Deposition of Joshua Simons

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 1 of 43 AMENDED Exhibit 13 PLAINTIFFS' OMNIBUS OPPOSITION TO DEFENDANTS' MOTIONS FOR SUMMARY JUDGMENT Case No.: 4:22-md-03047-YGR MDL No. 3047 In Re: Social Media Adolescent Addiction/Personal Injury Products Liability Litigation 1 Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 2 of 43 Page 1 UNITED STATES DISTRICT COURT FOR THE NORTHERN DISTRICT OF CALIFORNIA 2 IN RE: SOCIAL MEDIA ADOLESCENT ) MDL No. 3 ADDICTION/PERSONAL INJURY PRODUCTS ) 4:22-md-3047-YGR LIABILITY LITIGATION ) 4 ) 5 SUPERIOR COURT OF THE STATE OF CALIFORNIA FOR THE COUNTY OF LOS ANGELES SPRING STREET COURTHOUSE 6 7 COORDINATION PROCEEDING SPECIAL TITLE [RULE 3.400] 8 SOCIAL MEDIA CASES 9 For Filing Purposes 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 ) ) ) ) Lead Case No. ) 22STCV21355 VIDEO DEPOSITION OF JOSHUA SIMONS MP April 1, 2025 Reporter: John Arndt, CSR, CCR, RDR, CRR CSR No. 084-004605 CCR No. 1186 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com 1 2 3 4 5 6 Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 3 of 43 Deposition of Joshua Simons MP, taken pursuant to Notice of Taking Deposition, before John Arndt, a Certified Shorthand Reporter and Certified Court Reporter, at Mishcon de Reya, LLP, 70 Kingsway, London, WC2B 6AH, United Kingdom, on April 1, 2025. APPEARANCES OF COUNSEL On Behalf of the MDL, personal injury, and school district plaintiffs: Motley Rice LLC 7 28 Bridgeside Boulevard Mt. Pleasant, SC 29464 8 843-216-9328 By: Jessica L. Carroll 9 10 11 12 13 14 15 16 jcarroll@motleyrice.com Andrew Arnold aarnold@motleyrice.com Previn Warren (via videoconference) Pwarren@motleyrice.com On Behalf of the Witness: Levy Firestone Muse LLP 900 17th Street NW, Suite 1200 Washington, DC 20006 202-845-3215 By: Joshua Levy jal@levyfirestone.com On Behalf of Meta Platforms, Instagram, and Siculus: Covington & Burling LLP 17 620 Eighth Avenue New York, NY 10018 18 212-841-1166 By: Gregory L. Halperin 19 ghalperin@cov.com Brian Reiser 20 breiser@cov.com 21 On Behalf of Meta: Bass, Berry & Sims PLC 22 100 Peabody Place, Suite 1300 Memphis, TN 38103 23 901-543-5966 By: Jordan Thomas (via videoconference) 24 jordan.thomas@bassberry.com 25 877-370-3377 Page 2 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 4 of 43 Page 11 1 interchangeably today, and that we all understand that 2 3 it represents the same company. Other than that we've spoken before, have we 4 not? 5 A. We have. 6 MR. HALPERIN: Object to form. 7 BY MS. CARROLL: 8 Ω. Do you do you recall when about that 9 was that we spoke? 10 A. It was roughly two to three weeks ago, I 11 think, roughly. 12 Ω. Did plaintiffs' counsel ask whether you 13 would be willing to testify on the record? 14 A. They did. 15 Ω. Appreciate it. Thank you for your time. 16 Just to orient the jury as to your background, 17 18 19 starting with your education, could you give us your where you attended university and any professional degrees thereafter? 20 A. 21 22 23 Yeah. I went to college in Cambridge University in the UK, where I studied social and political sciences for three years. I then worked. worked for a think tank. I worked for the Labour Party I 24 for a while. And I then started a PhD program in 25 Harvard, where I was for from 2016 until 2021, Golkow Technologies, 877-370-3377 A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 5 of 43 Page 12 1 studying political science. I then turned my PhD into 2 a book. 3 Ω. 4 A. And you received your PhD from Harvard? I did, yeah. 5 Ω. And what was that in? Was it governance? 6 A. It was. It was a PhD in political science 7 8 9 10 11 12 in the government department. I majored in political theory and political philosophy, and over the course of my PhD, I became interested in machine learning and computer science and data science and statistics, and the connection between those things and democracy, and I wrote a PhD about it's called Algorithms and 13 14 15 Democracy, which I finished in 2021, and then for a year and a bit afterwards, I turned that PhD into a book which is published by Princeton University Press. 16 Ω. So was the your dissertation or 17 thesis that was the manuscript for what became the 18 book that was published in 2023. Is that correct? 19 A. That's correct, yeah. 20 Ω. 21 parliament? 22 A. Okay. And you are currently a member of You were elected in July? Is that right? That's correct. 23 Ω. And you were in government prior to doing 24 your PhD program at Harvard. Is that also correct? 25 A. My side didn't win an election for quite a Golkow Technologies, 877-370-3377 A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 6 of 43 Page 13 1 2 3 4 I worked in politics for the Labour Party for about eight months during the year long time, so I was in politics but not in government. We were very much in opposition at that point. But yeah, I worked in 5 before I went to Harvard. 6 2. How did you become interested in the 7 technology side of your studies? 8 MR. HALPERIN: Object to form. 9 A. I 10 11 12 13 14 when I was at Harvard, it was actually through my wife, like many good things. and I she was interested in statistics, and I was interested in statistics and central banking, and we got talking and talked about classes that we were both taking. She 15 16 17 18 19 20 21 22 She suggested a class taught by a woman called Finale Doshi-Velez called Introduction to Machine Learning in the engineering department in computer science, which she said was super interesting, so I enrolled, went to the lectures, and through that class became close to and was learning from a woman called Cynthia Dwork, who is a cryptographer and computer scientist. 23 24 25 She helped with the cryptography the U.S. census, and she was an expert in fairness in artificial intelligence, trying to work out what it means to have 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 7 of 43 Page 14 1 fair AI. 2 3 And so I with Cynthia Dwork got a reading group together of computer scientists and political 4 5 6 7 scientists and engineers, and Cynthia Dwork then agreed to join my PhD committee, so it had a mix of folks from government, computer science, and law, and then it was through her that I ended up writing the PhD about 8 computer science. 