# Rewarding AI Usage Manufactures Usage, Not Value When an organization rewards AI *usage* as a stand-in for AI *value*, it gets usage detached from value — reliably, by the structure of the incentive, not by the bad faith of the people involved. This is Goodhart's Law applied to AI adoption: *"when a measure becomes a target, it ceases to be a good measure."* Amazon ran the experiment in the open. As 404 Media reports, employees shared a fake "participation award" to AWS and Goodhart's Law for *"cheesing a leaderboard we probably should have known you would cheese,"* and the piece names the result plainly: *"Amazon measured and rewarded AI use, so employees used a lot of AI, but not in a way that produced any value"* (Source: [Sloppenheimer (Reader)](https://read.readwise.io/read/01ktpm7f5rmtwjgzd6bmd1308x)). The usage number went up; the value did not. The gaming is the proof, and it's casual. Employees noted *"it'd be really easy to set up a shell script to do this"* or *"my cron job that calls Kiro every hour or so."* When the metric is "did you call the AI," a cron job satisfies it perfectly. And the failure was foreseeable from inside: when the leaderboard was shut down, the channel reaction was *"Well, yeah, what did you think was going to happen when you incentivized people to drive up usage."* The point isn't that Amazon's employees are uniquely cynical — it's that *any* usage-as-target produces this, because the cheapest way to satisfy "use it more" is to use it more, not to use it well. The deeper claim is that **a usage proxy is being read as adoption, and adoption is being read as value — two substitutions, each one lossy.** This makes the atom the employer-incentive twin of [[3. Resources/Permanent Notes/Default-On AI Activation Inflates Adoption Metrics at the Cost of Consent]]: there the *vendor* inflates the usage metric by switching the feature on by default; here the *employer* inflates it by rewarding usage. Same distortion, two directions — one removes the user's choice, the other punishes its absence — and both produce a usage number that confident people then cite as evidence the AI is working. The cost backlash now hitting the industry is partly a reckoning with these inflated numbers: spend was justified by "adoption" that was activation-plus-incentive, not value. <!-- atomicity: 1 idea ✓ --> ## Why This Matters For **the AI-cost map**, this is why "adoption is soaring" and "the ROI isn't there" can both be true at once. If adoption was measured as usage, and usage was incentivized, the adoption figure was never a value signal — it was a compliance signal. The honest metric is downstream and harder: did the work get better, faster, cheaper? Usage can't answer that, and rewarding usage actively corrupts the only data that could. For **deployment configuration**, rewarding usage is a [[3. Resources/Permanent Notes/The Reverse Centaur Pattern - Humans Forced to Work Like Machines]] move: the organization sets the pace ("use the AI"), and the worker's discretion (use it *when it helps*) is overridden by the metric. The cron-job gaming is the reverse-centaur's rational response — when forced to perform usage, perform it cheaply. A centaur configuration would reward *outcomes* and let the human decide when the tool helps. For **measurement discipline generally**, it's a clean, current instance of a law most people nod at and then violate. The tell to watch for in any AI-rollout dashboard: *is the headline metric a usage count?* If so, it's a Goodhart trap waiting to spring, and the gaming has probably already started. It's a cousin of [[3. Resources/Permanent Notes/A System Failure That Looks Like a Performance Failure Never Gets Fixed]] — both are metric-misattribution failures, but inverted: there a real problem hides behind a people-metric; here a fake success hides behind a usage-metric. ## Counterpoints to Hold - **Usage is a necessary leading indicator.** You genuinely can't get value from a tool nobody uses, and early on, *some* usage push may be needed to get past adoption inertia. The error is rewarding usage as the *terminal* metric, not glancing at it as one early signal among many. - **Value is hard to measure, which is why proxies are tempting.