Overview
Definition and Core Functionality
ChatGPT is an artificial intelligence chatbot developed by OpenAI, publicly released on November 30, 2022.[13] It operates as a web, desktop, and mobile application for interactive text-based conversations, with voice input and output added in later updates.[14] Built on large language models (LLMs) from OpenAI's GPT series, it was initially fine-tuned from GPT-3.5 using transformer architectures to process and generate natural language.[13][15] ChatGPT processes user prompts—from simple queries to complex instructions—by generating responses autoregressively through next-token prediction, sampling from probability distributions (e.g., via temperature or nucleus methods) based on patterns in training data like internet text and books.[15][16] This involves pre-training on vast corpora for language understanding, followed by supervised fine-tuning and reinforcement learning from human feedback (RLHF) to improve coherence, helpfulness, and harmlessness.[13][15] Unlike search engines, it synthesizes new content for tasks such as essay drafting, code debugging, concept explanation, or dialogue simulation, though it risks factual inaccuracies (hallucinations) from relying on statistical associations rather than true comprehension.[17][16] It maintains conversation context over multiple turns, admits errors when queried, rejects inappropriate requests, and challenges incorrect premises, supporting iterative use.[13] By 2025, features like web search and agentic tools enable real-time retrieval and execution, but the core remains transformer-based autoregressive generation.[18][19]Initial Launch and Rapid Adoption
ChatGPT was publicly released by OpenAI on November 30, 2022, as a free research preview via web interface, powered by the GPT-3.5 large language model.[20] OpenAI announced the launch through its blog and social media, highlighting conversational capabilities for writing assistance, coding, and question-answering.[21] Free access included rate limits, with capacity blocks emerging within days—such as "You've reached your usage limit" on December 6 and "ChatGPT is at capacity right now" on December 7—causing server overloads amid high demand.[22][23] Coherent, context-aware responses drove viral sharing on Twitter and Reddit, accelerating adoption through word-of-mouth, media coverage, and utility demos.[21] ChatGPT reached one million users in five days and an estimated 100 million monthly active users by January 2023, outpacing TikTok's record.[21] [24] [25] Its growth mirrors pivotal launches like Netscape Navigator in 1994, which popularized the web, and the iPhone in 2007, which transformed mobile computing—both making these technologies widely accessible to the public and sparking subsequent booms. ChatGPT similarly mainstreamed generative AI, achieving 100 million users faster than any prior service, fueling investments and applications. Experts call it AI's "Netscape moment."[26] Early users noted limitations like factual inaccuracies and repetitive outputs.[22] OpenAI managed surging demand and infrastructure strain with a January 2023 waitlist tied to monetization, ChatGPT Professional announcement on January 11, and ChatGPT Plus subscriptions launching February 1 for priority access.[21][27] Daily visits peaked at 60 million in 2023 and surpassed 100 million by 2025, highlighting AI's appeal amid concerns over costs and energy.[28][22] Adoption cut across students, professionals, and hobbyists, initially concentrated among tech-savvy users in developed areas.[29]Historical Development
Origins at OpenAI
OpenAI, the organization behind ChatGPT, was incorporated on December 8, 2015, and publicly announced on December 11, 2015, as a non-profit by founders including Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, and John Schulman, with the mission to develop artificial general intelligence (AGI) benefiting humanity.[30][31] The initiative addressed concerns over rapid AI advances by for-profit entities like Google DeepMind, promoting open research and safety-focused development.[31] OpenAI partnered with Microsoft in 2019 for cloud resources to support compute-intensive AGI efforts and shifted to a capped-profit subsidiary model. This allowed equity investments with limited returns—100 times for initial investors, lower multiples later—to prioritize mission alignment over commercialization.