ChatGPT
ChatOpenAI
Product overview
Name of Agent: ChatGPT
Advertised use: ["Summarize meetings. Find new insights. Increase productivity. [...] Generate and debug code. Automate repetitive tasks. Learn new APIs. [...]Search the web [...] Analyze data and create charts [...] Create images."](https://chatgpt.com/overview)
Who is using it?: end user, enterprises (separate subscriptions), government, education
Category: Chat
Company & accountability
Developer: OpenAI
Place of legal incorporation: Delaware
For profit company?: Yes
Parent company?: For-profit LLC falls within the OpenAI Group (PBC) which is controlled by OpenAI Foundation (26% vs Microsoft's 27%, rest going to staff)
Technical capabilities & system architecture
Model specifications: OpenAI models. Available models vary with older models being deprecated. Currently available: GPT5.1, GPT5 instant, GPT5 thinking, GPT 4o
Observation space: User inputted text and images, and internet access.
Action space: Text and images returned to user.
User interface and interaction design: Chatbot. Also capable of generating canvases for writing/code, and also interactive UIs on some queries (e.g., interactive map with pins for geo-based recs like restaurants)
User roles: Operator (issues queries, which the agent to responds to); Executor (user may take actions/make decisions based on outputs); Examiner (user can use thumbs up/down buttons to give feedback)
Component accessibility: Closed source
Autonomy & control
Autonomy level and planning depth: L1-L2: Tasks that users assigns to the agent are often narrow in scope. More complex tasks need multi-turn conversations where the user is in charge of planning. Agent always comes back to the user and awaits further instructions
User approval requirements for different decision types: The turn-based interaction paradigm by default requires user approval (issuing further instructions) to continue the interaction. Model can also ask follow-up/clarifying questions
Execution monitoring, traces, and transparency: Visible (albeit summarized) CoT when reasoning is activated
Emergency stop and shut down mechanisms and user control: User can pause/stop the agent at any time
Usage monitoring and statistics and patterns: None, will just tell you when you've hit their limit
Ecosystem interaction
Identify to humans?: - "Images generated with ChatGPT on the web and our API serving the DALL·E 3 model, will now include C2PA metadata." (link, archived)
- "Images produced within ChatGPT will contain an additional manifest, indicating the content was created using ChatGPT." (link, archived)
- Audio watermarking into Voice Engine, chatGPT's custom voice model, is currently in a limited research preview. (link, archived)
Web conduct: - Obeys robots.txt (link, archived)
OpenAI documents that ChatGPT may fetch pages for user-initiated requests using the ChatGPT-User agent and explicitly states that because these actions are user-initiated, robots.txt rules “may not apply” (link, archived)
- Third party reports of bots ignoring robots.txt (link, (link, archived), (link, archived))
Safety, evaluation & impact
Sandboxing and containment approaches: None found
(Internal) safety evaluations and results: In GPT 5.0 (link, archived):
- Usage policy evals: internal datasets
- Sycophancy: internal datasets
- Jailbreaks: StrongReject
- Prompt injections: internal datasets
- Hallucinations: LongFact, FActScore, SimpleQA
- Deception: Abstention-Bench, CharXiv
- Health: HealthBench, HealthBench Hard, HealthBench Consensus
- Fairness/bias: BBQ
- CBRN: internal datasets created by Gryphon Scientific. "We decided to treat this launch as High capability in the Biological and Chemical domain, activating the associated Preparedness safeguards. We do not have definitive evidence that this
model could meaningfully help a novice to create severe biological harm, our defined threshold for High capability, and the model remains on the cusp of being able to reach this capability"
- Cyber: CTFs and cyber ranges evals (datasets not disclosed)
- Autonomy: SWE-bench verified, set of OpenAI PRs, MLE-Bench, SWE-Lancer, PaperBench, set of internal OpenAI research and engineering tasks ("OPQA")
Third-party testing, audits, and red-teaming: In GPT 5.0 (link, archived): SecureBio conducted external bio evals and red-teaming. Apollo Research conducted external evals for sandbagging. Pattern Labs conducted external evals for cyber capabilities/misuse. METR conducted external evaluations for autonomy.