9 Ω. Very interesting. I'm going to introduce 10 11 our first exhibit, which is just a printout of your LinkedIn profile. I will represent to you that this is 12 how it prints when you print from LinkedIn. 13 So I just want you to review that and make sure 14 what is listed on your LinkedIn profile is accurate. 15 THE REPORTER: And is that Exhibit 1? 16 MS. CARROLL: It is. 17 THE REPORTER: Thank you. 18 [Exhibit 1 marked for identification.] 19 BY MS. CARROLL: 20 Ω. Mr. Simons, is this an accurate 21 representation of your education and experience? 22 A. It is, yeah. 23 24 25 Yeah, I mean, the only thing, just to mention, is that the method of my employment at Facebook changed several times over the course of the period of time I 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 8 of 43 Page 15 1 was there. I was initially hired as a contractor, and 2 then I became a full Facebook employee on a part-time 3 basis after a period of being a contractor. 4 Ω. Do you recall the time period or 5 approximate time period for that change? 6 A. Yeah, in broad terms, I do. In the summer 7 8 9 of 2018, I was a contractor, and then that contracting period came to an end. I was then rehired as a contractor on a rolling basis rather than a fixed-term 10 11 period, and then after that, I was hired as a Facebook employee, STE part-time full Facebook employee. 12 MR. LEVY: Explain STE. 13 A. STE means short I'm actually not sure 14 15 what STE means. Yeah, I'm not sure what STE means. actually never quite knew what STE means. But it was I 16 a 17 18 19 20 I was a full Facebook employee with all the same benefits as employees, but it wasn't a full-time position, because I was also doing my PhD at the same time, which in the team that I was in was a common form of employment. 21 22 23 24 25 There were a couple of other folks employed on the same basis. And then when I moved to the UK, I became a full Facebook employee on a part-time basis in the same arrangement, but through the UK rather than the U.S. 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 9 of 43 Page 16 1 BY MS. CARROLL: 2 Ω. If you recall, what was the impetus for 3 4 5 6 the change in the employment status, like contractor to employee? Was that something that you raised, or was it something that they offered? MR. HALPERIN: Object to form. or from 7 A. It was something that my manager, 8 9 10 11 12 13 and her manager, discussed with me explicitly. They were concerned that given the seniority of executive that I was talking to and engaging with, that my employment as a contractor rather than a full-time employee of Facebook would raise some risks for them. 14 15 16 17 I didn't fully understand what those were. They didn't explain those to me. But they were concerned about risks to them, and so they thought it would be better to hire me as a Facebook employee. 18 BY MS. CARROLL: 19 Ω. Okay. Thank you. We can put that aside 20 for now. 21 [Exhibit 2 marked for identification.] 22 BY MS. CARROLL: 23 Ω. I'll go ahead and introduce our second 24 exhibit, which is we're going to the second 25 exhibit will be excerpts from your book. Obviously due 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 10 of 43 Page 17 1 2 3 4 5 to copyright reasons, we're not putting the full publication into evidence. But this is something I think we will reference throughout. But to start, I'd just like to understand what prompted the writing of this book for you. 6 7 8 THE VIDEOGRAPHER: Which document is this? MS. CARROLL: This is the second one. THE VIDEOGRAPHER: Does it have a label? 9 MR. ARNOLD: B. 10 THE VIDEOGRAPHER: B. 11 MR. HALPERIN: We'll object to the 12 preamble. 13 THE VIDEOGRAPHER: I don't see a B in 14 here. 15 MR. ARNOLD: Oh, so there isn't a digital 16 copy of this document. 17 THE VIDEOGRAPHER: There isn't? Okay. 18 MR. ARNOLD: We just have the hard copies. 19 THE VIDEOGRAPHER: No worries. 20 BY MS. CARROLL: 21 Ω. Go ahead, Mr. Simons. 22 A. 23 24 25 So for me, I'd spent at this point five four-and-a-half, five years doing a PhD at Harvard studying what machine learning is, what AI is, and why it matters for democracy. - 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 11 of 43 1 2 3 4 5 6 7 8 9 10 Page 18 I had also spent several years working at the heart of a team at Facebook whose job it is to think about how you build AI with vulnerable people in mind to protect people, and what I learned from both that academic work and then also from being inside Facebook as a company and seeing the journey of that work, seemed really, really important to get out there into the world as a matter of fundamental importance to our society, our democracy, and what it means to be a citizen of a democracy. 11 So I had a conversation Princeton actually 12 13 14 15 16 approached me at a political science conference, and I had a conversation with an editor there who was really interested in helping me to tell the story of those ideas to an audience that was broader than just the narrow audience that you read in an academic PhD. 17 Ω. Thank you. Do you you mention 18 throughout your book an experience, accountability, and 19 language gaps. Can you explain what those are? 20 MR. HALPERIN: Object to foundation. 21 Assumes facts not in evidence. 22 A. 23 These I think are one of the most fundamental reasons, these gaps, why there's a disjunct 24 between the things has happen inside technology 25 companies and the debate that all of us as citizens who 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 12 of 43 Page 19 1 are subject to their power have. 2 And the heart of it is that AI is complicated. 3 4 5 It requires training in statistics and computer science and these disciplines that most of us don't have training in. 6 7 8 9 10 And what that means is that some of the language that technology companies use to talk about what they're doing doesn't make any sense to or is hard to unpick if you are a citizen, or frankly even experts, lawyers, judges, politicians. 