The document presents a technical perspective on Soylent, a Microsoft Word plug‑in that embeds paid Amazon Mechanical Turk workers directly into the authoring workflow to perform complex writing tasks such as text shortening, proofreading, and arbitrary user‑defined edits (the Human Macro). It argues that crowds can be treated as a programmable information processor, extending the classic view of human cognition as information processing. The core contribution is the Find‑Fix‑Verify (FFV) crowd programming pattern, which decomposes open‑ended editing tasks into three stages: workers first locate problematic text (Find), then propose concrete edits (Fix), and finally a separate set of workers validates the proposals (Verify). This pattern is applied to three main components: Shortn (interactive paragraph compression via a slider that selects the best combination of crowd‑generated rewrites), Crowdproof (human‑augmented spelling/grammar checking that surfaces errors Word misses and offers multiple fixes), and the Human Macro (a natural‑language interface allowing users to request arbitrary tasks such as citation formatting or image sourcing). The paper details the system’s architecture (a VSTO/WPF Word add‑in communicating with a back‑end that runs TurKit scripts on Mechanical Turk) and the algorithmic handling of slider‑based length constraints as a multiple‑choice knapsack problem. Empirical evaluations on a variety of text samples show that Shortn reduces paragraph length to 78‑90 % of the original with median work time under 2 minutes per paragraph but median wait times of ~18 minutes; Crowdproof catches 67 % of manually labeled errors (versus 30 % for Word) and fixes 88 % of those. The Human Macro achieves an 88 % intention success rate and 71 % accuracy. The discussion highlights challenges such as latency, cost, privacy, legal ownership of crowd‑generated text, and the need for interdisciplinary research on incentives and reputation. Ethical considerations about the social impact of large‑scale crowd work are raised. The conclusion outlines future work: richer crowd‑driven features, optimization of latency and cost, and extending the approach to other creative authoring domains.
Topics
Crowd‑Powered Interfaces: Bridging Human Cognition and Computation
The overarching vision that interactive systems can permanently embed paid crowd workers to perform complex, creative tasks that current AI cannot handle.
Find‑Fix‑Verify: A Three‑Stage Crowd Programming Pattern
A design pattern that splits open‑ended editing into (1) locating problematic text, (2) generating concrete edits, and (3) independent verification, using agreement and voting to ensure quality.
Shortn: Interactive Crowd‑Based Paragraph Compression
A UI component where users slide to specify desired length; the system selects a combination of crowd‑generated rewrites via a multiple‑choice knapsack algorithm, achieving 78‑90 % of original length.
Crowdproof: Human‑Augmented Proofreading
A workflow that finds spelling/grammar/style errors missed by Word, presents multiple crowd‑generated fixes, and verifies them, achieving 67 % error detection and 88 % correction of flagged errors.
Human Macro: Natural‑Language Crowd Scripting
A flexible interface allowing users to write free‑form task descriptions (e.g., citation gathering, image sourcing) that are sent to the crowd; evaluated for intention success (88 %) and accuracy (71 %).
Challenges and Limitations: Latency, Privacy, Legal Ownership
Discussion of practical barriers—worker recruitment latency, exposure of confidential text, work‑for‑hire legal status, and the trade‑off between cost and productivity.
Ethical and Societal Implications of Large‑Scale Crowd Work
Reflection on how crowd‑sourced interfaces can be a social good or reinforce inequality, calling for design processes that consider worker perspectives and avoid dehumanizing treatment.