Machine learning and human capital complementarities: Experimental evidence on bias mitigation

• Published date: 01 August 2020
• Author: Rajshree Agarwal,Evan Starr,Prithwiraj Choudhury
• DOI: http://doi.org/10.1002/smj.3152
• Date: 01 August 2020

RESEARCH ARTICLE

Machine learning and human capital

complementarities: Experimental evidence on

bias mitigation

Prithwiraj Choudhury

1

| Evan Starr

2

| Rajshree Agarwal

2

1

Harvard Business School, Boston,

Massachusetts

2

Robert H. Smith School of Business,

University of Maryland, College Park,

Maryland

Correspondence

Evan Starr, Robert H. Smith School of

Business, University of Maryland, College

Park, MD.

Email: estarr@rhsmith.umd.edu

Abstract

Research Summary:The use of machine learning

(ML) for productivity in the knowledge economy

requires considerations of important biases that may

arise from ML predictions. We define a new source of

bias related to incompleteness in real time inputs,

which may result from strategic behavior by agents.

We theorize that domain expertise of users can comple-

ment ML by mitigating this bias. Our observational

and experimental analyses in the patent examination

context support this conjecture. In the face of “input

incompleteness,”we find ML is biased toward finding

prior art textually similar to focal claims and domain

expertise is needed to find the most relevant prior art.

We also document the importance of vintage-specific

skills, and discuss the implications for artificial intelli-

gence and strategic management of human capital.

Managerial Summary:Unleashing the productivity

benefits of machine learning (ML) technologies in the

future of work requires managers to pay careful atten-

tion to mitigating potential biases from its use. One

such bias occurs when there is input incompleteness to

the ML tool, potentially because agents strategically

provide information that may benefit them. We dem-

onstrate that in such circumstances, ML tools can

make worse predictions than the prior technology vin-

tages. To ensure productivity benefits of ML in light of

potentially strategic inputs, our research suggests that

Received: 10 September 2018 Revised: 9 February 2020 Accepted: 14 February 2020 Published on: 7 April 2020

DOI: 10.1002/smj.3152

Strat Mgmt J. 2020;41:1381–1411. wileyonlinelibrary.com/journal/smj © 2020 John Wiley & Sons, Ltd. 1381

managers need to consider two attributes of human

capital—domain expertise and vintage-specific skills.

Domain expertise complements ML by correcting for

the (strategic) incompleteness of the input to the ML

tool, while vintage-specific skills ensure the ability to

properly operate the technology.

KEYWORDS

bias, complementarities, domain expertise, human capital, machine

learning

1|INTRODUCTION

Artificial intelligence (AI) and machine learning (ML)—where algorithms learn from existing

patterns in data to conduct statistically driven predictions and facilitate decisions (Brynjolfsson &

McAfee, 2014; Kleinberg, Lakkaraju, Leskovec, Ludwig, & Mullainathan, 2017)—may well trans-

form the future of work, with questions regarding whether it would substitute or complement

human capital (Autor, 2015; Bughin et al., 2017; Frank et al., 2019). Despite the promise of ML in

increasing productivity, many firms have encountered significant challenges due to biases in

predictions,

1

often thought to result from biased training data and/or algorithms (Baer &

Kamalnath, 2017; Bolukbasi, Chang, Zou, Saligrama, & Kalai, 2016; Polonski, 2018).

2

In many important contexts, however, a third source of bias may arise because agents strategi-

cally alterthe input to the algorithm, perhapsbecause they stand to benefitfrom biased predictions.

For example, ML algorithms can speed up the reviewing of resumes in recruiting, or processingof

insurance claims.However, resumes and insurance claims are generated byapplicants who have a

strategic interest in positive outcomes. Can ML correct for such strategic behavior? Research in

“adversarial”ML examines attempts to “trick”ML technologies (Goodfellow, Shlens, &

Szegedy, 2014), and generally concludes thatit is challenging to “adversarially train”the ML tech-

nology to account for every possible input. The combination of strategically generated inputs and

imperfect adversarial training of MLcreates biased predictions that stemfrom what we term input

incompleteness. Accordingly, two important questions arise: How can firms mitigate such bias to

unlock the potentialof ML? And, how may human capital complementML to do so?

We examine answers to these questions in the context of ML technology used for patent

examination, a context rife with input incompleteness. Patent examiners face a time-consuming

challenge of accurately determining the novelty and nonobviousness of a patent application by

sifting through ever-expanding amounts of “prior art.”Moreover, patent applicants are permit-

ted by law to create hyphenated words and assign new meaning to existing words to accurately

reflect novel inventions (D'hondt, 2009; Verberne, D'hondt, Oostdijk, & Koster, 2010). However,

this freedom also allows patent applicants to strategically write their applications to enhance

1

The word “bias”evokes many connotations. We use bias to reflect inaccurate predictions (e.g., Types 1 and 2 errors).

2

Gender and racial biases resulted in Amazon discontinuing an ML-based hiring technology (Dastin, 2018), IBM and

Microsoft coming under fire for ML-based facial recognition (Buolamwini & Gebru, 2018), and the Apple credit card

scrutinized through a regulatory investigation (Knight, 2019).

1382 CHOUDHURY ET AL.