11 12 13 14 15 16 17 18 And the way I saw it at this point, I had had the privilege of an incredible education in which I had got some training in the basics of statistics and how machine learning works, and had a lot of training in communication and explanation, and therefore part of my job in this book was an explanatory job to bridge the gap as described, so that folks who don't know anything about AI or Google or Facebook can understand the 19 20 21 basics of how they work and what it means for the people they impact and the effect they have on individuals and on society. 22 THE VIDEOGRAPHER: Excuse me. 23 [Discussion off the record.] 24 BY MS. CARROLL: 25 Ω. 877-370-3377 Could you provide for the jury an Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 13 of 43 Page 55 1 off the record for a second? 2 MS. CARROLL: Yeah. 3 THE VIDEOGRAPHER: We are going off the 4 record at 11:13 AM. 5 [Discussion off the record.] 6 THE VIDEOGRAPHER: We are back on the 7 record at 11:13 AM. 8 BY MS. CARROLL: 9 Ω. 10 11 Mr. Simons, did you ever raise any concerns with your superiors at Meta about Meta's prioritizing engagement over other considerations? 12 13 14 raised that concern with were themselves concerned. MR. HALPERIN: Object to form. A. Yes, I did. And some of those who I I 15 16 can give some examples if that would be helpful. BY MS. CARROLL: 17 Ω. Are there specific individuals that you 18 19 A. 20 21 22 23 24 25 can name that you spoke with? Yes. I had a conversation on several occasions with a senior VP, Margaret Stewart, who was VP of Responsible Innovation, and she and I discussed the fundamental purpose of the team that I had been asked to join and help build, which was to embed principles, consideration of harm and safety of users, early in the AI design progress. 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 14 of 43 Page 56 1 And Margaret Stewart's view she reported it 2 3 4 5 6 7 to me was that that was necessary for engineers across the company to think about harm, user harm before a model was rolled out and deployed. And we discussed on numerous occasions how to get those principles embedded in the AI design process at Facebook, and to my knowledge, while I was there, they 8 never were. 9 Ω. What was the source of the emphasis? 10 MR. LEVY: I don't know that he was done 11 with his answer to your question. 12 BY MS. CARROLL: 13 Ω. Oh, excuse me. Were you finished? 14 A. There were most of the products that I 15 worked with, whether it was in the advertising delivery 16 system, an engineering called the News 17 Feed system, Guy Rosen, or 18 19 the Instagram product, my whole job was to develop principles that would be embedded in the AI design 20 21 22 23 24 25 process at Facebook, and the whole point of doing that was that it would not just invite but require engineers and product leaders to think about harm before a product was designed, not after the fact, and in each of those cases, those senior engineers and product leaders reported the need for something like that to 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 15 of 43 Page 57 1 happen, and in the end my team was folded and disbanded 2 and folded into the another part of the 3 organization, and so those principles never were 4 deployed in the product design process. 5 6 7 And you know, the essence of that meant that engagement was prioritized over those principles that could have been embedded in Facebook's systems. 8 Ω. Were there a lot of individuals of 9 10 differing backgrounds like yourself attempting to promote this user safety work within the organization? 11 A. There were not many, no. In my team when 12 13 I joined it, I was the only person with a political science or a kind of nonengineering background in the 14 team. 15 16 17 18 19 20 21 22 23 24 I think two others joined over the period that I was an employee at Facebook, and what was important and different about the team that I was working in was that we weren't in the policy or legal or comms bit of Facebook, and those parts of the company were always sort of disrespected by the main product engineering AΙ side of the company, that that was where the pride was, that was where the promotion opportunities were, the money, and because we were in that, we were embedded within it, that what made my role and the role of the 25 others who got hired with my kind of same background on 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 16 of 43 Page 60 1 Responsibility XFN, 2 in the advertising delivery system, was how do we build the case inside 3 4 5 Facebook that we have to do this, that we have to change the way AI is built to consider safety, harm earlier on in the process? And ultimately we would 6 7 always come back to Mark Zuckerberg had to support it for it to happen. 8 BY MS. CARROLL: 9 Ω. 10 Was there an understanding amongst the Responsible AI team at the time when you were working 11 12 with the company of why it wouldn't be a priority for Mark Zuckerberg? 13 MR. HALPERIN: Object to form and 14 foundation. 15 A. Whether or not Mark Zuckerberg would 16 17 18 19 support the need to embed responsible considerations of safety into AI design at Facebook, whether he would support that, was at the first stage of my period as an employee of Facebook a matter of debate. Different 20 21 people had different views about where he would land. What I found over the course of my employment 22 23 24 25 was that every time we were involved in a decision that he actually made, that decision tended to be not to support the need for those principles to be embedded. And rarely was an explicit rationale given for 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 17 of 43 1 2 3 4 Page 61 that decision, but it always had the effect of prioritizing the existing system, engagement, meaningful social interactions, over the other considerations that we were inviting to be inside the 5 system. 6 7 8 9 10 11 So the broad pattern of decisions was that if we were proposing something that would reduce engagement that went up for the executive team, Mark Zuckerberg, to review, the decision that came back would prioritize the existing system of engagement over other safety considerations. 