time.

3) We need better tools. In a small country such as Israel, it is feasible (albeit laborious) to run an election monitoring project based on screenshots. But this method doesn’t scale. The reason we used screenshot is that Facebook has shut down other public initiatives that monitored targeted messaging through automated browser add-ons. Researchers of political campaigns are directed to Facebook’s Ad Library, but the latter does not disclose targeting information. The Ad Library also does not provide any information on political targeting via Facebook Messenger. It is these targeted, unmonitored spaces, where manipulation might take place.

4) What would be the implications of this case on other countries? In the hate speech incident, Facebook was quick to respond before there was a judicial directive from the CEC. But for the first time, Judge Melcer’s directive on the polling message held Facebook responsible for hosting illegal propaganda. It remains to be seen how this precedent will affect Facebook’s policies and ties with other CECs in other countries with upcoming elections.

5) Chatbots are external applications that collect user data. The absurdity is that If Facebook has learned the lessons from the Cambridge Analytical scandal and shut down hundreds of thousands of apps that extracted data from the platform, they shouldn’t allow third-party data collection through chatbots. Don’t talk to chatbots. [1] All direct quotes in this document have been translated from Hebrew by the author.