Uprisings for racial justice are sweeping the country. Following the police murders of George Floyd, Breonna Taylor, and so many others, named and unnamed, America has finally reached its moment of reckoning. And politicians are starting to respond. That starts with banning facial recognition, a technology perfectly designed for the automation of racism. Tawana Petty is director of the Data Justice Program at Detroit Community Technology Project and co-leads the Our Data Bodies project.

Artificial intelligence (AI) is expected to play a major role in shaping global competitiveness and productivity over the next couple of decades, granting early adopters significant societal, economic, and strategic advantages. As the pace of AI innovation and development picks up--underpinned by advancements in big data and high-performance computing--the United States and China are both in the driver's seat.

As machine learning is increasingly deployed in high-stakes contexts affecting people's livelihoods, there have been growing calls to open the black box and to make machine learning algorithms more explainable. Providing useful explanations requires careful consideration of the needs of stakeholders, including end-users, regulators, and domain experts. Despite this need, little work has been done to facilitate inter-stakeholder conversation around explainable machine learning. To help address this gap, we conducted a closed-door, day-long workshop between academics, industry experts, legal scholars, and policymakers to develop a shared language around explainability and to understand the current shortcomings of and potential solutions for deploying explainable machine learning in service of transparency goals. We also asked participants to share case studies in deploying explainable machine learning at scale. In this paper, we provide a short summary of various case studies of explainable machine learning, lessons from those studies, and discuss open challenges.

The requirements on explainability imposed by European For decisions adopted by public authorities, two stronger laws and their implications for machine learning requirements are studied: motivation obligations for administrations (ML) models are not always clear. In that perspective, and for judges (imposed by European Convention our research (Bibal et al., Forthcoming) analyzes explanation on Human Rights). For administrative decisions, all factual obligations imposed for private and public decisionmaking, and legal grounds on which the decision is based should be and how they can be implemented by machine provided. For judicial decisions, judges have in addition to learning techniques. For decisions adopted by firms or individuals, we mainly The objectives of those explanation requirements are focus on requirements imposed by general European legislation twofold: first, allowing the recipients of a decision to understand applicable to all the sectors of the economy.

While it is still unclear if agents with Artificial General Intelligence (AGI) could ever be built, we can already use mathematical models to investigate potential safety systems for these agents. We present an AGI safety layer that creates a special dedicated input terminal to support the iterative improvement of an AGI agent's utility function. The humans who switched on the agent can use this terminal to close any loopholes that are discovered in the utility function's encoding of agent goals and constraints, to direct the agent towards new goals, or to force the agent to switch itself off. An AGI agent may develop the emergent incentive to manipulate the above utility function improvement process, for example by deceiving, restraining, or even attacking the humans involved. The safety layer will partially, and sometimes fully, suppress this dangerous incentive. The first part of this paper generalizes earlier work on AGI emergency stop buttons. We aim to make the mathematical methods used to construct the layer more accessible, by applying them to an MDP model. We discuss two provable properties of the safety layer, and show ongoing work in mapping it to a Causal Influence Diagram (CID). In the second part, we develop full mathematical proofs, and show that the safety layer creates a type of bureaucratic blindness. We then present the design of a learning agent, a design that wraps the safety layer around either a known machine learning system, or a potential future AGI-level learning system. The resulting agent will satisfy the provable safety properties from the moment it is first switched on. Finally, we show how this agent can be mapped from its model to a real-life implementation. We review the methodological issues involved in this step, and discuss how these are typically resolved.

Algorithm selection (AS) deals with the automatic selection of an algorithm from a fixed set of candidate algorithms most suitable for a specific instance of an algorithmic problem class, where "suitability" often refers to an algorithm's runtime. Due to possibly extremely long runtimes of candidate algorithms, training data for algorithm selection models is usually generated under time constraints in the sense that not all algorithms are run to completion on all instances. Thus, training data usually comprises censored information, as the true runtime of algorithms timed out remains unknown. However, many standard AS approaches are not able to handle such information in a proper way. On the other side, survival analysis (SA) naturally supports censored data and offers appropriate ways to use such data for learning distributional models of algorithm runtime, as we demonstrate in this work. We leverage such models as a basis of a sophisticated decision-theoretic approach to algorithm selection, which we dub Run2Survive. Moreover, taking advantage of a framework of this kind, we advocate a risk-averse approach to algorithm selection, in which the avoidance of a timeout is given high priority. In an extensive experimental study with the standard benchmark ASlib, our approach is shown to be highly competitive and in many cases even superior to state-of-the-art AS approaches.