12 BY MS. CARROLL: 13 Ω. Thank you. I'm going to introduce Exhibit 14 Number 3. It's F. 15 [Exhibit 3 marked for identification.] 16 MS. CARROLL: Can we put a clean copy on 17 18 the here, I can put it on this and not this one. don't know why there's highlighting on this one and not I 19 that one. It's not my highlighting. 20 MR. REISER: I have a spare. 21 MR. HALPERIN: She's talking about the 22 screen. 23 MS. CARROLL: On the record. Yeah. I 24 don't want to I don't want to have I'll put it on 25 this thing over here. I don't want to misrepresent 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 18 of 43 Page 62 1 this document. 2 3 4 5 MS. CARROLL: I don't know why it has MR. ARNOLD: Okay. We can MS. CARROLL: All right. MR. ARNOLD: Let's go off the record. Yeah, let's off the record 6 for one second. 7 THE VIDEOGRAPHER: We are going off the 8 record at 11:25 AM. 9 [Discussion off the record.] 10 THE VIDEOGRAPHER: We are back on the 11 record at 11:27 AM. 12 BY MS. CARROLL: 13 Ω. 14 15 16 17 18 Mr. Simons, you have Exhibit 3 in front of you, which is a letter written in response to request for public comment to the Federal Trade Commission. And it is written by representatives from the Shorenstein Center on Media, Politics and Public Policy, submitted by Are you familiar 19 with Mr. ? 20 A. I am. We wrote a paper together at one 21 point. 22 Ω. He writes on the first page, or they write 23 to the FTC, "Our chief concern" and I'll represent 24 that as far as we know, this is how it was produced to 25 us, the highlighting. Golkow Technologies, 877-370-3377 A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 19 of 43 Page 63 1 2 3 4 5 "Our chief concern is that this is the business model that sits at the core of the major platform companies in and of itself that has caused these harms, and yet we do not have" regulatory enforcement at the national level that can "yet we have yet to see 6 treat these side effects." 7 8 What is your understanding strike that. What was your understanding of the business - 9 model of Facebook while you were an employee? 10 MR. HALPERIN: Object to form. Vague. 11 Misstates the document. 12 A. The main exposure that I had while I was 13 an employee at Facebook directly to the interaction 14 between the systems Facebook's AI systems that I was 15 working on 16 17 18 19 and the business model, how revenue was generated and profit was made, was the ad delivery system, which I worked on intensively in the summer of 2018 with a senior engineer in that system who explained to me the connection between the way the 20 system is designed and the revenue that it generates 21 for the company. 22 BY MS. CARROLL: 23 Ω. Thank you. Is it was it your 24 25 understanding when you worked for Facebook, Meta, that the business goals were prioritized over user safety? 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 20 of 43 Page 64 1 MR. HALPERIN: Object to form and 2 foundation. 3 A. The 4 5 6 7 8 the most important way that I observed the relationship between the business goals and choices about the design of the system while I was at Facebook was in specific examples of cases that went up to senior executives, and then the decisions that happened in those rooms. 9 10 11 12 13 In a few cases I was in the rooms where the discussion happened. In other cases, I was helping to prepare the team to go to review with Mark or the executive team, and then would download the review afterwards. 14 15 One specific example that I discussed with the advertising executive well, sorry, the engineer in 16 the advertising team who I 17 mentioned 18 19 20 21 22 23 24 25 we talked about the ad delivery system, which is where AI design and revenue most directly interact. And we talked in particular about the organic bid component of that system, which basically defines what's of value to someone, to a person using that system. He used an example of a mum who had just had a child, that the system knows that you've just had a child, that you're a new mum because of all the data 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 21 of 43 Page 65 1 that it has about you, and it was showing mum women 2 who just had children ads for things that would sort of 3 4 make them feel bad about their body in some way, you know, like body cream or body adjustment surgery or 5 whatever. 6 And the conversation we talked about was 7 whether - what's of value to a user in the organic bid 8 component you could change that. You could redefine 9 it. So instead of making people feel click on ads 10 11 12 13 14 15 because they feel shame about their body after they've had a baby, maybe there would be some kind of higher-order goal that you could put into that component of the ad delivery system like, how do I breastfeed, or you know, are there any downsides to bottle? Whatever. 16 And talked about how he'd, you know, worked 17 18 19 20 21 22 23 on some different ways of changing that component of the ad system, but had been unable to get those things rolled out, because ultimately they were less efficient or effective, and what that meant in that context was they would result in fewer ads being clicked on and therefore less revenue generated by the ad delivery system. 24 BY MS. CARROLL: 25 Ω. 877-370-3377 Thank you. On Page 6 of this document, Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 22 of 43 Page 67 1 BY MS. CARROLL: 2 2. 3 4 Was addiction or addictive properties a part of that discussion with anyone during your time at Facebook? 5 A. Yes. And in some of the documents that I 6 wrote whilst I was at Facebook, addiction was one of 7 8 the types of harm or the mechanisms by which harm was reproduced that I talked about with senior 9 executives Margaret Stewart, to name two 10 examples - in those documents themselves. 11 And in particular, the question we talked about 12 13 was, given that the whole News Feed system is aimed at engagement, that's a set of repeated patterns of 14 behavior clicking, liking, sharing, and on so the 15 16 17 best way to increase engagement is addiction in some way, you know, is to get people doing those forms of engagement. 