Distinguishing Gaming from Genuine Improvement

40 concepts25 papers1 proposal

In standard machine learning settings, data points are treated as passive observers. In high-stakes social domains like lending, hiring, or admissions, however, the data points are human agents who actively modify their features to obtain a favorable classification. Early frameworks in performative prediction often treated all user adaptations uniformly as utility-maximizing attempts to circumvent a model's rules (Gavin Brown 0003 et al. in Performative Prediction in a Stateful World).

However, treating all behavioral responses as identical, superficial manipulation overlooks a critical distinction. When an individual alters their features, they might be engaging in gaming (manipulating proxy features to improve their score without changing their underlying qualifications) or genuine improvement (exerting productive effort that actually enhances the latent trait the evaluator cares about) (Jon Kleinberg et al. in Designing Evaluation Rules That Are Robust to Strategic Behavior). If a system cannot separate these two behaviors, it risks rewarding deceptive shortcuts while failing to provide meaningful pathways for actual self-improvement (Saba Ahmadi et al. in On classification of strategic agents who can both game and improve; Tian Xie et al. in Automating Data Annotation under Strategic Human Agents: Risks and Potential Solutions).

The Causal Mechanism of Strategic Behavior

To understand why some actions are beneficial and others are purely manipulative, we must look at the causal relationships between features and target labels. As demonstrated by John Miller et al. in Strategic Classification is Causal Modeling in Disguise, strategic classification is fundamentally a causal discovery problem.

Imagine a bank predicting loan default based on two features: credit card utilization and the number of open credit lines.

  • Genuine Improvement: If an applicant works extra hours to pay down their debt, they decrease their credit utilization. Because having less debt causally reduces the likelihood of default, this feature change represents genuine improvement—the underlying target label changes along with the feature.
  • Superficial Gaming: Conversely, if an applicant opens a new credit line purely to increase their total limit and artificially lower their utilization ratio, they have manipulated the proxy feature without altering their actual financial stability. The target label remains unchanged.

By framing strategic classification using structural causal models, we can formally separate genuine adaptation from fraudulent misreporting (Dylan Zapzalka et al. in Disentangling misreporting from genuine adaptation in strategic settings: a causal approach). If a feature lies on a causal path leading to the target outcome, modifying it constitutes improvement. If the target outcome causes the feature, or if they are merely correlated through a confounding variable, modifying the feature is gaming.

Shifting Thresholds to Alter the Cost Benefit Trade off

How can a classifier protect its predictive accuracy when it cannot directly observe whether an agent is gaming or improving? The primary tool is strategic threshold adjustment.

When a classifier anticipates that users will attempt to flip their outcomes, it can proactively shift its decision boundary upward (Hidde Fokkema et al. in The Risks of Recourse in Binary Classification). By implementing a manipulation-cost-adjusted threshold, the learner raises the decision boundary by an amount proportional to the maximum effort or budget an agent can exert (Saba Ahmadi et al. in The Strategic Perceptron; Tianrun Zhao et al. in Online Strategic Classification With Noise and Partial Feedback).

This works because manipulation is rarely free; agents incur costs to alter their features (Jing Dong 0008 et al. in Combinatorial Bandits under Strategic Manipulations). By elevating the threshold or introducing an abstention buffer, the principal ensures that the economic utility of gaming becomes negative for unqualified agents (Lina Alkarmi et al. in When In Doubt, Abstain: The Impact of Abstention on Strategic Classification).

Only agents whose true underlying attributes are already high enough—or who are willing to invest in high-cost, permanent self-improvement—will find it viable to cross the elevated boundary (Keegan Harris et al. in Strategic Apple Tasting; Association for Artificial Intelligence 2026 et al. in Testing Under Strategic Manipulation: Mechanism Design for Human and AI Institutions). This effectively filters out superficial gaming, allowing the model to make updates based on reliable data.

The Paradox of Over Regulation

While raising the bar seems like an intuitive defense, it introduces a dangerous game-theoretic paradox. One might assume that making a test harder always reduces cheating, but setting overly strict thresholds can paradoxically increase system manipulation (Tian Xie et al. in Learning under Imitative Strategic Behavior with Unforeseeable Outcomes).

Agents constantly weigh the cost of honest skill-building against the cost of deceptive manipulation (Tian Xie et al. in Algorithmic Decision-Making under Agents with Persistent Improvement). Honest improvement typically takes more time and effort than superficial gaming. If a principal sets an excessively high decision threshold, the total effort required for genuine improvement may skyrocket much faster than the effort required to cheat. When the net utility of honest work plummets, agents abandon genuine adaptation entirely and switch to gaming the system.

To maximize utility, decision-makers must balance standards with incentives. When the population has a high capacity for genuine self-improvement, setting a more lenient threshold can actually reduce manipulation by making honest progress the more economically attractive path (Tian Xie et al. in Learning under Imitative Strategic Behavior with Unforeseeable Outcomes). Furthermore, blindly tightening thresholds to block gaming imposes a heavy social burden that often falls disproportionately on disadvantaged groups, undermining fairness (Smitha Milli et al. in The Social Cost of Strategic Classification).

Designing Incentive Compatible Evaluation Rules

Ultimately, the goal is to design evaluation rules where the agent's self-interest aligns perfectly with the principal's objectives.

To achieve this, Yatong Chen et al. in Learning to Incentivize Improvements from Strategic Agents propose optimizing for "Constructive Adaptation risk," a training objective that explicitly rewards beneficial self-improvement while penalizing harmful manipulation based on a clear feature taxonomy. Additionally, institutions can deploy mechanisms like uniform auditing or verification subsidies to verify whether an observed feature change reflects true recourse or deceptive misreporting (Andrew Estornell et al. in Incentivizing Recourse through Auditing in Strategic Classification).

Surprisingly, complex rules are not always necessary to achieve this alignment. As proven by Jon Kleinberg et al. in Designing Evaluation Rules That Are Robust to Strategic Behavior, if any reasonable evaluation rule can successfully incentivize productive effort over gaming, a simple linear combination of the proxy features will suffice. By carefully structuring the weights of these features relative to their manipulation costs, the system can ensure that the cheapest path to a positive classification is the one that actually generates real value.

Key Foundational Work

Strategic Classification is Causal Modeling in Disguise (2020) — contributed 3 concept elements to this direction.
The Strategic Perceptron (2020) — contributed 1 concept elements to this direction.
The Social Cost of Strategic Classification (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_gaming_improvement_distinction_2535c01b,
  title = {gaming-improvement distinction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/gaming-improvement-distinction/rdr_2535c01b99914cd296913b6bf46f93cb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_gaming_improvement_distinction_2535c01b_v1,
  title = {gaming-improvement distinction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/gaming-improvement-distinction/rdr_2535c01b99914cd296913b6bf46f93cb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake