Behind the Interface: How 'Project Lily' Employs Hundreds to Read Raw Prompts
An investigative report published by 404 Media has uncovered internal operations at OpenAI operating under the internal codename 'Project Lily.' According to the findings, the AI vendor employs a workforce of hundreds of human contractors dedicated to reading, evaluating, and grading actual ChatGPT prompt logs submitted by everyday users. Contractors are brought on through an intermediary vendor known as Crossing Hurdles and processed via the AI staffing platform Mercor, where compensation rates for North American participants exceed $50 per hour.
The core responsibility of these contractors involves analyzing end-to-end conversation logs and scoring the model's outputs across a structured 1-to-7 rating scale. Leaked internal guidelines establish that these evaluations focus heavily on curbing sycophantic behavior—where the model excessively validates user biases or errors—and strictly preventing the language model from anthropomorphizing itself as an entity possessing human consciousness or emotions. The disclosure demonstrates that human feedback refinement continues to rely on real-world prompt distributions.
Automated Redaction Gaps: Leaked Summaries and Identifiable Details
To mitigate regulatory blowback, OpenAI deploys an automated machine learning tool referred to as the 'Privacy Filter' designed to scrub standard direct identifiers, including account usernames and registered email addresses, before data packets reach contractor queues. However, OpenAI's own internal compliance documentation explicitly concedes that the automated filter is not foolproof and can fail when confronted with uncommon names, localized terminology, or nuanced narrative contexts.
Exacerbating the privacy risk, contractor interfaces display full conversation transcripts alongside an active 'user memories summary' panel. This module, populated by ChatGPT’s persistent memory features, exposes contextual breadcrumbs such as geographic locations, exact professional occupations, family relationships, and specific project timelines. In practical terms, these rich contextual markers frequently allow contractors to infer the real-world identity of the user, rendering basic automated redaction inadequate.
Default Opt-In Architectures and Comparative Industry Standards
A major structural issue spotlighted by the investigation is that conversational routing into Project Lily relies entirely on OpenAI's default-on telemetry settings. Standard consumer accounts automatically feed prompts into the pipeline unless the user manually toggles off the 'Improve the model for everyone' setting within their profile controls. Direct acknowledgment that human contractors review sessions was historically confined to general explanations within an FAQ support article maintained since 2023, rather than surfaced clearly at the interaction layer.
Technically, human-in-the-loop review operations remain a foundational element of enterprise Reinforcement Learning from Human Feedback (RLHF) and related safety post-training paradigms aimed at reducing behavioral degradation. Competitors across the frontier ecosystem implement comparable review workflows: Google conducts human evaluations on Gemini data distributions, while Anthropic maintains review pipelines that require explicit opt-in confirmation. The divergence in default consent states has intensified scrutiny around how transparently user prompts are ingested.
Community and Practitioner Reactions: Alarm Versus Training Necessity
Reaction across practitioner circles and software engineers manifested as immediate alarm and deep frustration. Many professionals who had adopted ChatGPT as a private thinking environment, drafting desk, or informal debugging assistant expressed disbelief at the operational reality of human third parties parsing their logs. Step-by-step guides detailing how to navigate the account preferences menu to disable data sharing circulated rapidly across developer communities.
Conversely, machine learning practitioners and safety researchers emphasized that data curation using real conversational distributions represents a technical necessity. From their perspective, purely synthetic data and automated evaluators fail to capture edge cases, making human oversight critical for addressing model degradation and sycophancy. Nevertheless, technical observers largely agreed that the friction stems from user experience defaults: failing to communicate human inspection clearly at the input box undermines trust regardless of the architectural rationale.
Strategic Takeaways for Thai Enterprises Under PDPA Compliance
For enterprise leadership and compliance officers in Thailand, the revelations surrounding Project Lily pose immediate regulatory challenges under the Personal Data Protection Act (PDPA). When employees input sensitive operational files, localized customer records, or internal source code into consumer-grade ChatGPT accounts without modifying account defaults, those data points risk passing through incomplete automated filters directly onto overseas contractor monitors. This introduces immediate liability regarding unauthorized cross-border personal data transfers under Thai statutory standards.
Chief Information Officers (CIOs) and Data Protection Officers (DPOs) across Thailand must reinforce operational AI governance frameworks. Organizations should mandate the use of zero-data-retention enterprise agreements or standard API tiers, where OpenAI contractually guarantees prompts are excluded from model training pipelines. If enterprise accounts are unavailable, companies must enforce mandatory local opt-out configurations across all corporate browsers and implement strict data loss prevention (DLP) policies barring unredacted proprietary data from conversational interfaces.
For Thai enterprises and individual professionals routinely inputting operational data into ChatGPT, default-on data contribution settings expose trade secrets and PDPA-regulated personal identifiers to external human reviewers.