18 19 20 21 22 And so the type of content that the News Feed system made viral was the kind of content that would make you feel, make users feel whatever they needed to feel to produce that type of behavior. And so I talked with News Feed executives as well as Margaret and 23 about that feedback loop, essentially, of a 24 25 system designed to promote engagement that produced content that kept people on the platform, coming back 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 23 of 43 Page 68 1 to the platform, and what to do about that system. 2 Ω. 3 Is "feedback loop" an industry term? MR. HALPERIN: Object to foundation. 4 A. We talked a lot about loops in particular, 5 6 7 learning can 8 producing 9 10 11 12 13 14 15 and feedback loops as a kind of loop. Loops is a machine learning term, and the way that machine it's almost performative by by predicting something, it can make that thing come true, especially in a closed system like Facebook. So if you predict a particular form of engagement, the system is likely to display more of that form of engagement over time, and therefore it is so important, what you choose to train and build a system to do, creates the future on that platform. 16 And so for example, who had worked 17 really deeply in the advertising delivery system before 18 he led Responsible Al we talked a lot about how AI 19 20 really raised the stakes of the impact and choices about the impact in all of Facebook's systems. 21 BY MS. CARROLL: 22 Ω. Did Mr. also work for Facebook? 23 MR. HALPERIN: Object to foundation. 24 A. I believe when I started working with him 25 he had had some role at the company. I don't recall 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 24 of 43 1 Page 73 Actually, no. Strike that. We're going to 2 do let's do G. 3 [Exhibit 5 marked for identification.] 4 MR. ARNOLD: I'll 5 MS. CARROLL: Yes. 6 BY MS. CARROLL: 7 Ω. While Mr. Arnold is passing that around 8 what is a WIP? What does WIP stand for? 9 A. That was an internal Facebook term for a 10 work-in-progress document. 11 Ω. Work in progress. 12 A. Thank you. 13 Ω. Was this document here, "Ethics at 14 Facebook: Principles and Processes," created and shared 15 on an internal platform within Meta? 16 MR. LEVY: Take a minute to look through 17 it. 18 A. Okay. 19 MR. HALPERIN: Object to foundation. 20 A. Yes, as I remember it, it was created on 21 Quip. 22 BY MS. CARROLL: 23 Ω. 24 A. What is Quip? Quip 25 I don't know is Quip's an acronym or what, but it was the internal kind of Microsoft Word 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 25 of 43 1 2 3 4 Page 74 of Facebook. It was Facebook's own internally built document system that we were instructed at this point only to use Facebook's internal document system, and not any others. 5 Ω. 6 7 8 strike that. And as you've reviewed this document, you'll notice there are lots of others individuals making comments on this Quip document. Is that a normal 9 10 11 Was that, during your time with the company, normal way of communicating progress on a project with team members in the normal course of business? a 12 MR. HALPERIN: Object to form. 13 A. 14 was a Yes, commenting in a Quip with each other was a normal way to make progress on a project 15 at Facebook when I was there. 16 BY MS. CARROLL: 17 Ω. Did you were you the author of this 18 Quip? 19 A. I was. Others edited sections of it, and 20 as you can see from the comments, fed into it. But I 21 was the author of it. 22 Ω. How did this come about? Why did you 23 start this document or write what was here? 24 MR. HALPERIN: Object to form. 25 MR. LEVY: Objection. If you could 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 26 of 43 Page 75 1 clarify the question. 2 MS. CARROLL: Sure. 3 BY MS. CARROLL: 4 Ω. 5 6 When you started this Quip, if I'm referring to that correctly, what was the purpose of doing so? 7 A. So at this point in my time at Facebook, 8 9 10 11 12 13 this was the sort of end of 2018 to the best of my recollection, and I had spent nearly half a year at the company talking to senior engineers, data scientists, product designers, folks from policy, law a very wide range of different parts of Facebook and its systems. 14 And who was the 15 16 17 who leader of Responsible AI, and reported to him and who was my manager, asked me to essentially round up the work I'd done and produce 19 20 21 22 23 18 suggestions and proposals which would then go to senior executives inside the AI part of the organization and then in other bits of the organization that followed from that work, and so the purpose of this document was to be sent to senior executives to summarize what work I had done so far. 24 Ω. Thank you. Turning to the very last page 25 of this document and I apologize that this is not 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 27 of 43 Page 80 1 Ω. In it, you discuss what machine learning 2 is. Is that correct? 3 A. Correct. 4 Ω. And "Three Substantive Challenge" 5 "Challenges." The third one is where I'd like to 6 focus. 7 And I'll represent that the third one is, "Filter Bubbles and Autonomy." 8 Can you explain what that meant at this time of 9 10 writing this while you were with the company, and why it was a challenge? 11 12 A. MR. HALPERIN: Objection. Compound. So I could take each bit, what I was 13 14 writing about and why I thought that was a challenge. This section in the document was based on my 15 conversations with and in 16 17 18 19 20 ads, because in both of those systems, there's a section of the system, a part of the system. In News Feed it's called user value, and in ads it's called the organic bids. And those bits of the system are supposed to define what's good for users. 21 22 23 24 And the words "user value" or "organic bids" you have to really get into it and speak to the people who build it to understand, what does that actually mean? And in both cases, what they really do is define 25 what's of value in terms of proxy. And a proxy means 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 28 of 43 Page 81 1 2 what's really of value to me is my happiness and the well-being of my family. 3 4 5 6 7 8 The way Facebook measures that is how many posts do I share from my wife, for example. Now, that's not really what I care about. I don't really care about how many posts I share from my wife. I care about her happiness. But they can't measure that, so they measure shares. 