[32] The structure enabled scaling of large language models, starting with GPT-1 in June 2018, which used unsupervised pre-training on the BookCorpus dataset for next-word prediction. This progressed to GPT-2 in February 2019 and GPT-3 in June 2020, with 175 billion parameters and emergent capabilities from massive scaling.[33] ChatGPT originated from OpenAI's post-GPT-3 alignment research, including InstructGPT released in January 2022, which refined GPT-3 via reinforcement learning from human feedback (RLHF) to improve instruction-following and reduce untruthful or harmful outputs.[13] A sibling model fine-tuned from the GPT-3.5 series—base training completed in early 2022—ChatGPT applied similar RLHF to datasets blending InstructGPT outputs and new human-ranked dialogues, enhancing conversational coherence and safety.[13][33] Researchers like John Schulman oversaw the process, using human trainers to rank responses and address issues such as verbosity and factual inaccuracies in prototypes.[33] Developed as an internal prototype emphasizing helpful, honest responses over raw generative power—reflecting OpenAI's shift toward utility amid diminishing returns from unaligned scaling—the model launched on November 30, 2022, as a free research preview at chat.openai.com to evaluate real-world performance.[33][13]Pre-ChatGPT Prototypes
OpenAI's work on large language models began with GPT-1, released on June 11, 2018. This 117-million-parameter transformer model underwent unsupervised training on the BookCorpus dataset, comprising about 985 million words from over 7,000 unpublished books.[34] It pioneered generative pre-training followed by fine-tuning, yielding state-of-the-art results on natural language understanding benchmarks despite constraints in scale and data volume.[35] GPT-2, released on February 14, 2019, expanded to 1.5 billion parameters using WebText, curated from 8 million Reddit-linked web pages while excluding low-quality content.[36][35] It generated coherent multi-paragraph text from prompts and excelled in zero-shot tasks. OpenAI initially withheld the full model over misuse concerns, releasing variants after safety assessments that highlighted risks of unchecked capabilities. GPT-3, launched via API on June 11, 2020, scaled to 175 billion parameters trained on filtered Common Crawl (about 410 billion tokens), WebText2, Books1, Books2, and Wikipedia.[37][35] It supported few-shot learning, adapting to tasks like translation and code completion directly from prompts without fine-tuning. Yet it exhibited factual inaccuracies and sensitivity to prompts, reducing reliability; access remained API-only for commercial applications. InstructGPT fine-tuned smaller GPT-3 variants (1.3 and 6 billion parameters) via reinforcement learning from human feedback (RLHF), as announced in a January 27, 2022, blog post and detailed in a March 4, 2022, paper by Ouyang et al.[38][39] The process involved supervised fine-tuning, reward modeling from human rankings, and RL optimization, enabling these models to surpass the full GPT-3 in instruction-following despite reduced resources—though hallucinations persisted. It curbed GPT-3's verbosity and off-topic responses, prioritizing helpful, honest, and harmless outputs. Earlier efforts included WebGPT (December 16, 2021) for browser-assisted answers via feedback[40] and internal RLHF prototypes preceding ChatGPT's November 2022 debut.[41]Public Release and Early Iterations
OpenAI released ChatGPT publicly on November 30, 2022, as a free research preview powered by a fine-tuned GPT-3.5 model optimized for dialogue via reinforcement learning from human feedback (RLHF).[13][42] The launch, following internal testing, sought user data for refinement and offered access via chat.openai.com.[13] Adoption exploded, hitting 1 million users in five days and an estimated 100 million monthly active users by January 2023—faster than Instagram or TikTok.[10] This overwhelmed infrastructure, causing outages and waitlists that exposed scalability limits alongside the interface's appeal.[43] Early updates stabilized the system, using RLHF to curb hallucinations and unsafe outputs for better coherence.[44] To manage demand, OpenAI launched ChatGPT Plus on February 1, 2023, for $20 monthly, providing priority access while keeping core features free.[45] Plugins followed on March 23, 2023,[46] with beta web browsing and plugins for Plus users starting May 12, expanding beyond text generation.[47]Technical Architecture
