How Policymakers Can Foster Algorithmic Accountability Joshua a - PowerPoint PPT Presentation

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG How Policymakers Can Foster Algorithmic Accountability Joshua a New @josh sh_a _a_ne new Accountab abili lity y in the Algorithmic Economy May 22, 2018

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG OV OVERV RVIEW 1. Algorithms pose new challenges 2. Existing proposals are flawed 3. Algorithmic accountability is the right approach 4. Implementing algorithmic accountability 5. Impact 6. Additional steps

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG AL ALGORITHMS POSE E NEW EW CHAL ALLENGES ES Complexity:  Many ways bias can influence an algorithm  Difficult to interpret Scalability:  Risk amplifying flaws on a large scale Model of a neural network. Source: TeXample.net.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG EXISTING PROPOSALS AL EX ALGORITHMS AR ARE E FLAWED WED Mandates for algorithmic transparency and explainability:  Hold algorithms to a standard that does not exist for humans.  Incentivize the use of less effective AI.  Assume the public and regulators could interpret source code.  Are useful in select contexts; Angela Merkel discussing algorithmic transparency. Source: Medientage. ineffective or harmful in most others.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG EX EXISTING PROPOSALS AL ALGORITHMS AR ARE E FLAWED WED Master regulatory bodies:  Ignores the need for context-specific expertise.  Assumes regulators cannot develop the expertise to understand algorithms. Elon Musk at the NGA 2017 Summer Meeting. Source: National Governors Association.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG EX EXISTING PROPOSALS AL ALGORITHMS AR ARE E FLAWED WED Generalized regulatory proposals:  Lack specifics about how to operationalize.  Rely on platitudes that do not translate to effective governance. Theresa May at Davos 2018. Source: Number 10.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG EX EXISTING PROPOSALS AL ALGORITHMS AR ARE E FLAWED WED Doing nothing:  Market forces usually provide adequate incentives:  Bad decisions hurt a company  Consumer feedback and outrage  Harms are minimal in many cases A ProPublica investigation revealing racial bias  Some use-cases are less subject in COMPAS, a risk-assessment algorithm. Source: ProPublica. to these feedback mechanisms.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG ALGORITHMI MIC ACCOU OUNTABILITY IS THE HE RI RIGHT HT APPR PPROACH Algorithmic accountability is the principle that an algorithmic system should employ a variety of controls to ensure the “operator” (i.e., the party responsible for deploying the algorithm) can:  Verify it acts in accordance with the operator’s intentions; and  Identify and rectify harmful outcomes. Pepper the robot. Source:Tokumeigakarinoaoshima.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG DEFINING ALGORITHMIC C ACCOUNTABILITY Verify it acts in accordance with the operator’s intentions:  Transparency  Explainability  Confidence measures  Procedural regularity Identify and rectify harmful outcomes:  Impact assessment Datumbox Machine Learning Framework. Source: DatumBox.  Error analysis

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG IMPLEME MENTING A ALGORI ORITHMI HMIC ACCOU OUNTABILITY Was there unfair consumer injury? YES NO Did the operator have sufficient controls to No penalty verify its algorithm worked as intended? YES NO Did the operator identify and rectify Did the operator identify and rectify harmful outcomes? harmful outcomes? YES NO YES NO Low or no penalty Medium penalty High penalty

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG IMPA PACT  Operators have a clear understanding of regulatory oversight and would proactively embrace algorithmic accountability.  Market forces would encourage adherence to algorithmic accountability.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG ADDI DDITIONAL S STE TEPS  Adopt this as the U.S. approach and advocate for its adoption abroad.  Implement specific statutes for algorithmic accountability for specific applications when appropriate.  Increase regulators’ technical expertise.  Invest in new methods for achieving algorithmic Federal Trade Commissioner Joseph Simons. Source: Andrew Harrer/Bloomberg. accountability.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG CONCL NCLUSION • Algorithms pose new challenges, but existing proposals and the EU’s approach would not be effective and would harm innovation. • Algorithmic accountability is the right solution to the challenges posed by algorithmic decision- making. European Commission. Source: Pixabay/Jai79.

THE CENTER FOR DATA INNOVATION | DATAINNOVATION.ORG THA HANK YOU OU  How Policymakers Can Foster Algorithmic Accountability: http://www.datainnovation.org/2018/05/how-policymakers-can-foster-  algorithmic-accountability/  Email me: jnew@datainnovation.org @josh_a_new 