Slides for paper presented at Data for Policy Conference September 2017.
Abstract
Algorithmic decision-making systems based on artificial intelligence and machine learning are enabling unprecedented levels of personalisation, recommendation and matching. Unfortunately, these systems are fallible, and their failures have costs. I develop a formal model of algorithmic decision-making and its supervision to explore the trade- offs between more (algorithm-facilitated) beneficial deci- sions and more (algorithm-caused) costly errors. The model highlights the importance of algorithm accuracy and human supervision in high-stakes environments where the costs of error are high, and shows how decreasing returns to scale in algorithmic accuracy, increasing incentives to ’game’ popular algorithms, and cost inflation in human supervision might constrain optimal levels of algorithmic decision-making.
2. Motivation
An explosion of algorithmic decision-making An explosion of algorithmic error
Algorithms will always make mistakes
(Even without bias, manipulation or catastrophic failure)
How do we balance the benefits of more algorithmic decision-making and the costs of more
algorithmic errors? What are some important factors in these situations?
I approach the issue from an economics angle, focusing on 3 questions:
Risk, Supervision and Scale
3. Literature
"...in an information-rich world, the wealth of information means a
dearth of something else: a scarcity of whatever it is that
information consumes. What information consumes is rather
obvious: it consumes the attention of its recipients. Hence a wealth
of information creates a poverty of attention and a need to allocate
that attention efficiently among the overabundance of information
sources that might consume it" (Simon 1971, pp. 40–41)
• Attention is scarce and decision-making is not perfect
• Organisations invest on technologies that economise on attention:
routines, heuristics and algorithms (see also Agrawal el al, 2017).
• They design organisational structures to manage error (Sah and
Stiglitz, (1984, 1985, 1988) -> This provides the foundation for the
model I sketch in the rest of the talk.
4. Model #1 (Risk)
We consider an organisation where an algorithm (a1) needs to process informational
inputs, making decisions about their quality.
• The quality of the input pool is represented by α (share of good inputs)
• p11 is the true positive rate and p12 is the false positive rate.
• Accepting a good input creates a benefit r1, accepting a bad input has a cost -r2
Input pool quality α
a1
Rejected Accepted
r1 -r2
(1- α )p12α p11
The expected benefit of a
decision is positive if:
Algorithms operating in high stakes / low
quality environments should be accurate
5. Model #2 (supervision)
We introduce a human supervisor a2 who validates a share t of the algorithm’s decisions
with a true positive rate p21 and a false positive rate p22. Her cost is C2
Input pool quality α
Accepted
a1 Accepted
r1 -r2
α(1-t) p11
Validated
a1 decision-making
process
Not validated
r1
-r2
(1-α)(1-t) p12
α t p11 p21 (1-α) t p12 p22
The net contribution of the supervisor is
positive if
Supervisors more valuable in high stakes /low-quality input
environments, and if they are not costly or highly productive
!
6. Model #3 (scale)
We consider what happens when the number of inputs (and decisions) increase. We
assume t stays constant, and that r1, r2 do not change with more decisions.
We assume the organisation’s labour costs are a function of its developer and supervisor
workforces La1 and La2
Inputs (+)
? (-) (+) (+++)
Gains in accuracy vs
increase in variance
CostsBenefits
+ incentives to game the
algorithm
Highly productive
developers (skills
shortage?)
Low productivity
supervisors. Baumol’s
disease?
Eventually diminishing returns set in?
7. Implications
1. Finding the right algorithm-domain fit
• Domains with different stakes require algorithms with
different accuracies (e.g. recommendation engine vs
criminal justice system)
• Government by algorithm could get expensive if it requires
substantial human supervision
2. On the costs and benefits of human supervision
• Crowdsourcing can reduce supervision costs...but it creates
a new type of algorithmic unfairness.
• Human supervision can help detect & address sudden
declines in performance, specially where costs are harder to
measure or external to the organisation
8. Conclusion
Extensions
• Consider alternative and more complex organisational designs
• Extend to algorithmic discrimination situations
• Endogeneise quality α to cover games between platforms and bad
actors
• Explore effects of scale on benefits and costs (r1, r2)
Applications
• Operationalise, simulate and experiment
• Make Economics part of “a practical and broadly applicable social-
systems analysis [that] thinks through all the possible effects of AI
systems on all parties [drawing on] philosophy, law, sociology,
anthropology and science-and-technology studies, among other
disciplines” (Calo and Crawford, 2016)