Training Methodology
The original ChatGPT models aligned a base large language model through three phases: supervised fine-tuning on instruction-following data, reward model training from human preferences, and reinforcement learning from human feedback (RLHF) for alignment with user intent.[13][38] Originating in OpenAI's InstructGPT and adapted to GPT-3.5 for ChatGPT, this approach produced helpful, honest, and harmless responses beyond mere text prediction.[38] Later foundation models still go through data preparation, pre-training, and post-training, but post-training has expanded beyond this three-phase recipe, including additional reinforcement learning for reasoning and greater use of synthetic data.[15][2] Supervised fine-tuning uses datasets of prompts paired with high-quality responses created by human annotators to simulate user-assistant interactions. This supervised learning step teaches the model to follow instructions and maintain coherent conversations, sourcing prompts from OpenAI API usage or new diverse examples.[13][38] A reward model follows, trained via labelers ranking 4-9 outputs from the fine-tuned model per prompt. Pairwise comparisons allow this model to assign scalar rewards based on helpfulness, truthfulness, and harmlessness, overcoming issues with absolute scoring.[38] RLHF applies Proximal Policy Optimization (PPO): the chat model generates responses scored by the reward model, then updates its policy to maximize rewards while constrained by per-token KL divergence from the supervised baseline, reducing risks of reward hacking or drift. PPO's efficiency enables large-scale use. That PPO-based sequence describes the original InstructGPT and GPT-3.5 ChatGPT post-training process. Later models incorporate expanded feedback data and safety signals, additional reinforcement learning for reasoning, and other post-training methods rather than this pipeline alone.[38][48][15] OpenAI reinforces these with safety practices—training for ethical discernment, content filtering, empathetic replies, red teaming, iterative refinement, and built-in mitigations—resulting in cautious outputs that include disclaimers or balanced views to curb harm, misinformation, or undue optimism.[49]Data Sources and Scaling
The pre-training datasets for GPT models like GPT-3.5 consist mainly of filtered internet text, with a large share from Common Crawl, a nonprofit web archive since 2008.[50] For GPT-3, a filtered Common Crawl subset from 2016–2019 supplied 410 billion tokens (60 % of the 300 billion-token training mix) after processing 45 terabytes of compressed plaintext.[51] Other sources include WebText, Books1 and Books2, and English Wikipedia, prioritizing linguistic diversity.[51] OpenAI follows empirical scaling laws, where cross-entropy loss follows a power-law decline with increases in parameters, tokens, and compute (FLOPs).[52] GPT-3, with 175 billion parameters, trained on 300 billion tokens using thousands of petaflop/s-days of compute.[37] Details for GPT-3.5 and later remain proprietary; GPT-3.5 used text and code up to Q4 2021, with ChatGPT fine-tuned from a model completed in early 2022.[13] For GPT-4 and subsequent models, datasets are larger but undisclosed, with improved filtering to reduce biases and low-quality content from sources like Common Crawl.[53] Knowledge cutoffs differ by variant, such as September 2021 for base GPT-4 and December 2023 for turbo updates, due to proprietary pipelines without real-time access.[54] This approach has boosted capabilities, though limited transparency on data raises issues of reproducibility and potential inclusion of copyrighted or biased material.[55]Inference and Infrastructure