This week AI searches for war crimes, we ask how AI shifts power, one bank's approach to unifying their AI effort, deep learning used for automatic basketball video production, and new GPUs available in Google Cloud. Don't ask if artificial intelligence is good or fair, ask how it shifts power (Nature, 07/07/2020) NVIDIA's AI-focused Ampere GPUs are now available in Google Cloud (AI News, 07/08/2020)

While the COVID-19 crisis got most folks thinking about face masks and toilet paper, Chris Calabrese was pondering artificial intelligence and its implications for public policy. His aha moment came when he realized Facebook had sent home most of its human overseers and put AI in charge of policing the social forum for inappropriate content. "The result has been systems that don't work as well. They are taking down groups dedicated to sewing masks, just because they are falsely flagged," said Calabrese, vice president of policy at the Center for Democracy and Technology. "That's automation being used by one of the most influential companies in the world, and it's still not up to snuff. That gives me a sense of how far we have to go." Facebook's stuttering steps into automation reflect broader ethical challenges faced by public tech leaders as AI, biometrics and surveillance technologies increasingly enter the mainstream.

![Figure][1] GRAPHIC: DAVIDE BONAZZI/SALZMANART Modern governments gather information across an extraordinary range of activities and use this information to direct policy. Whether a central bank monitoring inflation or a health agency monitoring disease, these entities typically publicly disclose the information gathered so that their actions can be reviewed and evaluated by others. But in many respects, the justice system is a glaring exception. In the United States, a range of technical and financial obstacles blocks large-scale access to public court records—all but foreclosing their use to direct policy. Yet a growing body of empirical legal research demonstrates that systematic analyses of court records could improve legal practice and the administration of justice. And although much of the legal community resists quantitative approaches to law, we believe that even the skeptics will be receptive to quantitative feedback—so long as it is straightforward, apolitical, and incontrovertible. We offer an example of this kind of feedback as well as a collaborative research agenda to dismantle access barriers to court records and enable the public to analyze them. Although court records in the United States sit in the public domain, federal courts charge $0.10 per printed page to view any record online ([ 1 ][2]). Accessing a single case might cost $10 or more. Accessing all cases from a given year would cost millions of dollars ([ 2 ][3]). To be sure, the federal judiciary releases inhouse studies that use federal court records, as well as a database of basic information about each case, such as the subject matter (e.g., tort, contract, civil rights) and disposition (e.g., settled, transferred, jury verdict) ([ 3 ][4]). The federal judiciary has steadfastly refused, however, to make the underlying public court records freely accessible. Selective access is not the approach taken by the rest of the U.S. federal government: Congressional records are freely available at [congress.gov][5]. Executive agencies' records are freely available at [regulations.gov][6]. It's hard to conceive of a compelling argument for selective access to judicial records that does not apply equally to selective access to congressional records or federal agencies. More to the point, it's hard to conceive of a reason why public records should not generally be accessible to the public. There are some alternative sources for court records, but barriers to systematic analysis remain. Commercial legal services have directly purchased many court records, but they impose their own fees, prohibit bulk downloads, and thus foreclose systematic analysis even for subscribers. Individual judges and commercial services occasionally grant ad hoc fee reductions for research purposes, but these grants are rare, cumbersome to acquire, limited to subsets of the data, and always come with the condition that the underlying records are not disclosed to the public ([ 4 ][7]). An open alternative, Free Law Project , maintains a crowdsourced repository of free court records, but coverage remains too low to support systematic research. The lack of access to court records seemingly undercuts any claim that the courts are truly “open” ([ 5 ][8], [ 6 ][9]). It surely conflicts with researchers' conception of openness. Scientific practice is grounded on a commitment to sharing data and enabling others to replicate findings. But the law's conception of openness is different, a commitment to carrying out public acts in a public space. A scientist might restrict access to a lab and still claim that the research she conducts there is “open.” Closed proceedings in a legal setting, on the other hand, are only tolerated in extraordinary circumstances. Also in contrast to scientific practice, much of the legal profession resists quantitative or evidence-based approaches to improving legal practice and instead prefers to rely on personal experience and professional judgment ([ 7 ][10]). In a recent Supreme