Surrogation

In the games Goodhart, Campbell and Kerr describe, most of the people involved do not feel they are evading the rules. A conscientious person can chase a measure they know to be distorted with a clear conscience. Why? Because the measure quietly displaces the goal inside their head. The phenomenon is called surrogation. It explains why the most sincere people fall in too, and why moral education is basically useless against it.

Start with the scandal. Wells Fargo's strategy was supposed to be "deep customer relationships": the more a customer trusts you, the more business they do with you. Management found a proxy measure for that abstract strategy, the cross-sell ratio, meaning how many products each customer holds, with the slogan Eight is Great, eight per household. Then they wired that number to pay, internal rankings and the threat of dismissal, layer by layer.

The result: employees opened fake accounts on a massive scale without customers' knowledge. In September 2016 the consent order from the US Consumer Financial Protection Bureau (CFPB) revealed about 1.534 million unauthorized deposit accounts and about 565,000 unauthorized credit card applications. In August 2017 an expanded review commissioned by the bank (covering about 165 million accounts since 2009) revised the number of potentially unauthorized accounts up to about 3.5 million. What does 3.5 million mean? It is the entire population of a large city, each with one fake account opened in their name. The direct output of chasing the measure was the destruction of the very construct it was proxying for: the customer relationship.

The management accounting scholars Choi, Hecht and Tayler defined this class of phenomenon in 2012: people act as though the measures are the construct of interest, treating the measure of the strategy as the strategy itself. The measure was only ever a proxy for the goal, yet inside people's heads it quietly completed the substitution.

The foundation of the mechanism comes from psychology and is called attribute substitution, proposed by Kahneman and Frederick: faced with a hard question (is this company executing its strategy well), the cognitive system quietly swaps in an easy one (what was the survey score this quarter) and answers the easy one, while the person believes throughout that they are still answering the original.

There is no decision point anywhere in the substitution, no moment where a conscience could intervene. That is exactly the line between it and cheating: a cheat knows they swapped the goal, someone in the grip of surrogation does not. Choi and colleagues' contribution was to point out that a manager facing a measure sits in the ideal trigger conditions for this mechanism. The construct is abstract and hard to assess, the number is concrete and readily available. Once those conditions line up, substitution is close at hand, and unnoticed.

More importantly, they moved the mechanism into controlled experiments and measured its switches. Two sets of experiments in 2012 and 2013 produced a key pair of findings. Tying pay to a single measure significantly aggravates surrogation: pay someone to watch one number and that number turns into the goal in their head faster. Switching to multiple measures significantly relieves it, because several numbers side by side keep hinting that "no single one of these is the strategy." Note that the lever in the prescription is cognitive, not supervisory: multiple measures work not because cheating gets harder but because they dismantle the cognitive conditions for the illusion that the measure is the construct.

Back to Wells Fargo. What makes the case interesting is its two-layer structure, and the two layers are two different diseases.

The frontline layer is adversarial gaming: knowingly fabricating accounts, with internal slang for the techniques such as gaming, sandbagging and pinning, words recorded verbatim in the board's independent investigation report of April 2017. The management layer is surrogation: sincerely treating the cross-sell ratio as evidence of a healthy business, and showing investors that pretty curve year after year. One measure, driving the cheats and the sincere at the same time, two corrosions stacked. This is the full corporate version of the line in B01 about a measure quietly replacing the goal in your head.

The bill: fines totalling 185 million dollars (100 million from the CFPB, 35 million from the Office of the Comptroller of the Currency, 50 million from the Los Angeles City Attorney), about 5,300 employees dismissed; in 2018 the Federal Reserve imposed an asset cap (barring the bank from growing further, a fairly heavy instrument in the regulatory toolkit); in 2020 a further 3 billion dollar settlement with the Department of Justice and the SEC. The deferred prosecution agreement attached to the settlement covers conduct from 2002 to 2016: this collective illusion that "the measure is the strategy" ran for fourteen years. What do fourteen years tell you? Substitution is not a momentary lapse. It can become the normal cognition of a large company, passing through countless audits, board meetings and annual reports.

Experiment and case file testify for each other here. The laboratory side proves substitution needs no villains: subjects merely accepted a setup where "pay follows this number," and after a few rounds strategy in their judgment simply equalled that number. The case file side proves substitution scales: Wells Fargo wired the same cognitive switch to the pay, ranking and job security of tens of thousands of employees, and the illusion acquired industrial scale.

Mitigations have been measured out in the same literature, three of them.

First, use multiple measures, from the original 2012 experiment. Second, added by Bentley in 2019: requiring narrative reporting, writing sentences that explain performance rather than only reporting numbers, reduces distortion and surrogation. A number displaces the construct easily; a sentence forces the reporter back into contact with the construct itself. Black, Meservy, Tayler and Williams pushed further along that line in 2021. The third hides in the flip side of the first finding: since "incentives tied to a single measure" is the aggravating switch, unhooking rewards and punishments from the measure is flipping the switch back. Campbell's "keep indicators away from direct reward and punishment loops" (B02) meets this at the same point on the evidence chain. This body of experimental evidence is one of the foundations of G04 (multi-indicator checks and balances), while G02 (decoupling) directly dismantles the aggravating condition of measures being wired to rewards.

In the family division of labour, surrogation fills the psychology slot. Goodhart and Campbell describe the system-level consequence, Kerr describes the mismatch at the design end (B06), and surrogation explains the mind at the execution end, why rational and sincere people become the carriers of distortion. Principal-agent models (B12, the formal theory of information asymmetry between payer and worker) assume the agent knows what it is optimizing, and surrogation warns that even that assumption is optimistic: often the agent believes it is optimizing B while what is in its hands is already A. Push one step further and you are on the training floor. When a person substitutes, there is at least still a construct in their head that can be called back. The optimizer has only ever seen the proxy measure, so for it substitution is not a risk, it is the factory setting.

One open question, taken from Bentley's finding: why does switching to sentence-based reporting weaken substitution? If the answer is "narrative forces contact with the construct," then when the sentences in the report are drafted by a model and the reporter only reviews them, how much of that defence is left? What substitution fears is contact, and what it is best at is skipping contact.

The one-line takeaway: a measure does not just get gamed, it occupies your head; once the number has become the goal for you, cheating and hard work are no longer distinguishable from the inside.