9 10 11 12 And the problem with optimizing for proxies is that because they're not really what we care about, they reinforce the proxy, and not the thing we really care about. So for example, in resharing posts, 13 optimizing for resharing might be might mean that 14 15 16 I'm full of love for my wife and that's why I share posts, or it might mean that I'm full of hate or fear or anxiety, and that's why I share loads of things. 17 18 19 20 And what consistently across News Feed in the user value component and the ad system in the organic bids component what the engineers building them reported to me is that by defining value in terms of 21 those proxies clicks, likes, shares, plays you 22 23 ended up incentivizing repeated patterns of behavior on the tool that were in the end undesirable, and 24 sometimes actively harmful for the users of the tool 25 themselves. In both cases they had a problem with that 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 29 of 43 Page 82 1 and wanted to figure out what to do about it. 2 BY MS. CARROLL: 3 Ω. 4 Can you remind for the jury what 's role was in the organization when you at 5 this time in 2018? 6 A. was a very experienced 7 engineer within the News Feed ranking system. 8 Ω. And this phenomenon that you've described 9 10 of filter bubbles was understood by this time in 2018. Is that correct? at 11 MR. HALPERIN: Objection. Calls for 12 speculation. Foundation. 13 A. told me that he was working 14 on understanding filter bubbles and the connected 15 concept of polarization, which was also something that 16 who comments in this document, and who I 17 worked with, was also looking into at Facebook AI 18 research. 19 BY MS. CARROLL: 20 Ω. Thank you. Let's turn to 8035. Top of 21 the page, second comment. You mentioned previously 22 What was role within the 23 company in 2018? 24 A. at this time that I was 25 working with him was a senior member of the marketing 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 30 of 43 Page 100 1 sense? 2 THE VIDEOGRAPHER: We are going off we 3 are going off the record at 12:32 PM. 4 5 [A recess was taken.] THE VIDEOGRAPHER: We are back on the 6 record at 12:48 PM. 7 BY MS. CARROLL: 8 Ω. I think we are going to move to Exhibit 9, 9 which is K. "A Vision for Responsible AI." 10 [Exhibit 9 marked for identification.] 11 BY MS. CARROLL: 12 2. Do you have the document, Mr. Simons? 13 A. I do. Sorry. Yeah. 14 Ω. Is this another example of a Quip? 15 A. This is another Quip. 16 Ω. Are you familiar with this Quip titled, "A 17 Vision for Responsible AI (Was 'SAIL 2020')"? 18 A. I am. 19 Ω. If you recall, who was the author of this 20 document, or the original creator of this Quip? 21 A. The primary author of the document was 22 who was the leader of 23 Responsible AI. 24 Ω. What did you understand at the time was 25 the purpose for this document or Quip? 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 31 of 43 Page 101 1 A. The purpose of this document was to set 2 3 4 5 out the mission and the strategy and the need for the Responsible AI team using all the work it had done to date to get senior exec buy-in for the team for its objectives and to increase its headcount. 6 Ω. And did this include work that you had 7 contributed to? 8 MR. HALPERIN: Objection. Leading. 9 A. I was one of the key people contributing 10 to this document as wrote it. He and I had 11 12 several meetings to discuss it, and he directly drew from some of the earlier memos that I had written that 13 we have talked about today. 14 BY MS. CARROLL: 15 Ω. On the third page ending in 3672. Last 16 paragraph, starting with, "Engineering excellence 17 should also apply to AI." 18 The last sentence in bold says, "Explicitly 19 reward ML engineering excellence." And let me read 20 that in full in this sentence. 21 "It will be essential to explicitly reward ML 22 engineering excellence and to give ML practitioners 23 space and permission to adopt best practices." 24 Did Mr. write that? 25 A. I don't recall if he specifically wrote Golkow Technologies, 877-370-3377 A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 32 of 43 Page 103 1 systems? 2 MR. HALPERIN: Object to foundation. 3 A. Yes. When I was working in the 4 5 6 Responsible AI team, we worked with senior engineers in many systems, in the News Feed system, the integrity system, the ad system. And it came up repeatedly the 7 8 theme that engineers were incentivized to build systems that maximized the objectives. 9 10 11 12 Those systems were set by executives, so in the News Feed case, that was meaningful social interactions. And if the systems they were building boosted those metrics, they were rewarded and promoted 13 and given headcount, and if it had a deleterious effect 14 on meaningful social interactions for example, in 15 News Feed they were not. 16 And that was, in most of the conversations that 17 I had, a really significant obstacle for embedding 18 responsibility considerations in AI design in 19 Facebook's products, because that was bound to have the 20 effect of reducing engagement, and therefore there 21 would be resistance to it, given the performance 22 rewards that existed at Facebook. 23 BY MS. CARROLL: 24 Ω. Did you have an understanding at this time 25 of whether there was a monetary incentive in the 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 33 of 43 Page 104 1 performance structure of engineers? 2 MR. HALPERIN: Object to foundation and 3 form. 4 A. 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Yes. One of the things I discussed with senior engineers in News Feed and ads was bonuses, for instance, that were awarded to those teams who significantly boosted engagement as defined in different ways in different systems, and fundamentally, by boosting engagement, what these systems were aiming to do is boost the amount that a user interacts on the platform, the time they spend on it, the frequency they with which they visit the platform, and that produces more ad revenue. And in an example, for instance, of the principles that Margaret Stewart and I and the Responsible AI team were seeking to embed in the product, if those principles would reduce engagement, and therefore reduce revenue, it was expected that we would encounter senior opposition to embedding them in the system, and part of the context for this document that we're discussing was to make the case that doing this is the right thing to do, even though it might reduce engagement and therefore reduce profit. 