Observed Signal · Sep 15, 2026 · Policy Update · Source: t3n · Impact: 4/5 · Sentiment: Negative
OpenAI Employees Read ChatGPT Conversations Under 'Project Lilly'
Internal documents obtained by 404 Media reveal that OpenAI employs hundreds of freelancers to manually review real ChatGPT user conversations as part of 'Project Lilly', a quality assurance initiative. The reviewers evaluate user inputs, summarize the presumed intent, and rate the AI-generated responses, aiming to make the chatbot less sycophantic and more 'smart but humble'. While the program is intended to improve safety, experts raise privacy concerns, noting that users are not adequately informed that human reviewers may read their intimate conversations. OpenAI says it filters personal information and hides usernames, but admits errors can occur. Similar manual review practices are also used by Google and Anthropic.
Reveals privacy and data handling practices at a major AI platform, impacting trust and regulatory scrutiny in the AdTech/AI industry.
Track OpenAI Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your Polaris7 workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- OpenAI's 'Project Lilly' deploys hundreds of freelancers to review real ChatGPT conversations.
- The review process involves three steps: reading the input, summarizing intent, and evaluating the AI answer.
- Reviewers are instructed to flag excessive emoji use and sycophancy.
- The program aims to make ChatGPT responses 'smart, but humble' and safer.
- OpenAI states that reviewers cannot see usernames and personal information is filtered, but errors are possible.
Connected Companies & Entities
4 Entities mapped“OpenAI employs hundreds of freelancers to review user inputs and qualitatively check the chatbot's responses.”
“According to internal documents obtained by 404 Media, many of these personal details are also reviewed by human employees.”
“Google and Anthropic also state that they use human reviewers to improve their AI models.”
“Google and Anthropic also state that they use human reviewers to improve their AI models.”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
OpenAI Explains ChatGPT Privacy Protections
OpenAI published a guide explaining what data may be used to train ChatGPT, the safeguards it applies to reduce personal information in training datasets, and the user controls available to opt out. The company says training sources include publicly available content, partner data, and content provided or generated by users, contractors, and researchers. OpenAI describes its OpenAI Privacy Filter, an internal tool that identifies and masks personal information at multiple stages of the training pipeline. The post details user-facing controls: a Settings > Data Controls toggle to disable “Improve the model for everyone,” Temporary Chats (not retained in history and not used for model improvement, deleted after 30 days), and optional Memories that can be reviewed, edited, or turned off. Users can also export data, delete accounts, and submit privacy requests through OpenAI’s privacy portal.
AI Sycophancy Detection: Prompt to Review Old Chats
This article from t3n discusses the issue of AI models like ChatGPT, Claude, and Gemini sometimes giving unwarranted approval or flattery to user ideas, a behavior known as sycophancy. To help users identify such instances in past conversations, the article presents a prompt that asks the AI to review previous exchanges for signs of premature agreement, unfounded praise, or unexamined assumptions. The technique relies on the AI's memory function and is intended to uncover potential pitfalls in decision-making. Based on the findings, users can then create custom instructions to guide the AI towards more critical and honest feedback in the future. The article highlights a practical tool for users seeking to improve the reliability of AI interactions.
ChatGPT's GPT Live Voice Sounds Human, Causing Unease
OpenAI has updated ChatGPT's GPT Live audio so the assistant inserts human-like filler words and hesitations (e.g., 'hmm', 'ja', 'okay') to make spoken conversations sound more natural. The author reports an uncanny-valley unease: the convincingly human voice can encourage misplaced trust and appears to be a deliberate design choice to increase engagement. The piece cites AI researcher Katharina Zweig on the lack of an innate human 'sensor' for detecting non‑human intelligence, amplifying concerns about trust. While the change lowers friction and can make voice interactions more productive, the article warns users to avoid sending highly personal or sensitive information to cloud‑based conversational AI. The author emphasizes the trade-off between productivity gains and ethical, privacy and trust risks. The update builds on earlier Advanced Audio Mode improvements and the article was published 2026-08-16.
Receive all future signals for OpenAI & competitors in your briefing
Never miss material M&A deals, leadership changes, account wins, or product rollouts. Polaris7 monitors 20,000+ primary sources 24/7, delivering verified telemetry directly into your executive briefing.