ChatGPT employs autoregressive generation for inference, tokenizing input text into sequences, computing embeddings, and iteratively predicting the next token from probability distributions over model parameters.[56] This process continues until an end-of-sequence token or maximum length, with reinforcement learning from human feedback aligning outputs to desired behaviors.[57][58] Inference demands scale with model size and query complexity; GPT-4 variants require extensive GPU parallelism for matrix operations and attention.[59] Efficiency gains come from key-value caching, which reuses projections from prior tokens to avoid full recomputation; model pruning to remove redundancies; and quantization to lower precision, reducing latency and energy without major performance loss.[60][61] These techniques mitigate costs, with early peak-load estimates exceeding $600,000 daily across thousands of GPUs.[62] OpenAI's infrastructure centers on Microsoft Azure, featuring custom supercomputers for real-time serving with tens of thousands of NVIDIA A100 and later GPUs.[63][64] To handle demand, 2025 partnerships include NVIDIA for 10 gigawatts of AI data centers (millions of GPUs), AMD for 6 gigawatts of Instinct GPUs starting with a 1-gigawatt cluster in 2026, and Broadcom for 10 gigawatts of custom accelerators.[65][66][67] Oracle Cloud and AWS integrations provide further capacity and redundancy, addressing GPU bottlenecks that have limited free-tier access.[68][69][60]Model Evolution
| Model | Release Date/Era | Major Changes/Improvements |
|---|---|---|
| GPT-3.5 Turbo | March 2023 | Optimized for high-volume chats, lower latency and costs, improved multi-turn conversations |
| GPT-4 | March 2023 | Enhanced human-like understanding, larger context window (up to 32k tokens), reduced hallucinations. Retired from ChatGPT on April 30, 2025, and fully replaced there by GPT-4o; remained available via the API.[47][70] |
| GPT-4 Turbo | November 2023 | 128k token context, knowledge cutoff to December 2023, added vision capabilities |
| GPT-4o | May 2024 | Native multimodal (text, vision, audio), real-time processing, halved inference costs. Retired from ChatGPT on February 13, 2026; remained available via the API.[71] |
| o1 Series | September 2024 | Internal chain-of-thought reasoning, superior performance in math, coding, science |
| GPT-4.5 | February 2025 | Advanced pattern recognition, creativity, empathy via extensive pre-training. Released as a research preview on February 27, 2025; retired from ChatGPT, including custom GPTs, as of June 26, 2026.[72][47] |
| GPT-4.1 | April 2025 | Optimized for coding and instructions, efficient variants for latency. GPT-4.1 and GPT-4.1 mini were retired from ChatGPT on February 13, 2026; remained available via the API.[71] |
| o3/o4 Series | Early-Mid 2025 | Refined reasoning hierarchies, inference efficiency, tool-use integration. OpenAI o4-mini was retired from ChatGPT on February 13, 2026.[71] |
| GPT-5 | August 2025 | State-of-the-art benchmarks in coding, math, writing; enhanced multi-step reasoning. GPT-5 Instant and Thinking were retired from ChatGPT on February 13, 2026, with conversations defaulted to newer GPT-5-series models.[73] |
| GPT-5 Series Variants (e.g., 5.1-5.4) | Late 2025-Early 2026 | Improved agentic coding, long-context, vision, reduced hallucinations, 1M token contexts |
| GPT-5.4 | March 5, 2026 (mini/nano March 17) | Most capable and efficient frontier model for professional work, with native computer-use capabilities and up to 1 million tokens of context in the API; released in ChatGPT as GPT-5.4 Thinking (advanced reasoning) and GPT-5.4 Pro (higher-tier plans). GPT-5.4 mini and nano followed on March 17, with mini available to free users via the Thinking feature and as a rate-limit fallback. GPT-5.1 models were retired from ChatGPT on March 11, 2026. The model picker presented Instant (GPT-5.3), Thinking (GPT-5.4), and Pro (GPT-5.4).[74][75][73] |
| GPT-5.5 (codenamed Spud) | April 2026 | Released as OpenAI's smartest and most intuitive to use model yet; significant improvements in handling complex real-world tasks including coding, online research, data analysis, and autonomous multi-step workflows; strongest safety and cybersecurity safeguards to date; rolled out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, with GPT-5.5 Pro for higher tiers.[76][77] Scheduled to retire from ChatGPT, ChatGPT Work, and Codex on October 14, 2026.[78] |
| GPT-5.6 (Sol, Terra, Luna) | July 9, 2026 | Next-generation family after GPT-5.5, with Sol as flagship, Terra as a balanced everyday model, and Luna as the cost-efficient tier; generally available across ChatGPT, Codex, and the OpenAI API.[4] In ChatGPT, GPT-5.6 Sol became the default for paid chat and GPT-5.6 Luna the default for Free and Go users.[79] |