Court case challenging the constitutionality of partisan gerrymandering, Chief Justice John Roberts summarily dismissed empirical approaches to gerrymandering as “sociological gobbledygook” that any “intelligent man on the street” would denigrate as “a bunch of baloney” ([ 8 ][11]). Such skepticism is by no means confined to the United States. France, for example, has recently prohibited the publication of any statistical analysis of a judge's or clerk's decisions “with the object or effect of evaluating, analyzing, comparing or predicting their actual or supposed professional practices.” Violators face up to 5 years in prison ([ 9 ][12]). We believe that these differences help explain why the lack of large-scale access to data is not viewed as a priority—or even as a concern—by much of the legal community. The differences in priorities reflect not just commitments to different values but different conceptions of the same values. Yet, if court records are to be truly accessible and evaluable by the public, the legal and scientific communities must cooperate, and appreciate the values that the other holds dear. Access to justice is a fundamental right and the foundation of any fair and legitimate justice system. But how can one quantify and empirically evaluate this concept? Consider court fees. For a litigant without means, court fees are a substantial barrier to the civil justice system. Anyone who files a lawsuit in federal court must pay a $400 filing fee, along with other costs related to litigation such as formal service of the complaint. Litigants in need can file an application to waive court fees, but there is no uniform standard to review these requests ([ 10 ][13]). Application forms differ by district. Most ask the applicant to list sources of income, assets, and cash on hand—and then leave the decision to the judge's discretion. Individual judges thus have considerable power over whether to grant or deny access to the justice system. How do judges exercise this power? This is but one of the myriad questions that is difficult, and arguably impossible, to answer without easy access to structured court records. Even with free access to the data, the answer would be difficult to infer without being able to computationally analyze the text of the court records. In this case, the analysis is straightforward. When a party submits a fee waiver request, the case docket report adds a separate entry for that request, and the textual summary accompanying the entry typically includes some reference to whether the request was granted or denied. We analyzed these entries to compute the grant rate of each federal judge in 2016. Average grant rates naturally differ among federal districts because cases are not randomly assigned to districts. However, once a case is filed in, say, San Francisco, it is then randomly assigned to one of the judges sitting in the federal district that includes San Francisco. Thus, if all judges reviewed fee waiver applications under the same standard, then grant rates should not systematically differ within districts. We find, however, that they do (see the figure). At the 95% confidence level, nearly 40% of judges—instead of the expected 5%—approve fee waivers at a rate that statistically significantly differs from the average rate for all other judges in their same district. In one federal district, the waiver approval rate varies from less than 20% to more than 80%. These findings were recently presented to a group of federal judges who are responsible for amending the rules in their local district. On learning of the inconsistent treatment of fee waiver requests, these judges expressed interest in using our data to improve the decision-making process ([ 11 ][14]). We count this as an early and encouraging validation of our claim that judges will be especially receptive to quantitative feedback that is straightforward, apolitical, and incontrovertible. Going forward, we believe that the best way to provide the judiciary with quantitative feedback is to develop a forum where individuals can collaborate and build on each other's efforts. With this vision in mind, we propose a three-pronged collaborative research agenda to empower the public to access and analyze court records. ### Make court records free In theory, Congress could make federal records free by repealing the laws that authorize the judiciary to charge for access ([ 12 ][15]), or the Judicial Conference of the United States (the policy-making body of the federal judiciary) could stop charging fees. Both Congress and the courts have rejected calls to do so. A principal reason, it seems, is money. About 2% of the federal judiciary's budget comes from online record access fees ($145 million in fiscal year 2019). The judiciary is naturally unwilling to forgo this revenue without a commensurate increase from Congress, and Congress, for its part, is unwilling to increase funding. The stalemate persists because not enough judges, members of Congress, and people realize that this is an issue of legitimacy, not just an issue of money. To break this impasse, we believe that organizations outside government should directly purchase and publicize court records. The most impactful first step is to make docket reports accessible. A docket report is essentially a lawsuit's table