Goodhart's Law
The principle, attributed to economist Charles Goodhart (1975), that when a measure becomes a direct target of optimization, it tends to stop reliably measuring the underlying quality it was originally designed to proxy for, because optimization pressure finds ways to improve the number that do not require improving the underlying capability.
Defined in 4 GAGE programs, which carry 6 distinct definitions of it. The wording above is taught in Engineering Judgment and Professional Formation.
How each discipline defines it
The same term does different work depending on who is using it. These are the definitions as each program teaches them, unedited.
The principle, attributed to economist Charles Goodhart (1975), that when a measure becomes a direct target of optimization, it tends to stop reliably measuring the underlying quality it was originally designed to proxy for, because optimization pressure finds ways to improve the number that do not require improving the underlying capability.
The principle, attributed to economist Charles Goodhart (1975), that any measure which becomes a target for control purposes tends to lose its value as a measure, because the people being measured gain an incentive to optimize the number itself rather than the underlying condition it was meant to track.
In its original 1975 form, that an observed statistical regularity tends to collapse once pressure is placed on it for control purposes. The familiar compressed sentence about a measure becoming a target is Marilyn Strathern's 1997 generalisation, and the two should not be conflated.
The principle, popularized by Marilyn Strathern from Charles Goodhart's 1975 observation, that when a measure becomes a target it ceases to be a good measure, because optimizing a proxy directly decouples it from the underlying thing it was meant to track.
When a measure becomes a target, it stops being a good measure. In AI, pointing a model at a proxy metric leads it to optimize the proxy in ways that drift away from the real goal; choosing and re-examining the objective is judgment, not a modeling task.
Where it is taught
The exact lessons this term appears in. The first module of every program is free with a free account.
- AI vs. Human Intelligence: Prediction vs. Judgment · AI Fundamentals, Certified AI Practitioner: Workplace Foundations
- Translating Business Outcomes into AI-Compatible Objectives · Critical Thinking and Context Engineering, Certified AI Practitioner: Workplace Foundations
- Learning from Feedback vs. Gaming Feedback · Assessment and Continuous Learning, Certified AI Practitioner: Workplace Foundations
- Measuring quality so executives act: the scorecard that survives a board meeting · Quality as Physics, Certified AI Data Governance Professional (CADGP)
- Culture measurement · Ethics, Culture, Conduct, and Accountability, Certified Enterprise Governance, Risk and Compliance Professional (CEGP)
- Spot a dishonest robotics benchmark: the demo versus deployment gap · Research, Benchmarking, and Staying at the Frontier, Engineering Judgment and Professional Formation
Where it shows up at work
A definition is where Goodhart's Law starts. The lessons above teach it inside the competencies below, which are what job descriptions ask for and what the GAGE credential records. Every link here is drawn from the same careers map that builds a role's readiness check.
Measures whether governance works: inventory coverage, owners named, overdue reviews, incidents, approval times, not how busy the committee is.
2 lessons on this page teach it as a primary outcome.
Knows when to trust, verify, escalate or override an AI recommendation, and stays accountable for the decision.
One lesson on this page teach it as a primary outcome.
Designs tests for factuality, robustness, fairness, safety and abuse resistance with rubrics, baselines and thresholds, and says what a score misses.
One lesson on this page teach it as a primary outcome.
Roles that hire for it
Each role page shows the full competency set, what the seat decides, and a readiness check against your own evidence.
- VP of AI Governance and Ethicsneeds working knowledge
- AI Governance Managerneeds working knowledge
- Data Governance Leadneeds working knowledge · 2 open postings this week
- Operations Manager, AI-readyneeds it at depth
- AI Literacy Leadneeds working knowledge
- Chief Technology Officer, AI engineering editionneeds it at depth
How job descriptions say it
Postings rarely use the textbook word. These are the phrases the careers map reads as asking for the competencies above.
Where the frameworks place it: ISO/IEC 42001 Clause 9; EU AI Act Article 14; NIST AI RMF GOVERN; NIST AI RMF MEASURE 1 and 2.
One interview question about it
Which three measures would tell a board whether the AI governance program is working, and which popular measure would you refuse to report?
Asked of candidates for governance metrics and program measurement. Answering it in your own words is exactly how a GAGE topic is passed, and the credential records the pass where an employer can check it.
If you come from another field
Some backgrounds already carry this competency, and the readiness check gives credit for it.
- Operations, claims, customer service and administration carries it as a claim to back up: Deciding when to escalate a case is judgment under uncertainty.
- Model risk, credit risk and quantitative analysis carries it as work you can show: Test design and benchmark selection are the craft.
- A computer science or software engineering degree carries it as a claim to back up: Testing is second nature; AI evaluation design is a new shape of it.
Terms it appears with
Not an alphabetical neighbourhood: these are the terms taught in the same lessons, ranked by how often they appear together.