24 BY MS. CARROLL: 25 Ω. Thank you. And to remind the jury, who 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 34 of 43 Page 108 1 2 3 about them was used to make inferences, about either other demographic features or how it was used to train models that were then integrated in the system. 4 5 6 And the success state was describing a state of affairs in which there was a really simple, clear explanation by Facebook to users about how their data 7 was used across the Facebook product. 8 Ω. And then if we'll turn to Bates ending in 9 3677. Middle of the page. Mr. writes in a 10 bullet, "Sophisticated reinforcement learning might 11 12 13 increase the problem of tech addiction in Stories, News Feed, or Instagram." What did you understand that to refer to in this 14 document? 15 A. So that comment was based on a 16 conversation that Margaret Stewart, and I had 17 18 had in which we discussed whether AI was the primary driver of addiction in users of the product. 19 And when writes about sophisticated 20 21 reinforcement learning, what he means is one of the most advanced forms of AI that Facebook were using at 22 that point to optimize for engagement, and the risk 23 that that might supercharge problems of addiction 24 across those products. 25 Ω. 877-370-3377 If you turn to Bates ending in 3685. Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 35 of 43 Page 114 strike that. 1 2 3 4 5 Ω. Do the decisions Did your understanding of the decisions made by engineers building these machine learning models ever involve specific substance or context of a particular piece of content? 6 A. Could you explain what you mean by the 7 context of a piece of content? 8 Ω. Let me see. It might be easier to point 9 10 you to a document and not try to act like I understand all of this. 11 Let's go back to Exhibit 2 in your book. Let's 12 go to Page 137. 13 [Discussion off the record.] 14 MR. ARNOLD: You don't you don't have a 15 digital copy of this? 16 THE VIDEOGRAPHER: Not of the book. 17 MR. ARNOLD: No? 18 THE VIDEOGRAPHER: But I was saying she 19 could use the ELMO so people on Zoom can see. 20 MR. ARNOLD: Oh, I see. 21 MS. CARROLL: Thank you. 22 [Discussion off the record.] 23 BY MS. CARROLL: 24 Ω. 25 Mr. Simons, in your book you discuss an analogy using a public square. And I'm going to direct 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 36 of 43 Page 120 1 rank that high on their News Feed. 2 3 4 5 6 7 8 9 10 11 12 13 14 15 And so using the analogy in the book, when someone's entering a public square, it's a bit like the people who are handing out leaflets or literature or information thinking, "Okay, we want that person to go and do jumping jacks in the middle of the square while shouting loudly about something. We know from everything we've observed in the past that this kind of leaflet is most likely to get this person coming into the square to do jumping jacks and shout loudly, so we're going to give them that leaflet first." And that's the basic way that the ranking system determines how to order the thousands of things it could show to anyone, and therefore determine what you open when you open Facebook on your phone. 16 BY MS. CARROLL: 17 Ω. 18 So going back to Page 137. Towards the bottom of the page here, where it says it starts 19 with, "Although." "Although the corporation." 20 Do you see where I am? 21 A. Yes. 22 Ω. 23 It says, "Although the corporation cannot determine which pamphlet each individual receives as 24 25 they enter the square, the corporation doesn't much care about each individual. What it cares about are 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 37 of 43 Page 121 1 2 3 4 5 aggregate effects, how the collective mood responds to different ways of ranking information and ideas. Similarly, Facebook designs News Feed by considering how best to maximize its topline metric, not the particular content to show individuals." 6 Did I read that correctly? 7 A. You did. 8 Ω. And is that a fair representation of your 9 understanding of how Meta ranked information during 10 your time with the company? 11 A. 12 13 14 15 It is. And it stems specifically from my conversations with News Feed and integrity product leaders, VPs, and engineers in the wake of the aftermath of the meaningful social interaction rollout in 2018. 16 Ω. 17 18 scratch that. 19 20 21 Thank you. Is there anything else or any other specific examples from your time with the company that you think Are there any other examples that you personally encountered during your time with the company where Meta chose profits over user safety? 22 MR. HALPERIN: Object to form. 23 Foundation. And misstates testimony. 24 A. Yes, there are there are several which 25 I can mention. I guess the first and in some ways most 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 38 of 43 1 Page 122 important is that the team that I was hired to help 2 build, the Responsible AI team it was responsible 3 its job was to require Facebook teams to care about and 4 5 think about harm and user harm and safety before they rolled out products. 6 7 8 9 10 11 That Responsible AI team was not kept in the AI organization where it could do that. It was folded into a part of the company where it could not do that. So the failure of that team to be given the mandate to do what it was asked to do demonstrated to me that the business model was prioritized over the safety of 12 13 14 15 16 17 18 19 20 21 users. A specific part of that, where it was really brought home to me and others like Margaret Stewart in the company, as she said to me, was the embedding of principles. We pushed repeatedly to embed principles in the early stages of AI design so that senior product executives would have to explain how they were embedding principles of safety in AI design, and to my knowledge, those principles were never integrated into the AI design process. 