of contents. It lists the case title, presiding judge, subject matter of the suit, and information on the plaintiffs, defendants, and their attorneys. A docket report also gives the date that a document was filed, along with a summary of the document that can be analyzed to extract important features of a case. The data for the figure, for example, were constructed by parsing docket reports, not the underlying court records. Though docket reports represent only a fraction of all court records, acquiring them will be expensive. The docket reports used in the figure, which cover all cases filed in 2016, cost more than $100,000. ![Figure][1] Inconsistency in judicial fee waiver decisions Litigants filed 34,001 applications to waive court fees in U.S. federal courts in 2016. For visual simplification, we show only the 294 judges (out of 1742 total) who ruled on at least 35 applications. We would expect 5% of judges to differ from their within-district peers at 95% confidence. Instead, we find that nearly 40% of judges differ. GRAPHIC: X. LIU/ SCIENCE ### Link data in a knowledge network Because court records are mostly unstructured text, researchers will need to dedicate extensive time and resources to organizing the data. Documents must be analyzed using natural language processing; entities must be disambiguated; and events, such as the filing of a fee waiver, must be classified using machine learning. The docket reports should also be linked to external metadata such as information on judges, litigants, and lawyers. By linking court records to outside data sources, individual users can conduct more powerful searches, such as for litigation against big tech firms or for suits currently pending against the federal government. Although we already have solutions to many of the problems associated with organizing and classifying the data, for many more we will need additional research. For example, it is straightforward to link the presiding judge of each case to outside data on the judge's characteristics such as age, gender, and appointing president. By contrast, to assemble information about litigants and lawyers, researchers will need to make considerable progress on named-entity recognition techniques while protecting litigants' and third parties' privacy. We believe that an open and collaborative platform is the best way to make substantial and rapid progress on these challenges. ### Empower the public The ultimate goal must be to enable the public to directly evaluate and engage with the work of the courts. To this end, we should create applications that not only support scholars and researchers who may want to analyze the data but also enable members of the judiciary, entrepreneurs, journalists, potential litigants, and concerned citizens to learn more about the functioning of the courts. To support inquiries made by the public, we should develop applications that can process natural language queries such as “What are the most recent data privacy cases?” or “How often do police officers invoke qualified immunity?” Funding the efforts we propose will be challenging because the cause does not slot nicely into standard philanthropic categories. To carry out our proposals, the academic community should partner with other stakeholders such as nongovernmental organizations, law firms, legal clinics, and other advocacy groups. Indeed, we believe that one of the main reasons why past calls for change failed is because they were not coordinated. Opening up court records could lead to some flawed or misleading analyses, yet such problems apply to any setting with open data. No one can control what people do with congressional records, federal agency records, census data, etc. Nevertheless, these data are—and should remain—available to everyone. As in any discipline, standards and best practices eventually emerge, and there is already a thriving literature of empirical legal studies. Many scholars have engaged with these data, albeit on a smaller scale. Thus, for the most part, standards and best practices already exist ([ 13 ][16]). We believe that the judiciary should be shielded from outside pressures so that it can decide cases according to the law, not the latest poll. But the judiciary also acts on behalf of the public. Its independence must therefore be balanced with commensurate transparency. Ultimately, the judiciary's principal asset is not its annual appropriation from Congress or the revenue generated by access fees, but the public trust. And the most effective way to cultivate this trust—to promote transparency, dismantle barriers to access ([ 14 ][17], [ 15 ][18]), and build an open knowledge network—is to do it together. 1. [↵][19]Public Access to Court Electronic Records (PACER), “PACER user manual for CM/ECF courts” (United States Courts, 2019). 2. [↵][20]United States Courts, Federal judicial caseload statistics 2018 (2018); [www.uscourts.gov/statistics-reports/federal-judicial-caseload-statistics-2018][21]. 3. [↵][22]1. W. Hubbard , J. Empir. Leg. Stud. 14, 474 (2017). [OpenUrl][23] 4. [↵][24]1. J. B. Gelbach , Yale Law J. 121, 2270 (2011). [OpenUrl][25] 5. [↵][26]1. A. Bronstad , “PACER fees harm judiciary's credibility, Posner says in class action brief,” 25 January 2019; [www.law.com/2019/01/25/pacer-fees-harm-judiciarys-credibility-posner-says-in-class-action-brief/][27]. 