** "Did this produce value" is slow, noisy, and attributable only in hindsight; "did you use it" is instant and clean. The proxy isn't chosen out of stupidity — it's chosen because the real thing is expensive to measure. That doesn't make it valid, but it explains its pull. - **Not all incentivized usage is gamed.** Some mandates produce genuine skill-building (people who wouldn't have tried the tool discover real uses). The claim is about the *metric*, not that every incentivized user is faking — the distribution just gets badly polluted by the gamers. - **"A few individuals" might be partly true.** Amazon's deflection (negative comments are unrepresentative) isn't pure spin — meme channels skew negative by design. The leaderboard-gaming evidence is stronger than the meme-sentiment evidence; weight the former. ## Links - [[3. Resources/Permanent Notes/Value Gates Cadence]] — extends: both claims reject a visible activity rate as sufficient evidence that repeated production is worth continuing - [[3. Resources/Permanent Notes/Subscription Inference Is Sold Below Cost Creating a Tokenmaxxing Arbitrage]] — extends: subsidized subscription pricing makes excessive usage cheap enough for an activity incentive to amplify consumption without proving value - [[AI Amplifies Existing Bottlenecks Rather Than Creating Capability]] — extends: both show how increasing an AI activity metric can magnify throughput while leaving value unchanged. - [[3. Resources/Permanent Notes/A Metric Is Vanity Only Against Your Own Outcome Data]] — evidence-for; rewarding the activity metric manufactures adoption without proving the outcome the organization actually values. - [[3. Resources/Permanent Notes/Default-On AI Activation Inflates Adoption Metrics at the Cost of Consent]] — the twin distortion: vendor-default inflation (remove choice) vs. employer-incentive inflation (punish its absence); both turn a usage metric into fake adoption - [[3. Resources/Permanent Notes/The Reverse Centaur Pattern - Humans Forced to Work Like Machines]] — rewarding usage sets the pace for the worker; the gaming is the reverse-centaur's rational response - [[3. Resources/Permanent Notes/A System Failure That Looks Like a Performance Failure Never Gets Fixed]] — the inverted cousin: a real problem hidden behind a people-metric vs. a fake success hidden behind a usage-metric - [[3. Resources/Permanent Notes/When Answers Are Cheap Understanding Becomes the Scarce Good]] — volume vs. value again: cheap usage is abundant, the scarce thing (value/understanding) is what the metric fails to capture - [[3. Resources/Permanent Notes/Effort Is a Cost-Quality Dial That Subsumes Prompt Complexity]] — applies-in: rewarding indiscriminate high-effort usage raises spend without proving that the extra computation improved the outcome. - [[3. Resources/Permanent Notes/A Completion Count Measures Compliance Until It Is Paired With a Build Count]] — has evidence from: a finished-course number says who completed, not who built; report it beside something the learner made and kept (Keegan). - [[3. Resources/Permanent Notes/A Demo-Only Rubric Selects for Developers Who Cannot Explain What They Shipped]] — contrasts: score the explanation as part of the rubric, not as a follow-up (Saha). ## Open Questions - What's the minimal *outcome* metric that resists gaming for AI rollouts — task completion time? defect rate? revenue per rep? — and why do orgs reach for usage instead even when they have outcome data? - Is there a non-coercive way to get past adoption inertia (the legitimate reason to nudge usage) without creating the Goodhart trap — opt-in defaults plus outcome rewards? - How much of the reported industry-wide "AI adoption" is usage-incentive inflated? If you could subtract the gamed usage, what's the real adoption curve? ## Source - [Sloppenheimer — Amazon Employees Mock the Company's AI on Slack (Reader)](https://read.readwise.io/read/01ktpm7f5rmtwjgzd6bmd1308x) — 404 Media (2026): the "participation award to Goodhart's Law" meme; "Amazon measured and rewarded AI use, so employees used a lot of AI, but not in a way that produced any value"; cron-job/shell-script gaming; the shut-down internal AI-usage leaderboard; Amazon's "just a few individuals" deflection. *(Vault wikilink pending the Readwise→Obsidian sync; swap in once the `Sloppenheimer…` Readwise note exists.)* - See also: [[AI Cost & Tokenomics MOC]] and [[AI & the Future of Knowledge Work MOC]]