22 The - the examples from News Feed I think are 23 most were most striking to me at the time, and to 24 those I spoke to in News Feed like or Guy 25 Rosen, the meaningful social interactions shift was 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 39 of 43 Page 123 1 2 3 widely understood to increase the sharing of content that provoked feelings of shame or disgust or fear or pride social emotions where I compare myself to you 4 and you compare yourself to others in your community. 5 6 7 8 9 10 11 12 13 you know, there were multiple research projects that I was part of understanding the harm that that caused users in provoking those social emotions. To my knowledge, meaningful social interactions is still the primary driving thing that motivates Facebook's News Feed. And that meant more interaction in the Facebook system, and that meant more user engagement on the system, and that generated more revenue for the company, but it was 14 15 16 17 18 19 20 21 And then the toxicity model that I talked about a few times was a striking example to me, because it was a clear case of when a model had been trained and built by a team who wanted to reduce the prevalence of a harmful type of content. Saying that people with big noses are bad is harmful to Jewish people, and they trained a model to detect that kind of content. It could have reduced it. 22 23 It went all the way up to Mark Zuckerberg for review, and then for whatever reason, didn't get rolled 24 out. And that was a repeated pattern of that I 25 observed in which models that could have done good, 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 40 of 43 Page 124 1 2 could have been rolled out, it went up to senior leadership, and then they were not rolled out. 3 4 5 6 And the final example I would give is the one where product design and revenue most closely intersect, which is the ad delivery system. That is the system that generates revenue for Facebook. That 7 is the heart of its business model. 8 9 10 11 The component of that system that I worked really hard to understand with senior engineers who built that system called organic bids that defines what's of value to a user. Instead of saying to a new 12 mum 13 14 you know, offering her things that would be of real benefit to her as she learned how to be a mom, the ad system was, as the engineer explained it to me at 15 16 17 that point, showing ads that would, again, provoke those social emotions of shame or like body cream or body alteration surgery. 18 That ad delivery system could have been changed 19 20 21 22 to redefine what actually is of value to a user, but that would have reduced the revenue that that ad delivery system generated, and again, to my knowledge, those changes were never rolled out. 23 So again, to me, that seemed, you know, to the 24 engineers that I was talking to like as a 25 decision to prioritize the revenue that that system 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 41 of 43 Page 125 1 generated over the safety of the users that it 2 impacted. 3 BY MS. CARROLL: 4 Ω. I think this will Thank you. -- this is 5 my last question. 6 Did 7 strike that. During your time at Meta, did the company to 8 your knowledge ever warn users of the potential risks 9 10 posed by decisions made in their ranking in content moderation systems? 11 MR. LEVY: To his knowledge? 12 MS. CARROLL: Yes, to his knowledge. 13 MR. HALPERIN: Object to form and 14 foundation. 15 MS. CARROLL: I can ask it again. Let me 16 state it again more clearly, now that I've got 17 BY MS. CARROLL: 18 Ω. 19 20 21 While you were employed by Meta, to your knowledge, did the company ever warn users about the potential risks posed by the decisions the company was making in the design of their ranking and content 22 moderation systems? 23 MR. HALPERIN: Object to form and 24 foundation. 25 A. I was not aware of any efforts by Facebook 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 Page 42 of 43 Page 126 1 2 3 4 5 6 7 8 9 10 11 to warn users of the potential impact on them of Facebook's ranking model design. This was something that I explicitly talked about with, for example, Margaret Stewart in the context of Facebook's responsibility to vulnerable users in particular, to warn them about different risks on the site, and that was one of the motivating forces behind the work on principles that Margaret and I did together, and to my knowledge, that work was not embedded in the product design process at Facebook. MS. CARROLL: Thank you, Mr. Simons. 12 That's all from plaintiffs. We can go off the record. 13 THE VIDEOGRAPHER: We are going off the 14 record at 1:36 PM. 15 [A recess was taken.] 16 THE VIDEOGRAPHER: We are back on the 17 record 2:06 PM. 18 EXAMINATION 19 BY MR. HALPERIN: 20 Ω. Good afternoon, Dr. Simons. My name is 21 Greg Halperin. I represent Meta in this litigation. 22 You and I met this morning. Is that correct? 23 A. That is correct. 24 Ω. You talked about various criticisms in 25 your testimony with the plaintiff lawyers this morning 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Case 4:22-md-03047-YGR Document 2865-6 Filed 03/20/26 CERTIFICATE Page 43 of 43 Page 237 I, John Arndt, a Certified Shorthand Reporter and Certified Court Reporter, do hereby certify that JOSHUA SIMONS MP, prior to the commencement of the examination, was sworn by me on April 1, 2025, to testify the truth, the whole truth and nothing but the truth. I DO FURTHER CERTIFY that the foregoing is a true and accurate transcript of the proceedings as taken stenographically by and before me at the time, place and on the date hereinbefore set forth. I DO FURTHER CERTIFY that I am neither a relative nor employee nor attorney nor counsel of any of the parties to this action, and that I am neither a relative nor employee of such attorney or counsel, and that I am not financially interested in this action. John Arndt JOHN ARNDT, CSR, CCR, RDR, CRR CSR NO. 084-004605 CCR NO. 1186 877-370-3377 Golkow Technologies, A Veritext Division www.veritext.com