6. [↵][28]1. L. Doggett, 2. M. J. Mucchetti , Tex. Law Rev. 69, 643 (1990). [OpenUrl][29] 7. [↵][30]1. H. F. Lynch et al ., Science 367, 1078 (2020). [OpenUrl][31][Abstract/FREE Full Text][32] 8. [↵][33]Gill v. Whitford, Transcript of oral argument at 38 and 40, no. 16-1161, 138 S. Ct. 1916 (2018). 9. [↵][34]1. J. Tashea , “France bans publishing of judicial analytics and prompts criminal penalty,” ABA Journal, 7 June 2019; [www.abajournal.com/news/article/france-bans-and-creates-criminal-penalty-for-judicial-analytics][35]. 10. [↵][36]1. A. Hammond , Yale Law J. 128, 1478 (2018). [OpenUrl][37] 11. [↵][38]Owing to the preliminary nature of discussions, the identities of courts and judges are not reported, but Science has confirmed this claim. 12. [↵][39]28 U.S. Codes §§ 1913, 1914, 1926, 1930, 1932. 13. [↵][40]1. W. Baude et al ., Univ. Chic. Law Rev. 84, 37 (2017). [OpenUrl][41] 14. [↵][42]1. A. Madison , “Team tapped to review PACER amid fee dispute (corrected),” Bloomberg Law, 9 January 2020; . 15. [↵][43]1. A. Kragie , “Court transparency bill calls for live audio, free PACER,” 2 March 2020; [www.law360.com/articles/1249148][44]. Acknowledgments: We thank K. Sanga for valuable feedback. This research was supported by a gift from John and Leslie McQuown and by the National Science Foundation Convergence Accelerator Program under grant no. 1937123. The data and code used for this article, along with full replication instructions and additional discussion of the analyses, are available at and at Zenodo (10.5281/zenodo.3905128). [1]: pending:yes [2]: #ref-1 [3]: #ref-2 [4]: #ref-3 [5]: http://congress.gov [6]: http://regulations.gov [7]: #ref-4 [8]: #ref-5 [9]: #ref-6 [10]: #ref-7 [11]: #ref-8 [12]: #ref-9 [13]: #ref-10 [14]: #ref-11 [15]: #ref-12 [16]: #ref-13 [17]: #ref-14 [18]: #ref-15 [19]: #xref-ref-1-1 "View reference 1 in text" [20]: #xref-ref-2-1 "View reference 2 in text" [21]: http://www.uscourts.gov/statistics-reports/federal-judicial-caseload-statistics-2018 [22]: #xref-ref-3-1 "View reference 3 in text" [23]: {openurl}?query=rft.jtitle%253DJ.%2BEmpir.%2BLeg.%2BStud.%26rft.volume%253D14%26rft.spage%253D474%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [24]: #xref-ref-4-1 "View reference 4 in text" [25]: {openurl}?query=rft.jtitle%253DYale%2BLaw%2BJ.%26rft.volume%253D121%26rft.spage%253D2270%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [26]: #xref-ref-5-1 "View reference 5 in text" [27]: http://www.law.com/2019/01/25/pacer-fees-harm-judiciarys-credibility-posner-says-in-class-action-brief/ [28]: #xref-ref-6-1 "View reference 6 in text" [29]: {openurl}?query=rft.jtitle%253DTex.%2BLaw%2BRev.%26rft.volume%253D69%26rft.spage%253D643%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [30]: #xref-ref-7-1 "View reference 7 in text" [31]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DLynch%26rft.auinit1%253DH.%2BF.%26rft.volume%253D367%26rft.issue%253D6482%26rft.spage%253D1078%26rft.epage%253D1080%26rft.atitle%253DOvercoming%2Bobstacles%2Bto%2Bexperiments%2Bin%2Blegal%2Bpractice%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aay3005%26rft_id%253Dinfo%253Apmid%252F32139532%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [32]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEzOiIzNjcvNjQ4Mi8xMDc4IjtzOjQ6ImF0b20iO3M6MjI6Ii9zY2kvMzY5LzY1MDAvMTM0LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ== [33]: #xref-ref-8-1 "View reference 8 in text" [34]: #xref-ref-9-1 "View reference 9 in text" [35]: http://www.abajournal.com/news/article/france-bans-and-creates-criminal-penalty-for-judicial-analytics [36]: #xref-ref-10-1 "View reference 10 in text" [37]: {openurl}?query=rft.jtitle%253DYale%2BLaw%2BJ.%26rft.volume%253D84%26rft.spage%253D37%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [38]: #xref-ref-11-1 "View reference 11 in text" [39]: #xref-ref-12-1 "View reference 12 in text" [40]: #xref-ref-13-1 "View reference 13 in text" [41]: {openurl}?query=rft.jtitle%253DUniv.%2BChic.%2BLaw%2BRev.%26rft.volume%253D84%26rft.spage%253D37%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [42]: #xref-ref-14-1 "View reference 14 in text" [43]: #xref-ref-15-1 "View reference 15 in text" [44]: http://www.law360.com/articles/1249148

RICHARD DAWKINS: When we come to artificial intelligence and the possibility of their becoming conscious, we reach a profound philosophical difficulty. I am a philosophical naturalist; I'm committed to the view that there is nothing in our brains that violates the laws of physics, there's nothing that could not, in principle, be reproduced in technology. It hasn't been done yet; we're probably quite a long way away from it, but I see no reason why in the future we shouldn't reach the point where a human-made robot is capable of consciousness and of feeling pain. JOANNA BRYSON: So, one of the things that we did last year, which was pretty cool, the headlines, because we were replicating some psychology stuff about implicit bias--actually, the best one is something like'Scientists show that AI is sexist and racist and it's our fault,' which, that's pretty accurate because it really is about picking things up from our society. Anyway, the point was, so here is an AI system that is so humanlike that it's picked up our prejudices and whatever and it's just vectors.