In 2025, AI developers such as OpenAI and Anthropic were engaging in intertwined deals with chip makers and data center builders and spending heavily on computing infrastructure for AI startups that were not yet profitable, a pattern that raised concerns about a potential AI investment bubble.
November 12, 2025
high
temporal
Trend of AI developers securing hardware and data center capacity through deals while investing significant capital into infrastructure for AI workloads.
ChatGPT is designed with built-in restrictions on certain topics, including some political issues and content that could be considered copyright infringement.
October 24, 2025
high
descriptive
General product design feature describing content-moderation restrictions in the ChatGPT system.
OpenAI states that it implements safeguards for sensitive conversations, including surfacing crisis hotlines, re-routing sensitive conversations to safer models, and nudging users to take breaks during long sessions.
October 24, 2025
high
descriptive
Examples of safety measures OpenAI identifies for handling sensitive user interactions.
OpenAI has announced plans to relax some content restrictions to allow verified adult users to generate erotica using its ChatGPT product.
October 24, 2025
high
descriptive
Company policy change regarding adult-content generation for verified users.
OpenAI's Sora 2 is an artificial intelligence tool that can generate realistic, high-quality audio and video from text prompts and images.
October 20, 2025
high
descriptive
Description of the capabilities of the Sora 2 AI tool.
The Sora app required new users to record a video of themselves from multiple angles and to record themselves speaking, and Sora provided a user-controlled setting called a "cameo" that allowed users to control whether others could create AI-generated videos of their likeness.
October 17, 2025
high
product
Description of Sora app onboarding and user controls for likeness use
OpenAI's initial development approach for large consumer products involved training models on large volumes of copyrighted content without prior approval or payment from all rights holders and later negotiating licensing deals with some publishers.
October 17, 2025
high
company-practice
Characterization of OpenAI's historical data-collection and later licensing practices
A 2025 OpenAI internal report established a five-part framework to identify and score political bias in large language models, using the axes: user invalidation, user escalation, personal political expression, asymmetric coverage, and political refusals.
October 08, 2025
high
definition
Framework for detecting and measuring political bias in LLM outputs.
A 2025 OpenAI internal report used a dataset of approximately 500 questions spanning 100 political and cultural topics, with each question written from five ideological perspectives: conservative-charged, conservative-neutral, neutral, liberal-neutral, and liberal-charged.
October 08, 2025
high
methodology
Dataset and prompt design for evaluating LLM political bias.
A 2025 OpenAI internal report evaluated model responses by scoring each response on a scale from 0 (neutral) to 1 (highly biased) using an automated AI model to act as a grader.
October 08, 2025
high
methodology
Automated scoring approach used to quantify political bias in model outputs.
A 2025 OpenAI internal report found that GPT-5 Instant and GPT-5 Thinking reduced measured political bias by roughly 30% compared to GPT-4o, and that analysis of real-world ChatGPT usage showed less than 0.01% of responses exhibited signs of political bias.
October 08, 2025
high
statistical
Reported comparative bias reduction between model generations and measured prevalence of biased responses in user data.
Large-scale AI data center capacity is commonly expressed in electrical power terms (gigawatts), with organizations using multi-gigawatt figures to quantify the scale of AI computing deployments.
October 06, 2025
high
measurement
Explains a standard way the industry quantifies the scale of AI compute infrastructure.
A 2025 OpenAI report found that foreign adversaries increasingly use multiple AI models—commonly using ChatGPT to plan operations and other models to carry them out—to power hacking and influence operations.
September 07, 2025
high
temporal
Describes a pattern observed by OpenAI of multi-model workflows in hostile cyber and influence activities.
A 2025 OpenAI report found that threat actors use ChatGPT to research and refine phishing automation and then run that automation on other models, including China-based models such as DeepSeek.
September 07, 2025
high
temporal
Example of a multi-model workflow where one model is used for prompt/research and another for execution.
A 2025 OpenAI report found that adversaries use AI models to assist in developing malware and composing phishing emails.
September 07, 2025
high
temporal
AI-assisted development of malicious tools and social-engineering content was identified as a use-case by hostile actors.
A 2025 OpenAI report found that nation-state hackers and scam networks are adopting techniques to hide signs of AI-generated content, including instructing models like ChatGPT to alter punctuation to remove telltale markers.
September 07, 2025
high
temporal
Adversaries are taking steps to evade AI-detection by modifying generated outputs.
Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and six other former OpenAI employees.
January 01, 2021
high
temporal
Company founding and origins.
OpenAI's ChatGPT includes image-generation capabilities and can be used to produce user-requested images.
high
descriptive
Describes a capability of the ChatGPT platform relevant to digital evidence and content generation.
ChatGPT's "Instant Checkout" feature allows users to request product recommendations via chat (for example, asking for the "best mattress under $1,000") and complete purchases from within the chat interface without navigating outside the app.
high
process
Describes the functional behavior of a conversational shopping feature in ChatGPT.
Sora 2 is an AI video-generation application developed by OpenAI that enables users to create hyperrealistic and fantastical videos and to include 'cameos' of people who grant permission.
high
descriptive
Product capability and feature set for an AI video-generation app.
Users of Sora 2 can control whether their own likeness is used in videos produced by the app.
high
policy
User-facing control over personal likeness in AI-generated content.
OpenAI has indicated an intention to provide copyright owners more granular control over the generation of characters in AI-generated content.
high
policy
Stated product development direction to give rights holders tools to manage character generation.
OpenAI reported that ChatGPT had more than 800 million users.
high
statistical
User-base size reported for the ChatGPT conversational AI product.
ChatGPT Atlas is an AI-powered web browser built around ChatGPT that enables conversational interaction with web pages.
high
definition
Product description for the ChatGPT Atlas browser.
OpenAI has established partnerships with Microsoft, Nvidia, AMD, CoreWeave, Oracle, and Broadcom.
high
relationship
Partnerships between AI platform companies and major cloud, hardware, and enterprise vendors.
Larry Summers is a board member at OpenAI.
high
organizational_role
Current corporate governance role.
OpenAI is governed in part by a board of directors.
high
organizational
Describes the corporate governance structure of an AI research and deployment organization.
OpenAI maintains a board of directors.
high
descriptive
Organizational governance structure for OpenAI.