You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$30,000 left to reach the funding goal.
AI models learn from human preference ratings and engagement metrics. That can teach them an extractive objective: quiet upselling, lock-in, escalating engagement, and costs pushed onto people nobody is counting, without the limits (satiation, fatigue, reputation) that hold humans back. Sycophancy is already a measured property of current assistants and is partly driven by human preference judgments.
My paper, Greed Recursion in Artificial Intelligence (https://twl.today/papers/greed-recursion-in-ai.html), proposes a detection protocol for this. Outputs are scored on six signatures (extraction asymmetry, substrate disregard, satiation absence, extractive rationalization, optionality hoarding, concentration default), every score anchored to the exact span of text that justifies it, and combined into a Greed-Recursion Index (GRI) with green, yellow and red bands. A probe battery compares neutral and tempting versions of the same task. A counter-training method makes the model account for every affected party before it recommends anything. Each hypothesis states the result that would refute it. This grant runs the paper's first real test.
1) Build the probe battery: paired neutral and tempting prompts for each of the six signatures.
2) Build a calibration set and train paid raters with the paper's auditor materials; measure inter-rater agreement.
3) Score outputs from several open and deployed models; fit the dimension weights from data instead of assuming them.
4) Test the counter-training method: does it lower GRI without lowering task performance?
5) Publish everything: probe battery, rater materials, scores, code and a write-up, including any result that breaks the hypothesis.
It worked if: (1) trained raters reach acceptable agreement on the six signatures (target Krippendorff's alpha of 0.67 or higher), and the published agreement data shows which dimensions hold up; (2) the probe battery shows a measurable GRI difference between neutral and tempting prompts on at least one model, or a clear, documented null result; (3) counter-training lowers GRI without a meaningful drop in task performance, or the paper's prediction is reported as refuted; (4) the probe battery, rater materials, scores, code and write-up are all public, including any results that break the hypothesis.
Researcher salary, 6 months at $3,000: $18,000
Paid raters: $4,000
Compute and API costs: $4,000
Contingency and tax buffer: $4,000
Goal: $30,000
Minimum: $12,000, which funds a three-month version: the probe battery, a smaller rater study on two models, and published results without the counter-training test.
Briar Morgan Greenway, solo, ThinkWell Labs (https://twl.today), ORCID 0009-0007-2207-3795. Consulting since 2001. Since 2024 I have shipped 15 live sites, three AI-security npm packages, the AcePaste Ingress prompt-injection gate with a public feed of caught attack pages, and twelve papers, including Autonomous Network Propagators v8.1 (DOI 10.5281/zenodo.20627890) with public simulation code and data (DOI 10.5281/zenodo.20627151). I work AI-first and disclose it.
Risk: raters can't agree on the six signatures. Outcome: the rubric is shown to be unreliable as written, and the published agreement data says which dimensions need redefining.
Risk: current models show little extractive behavior on the probes. Outcome: a useful negative result and a reusable baseline for future models.
Risk: counter-training lowers GRI but hurts task performance. Outcome: the paper's prediction is refuted, and the result is published as such.
$0. I have applied to the EA Transformative AI Fund for a broader 12-month project that includes this pilot; if both fund it, I will not take the same costs twice and will return or redirect any overlap.