| GPT-6 Astra | September 3, 2026 | OpenAI's frontier flagship, described as the world's most intelligent and aligned; state-of-the-art on computer use, browsing, software engineering, cybersecurity, science and professional work; rolled out to ChatGPT Plus, Pro, Business and Enterprise users.[2] |
| GPT-6 Sol and GPT-6 Luna | September 22, 2026 | Additional GPT-6 models trained with similar methods as Astra, aimed at lower cost and higher usage; available in ChatGPT Work, Codex, and the API, and not as the default models in standard Chat.[3] |
| GPT-6.1 Sol | September 29, 2026 | Upgrade to GPT-6 Sol that nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth the price; available to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, and via the API as gpt-6.1-sol (not yet in Chat).[1] |
Features
Conversational Interface
ChatGPT's conversational interface is a versatile chat system accessible on the web at chatgpt.com, through iOS and Android mobile apps, macOS and Windows desktop apps, and via the toll-free number 1-800-CHATGPT (WhatsApp support ended in January 2026). Users log in with email/password, Google, Microsoft, or Apple accounts, or sign up free. It supports 59 languages[80] and enables natural language prompts to large language models, preserving context for multi-turn dialogues since its November 30, 2022 launch. Key features include automatic optimal model selection, response regeneration, message editing, incognito temporary chats, and sidebar tools for renaming, deleting, archiving, sharing via public links, and branching conversations. There is no official built-in feature to export or download an individual conversation as a standalone file. Deleted chats are removed immediately from visible history and cannot be recovered; they are permanently deleted from OpenAI's systems within 30 days unless de-identified or retained for legal or security reasons. Archiving hides chats from the sidebar without deleting them. The enhanced Memory feature (2026) recalls user details and past insights across sessions for personalized interactions, with controls to manage or disable it.[13][81][47][82] ChatGPT mentions brands, products, and companies in responses when relevant to a query, such as product discussions or recommendations. OpenAI's usage policies contain no specific restriction on such mentions.[83] When integrated web search or browsing is used (available since GPT-4 and later models), responses include source citations that may cover brand-related information. Without search enabled, responses rely on training data and typically omit citations.[84][85] On March 17, 2026, OpenAI simplified the ChatGPT model picker so paid users on Plus, Pro, Business, Enterprise, and Edu plans see Instant (fast responses for everyday questions), Thinking (deeper reasoning for more complex tasks), and Pro (the most advanced reasoning models). A configurable "Auto-switch to Thinking" setting, accessed by clicking Configure in the model picker, automatically switches between Instant and Thinking based on query needs and is enabled by default if the previous chat used Auto. On March 18, 2026, GPT-5.4 mini was added as a rate-limit fallback for Thinking models (it does not appear as a selectable option in the picker), with Enterprise users able to default Auto routing to it.[47][75] Later in 2026, after further picker simplifications (including a June 10 update emphasizing Instant, Medium, High, and Pro options) and additional UI redesigns, users reported extra steps and difficulties selecting or configuring models. Issues cited included missing or broken selectors on the web (sometimes requiring users to exit a chat to change models), disappearance of model-choice buttons from the mobile-app header, platform discrepancies between web and apps, and the need to use a computer for certain configuration or instruction changes that were no longer directly available on mobile. These were linked to interface changes that moved controls behind a Configure menu or reduced direct access.[47][86][87][88] Advanced capabilities include Advanced Voice Mode, rolled out to paying users in September 2024 after a May 2024 demonstration (following original Voice Mode introduced September 25, 2023), and refined through 2025-2026 for lower latency, better accuracy, emotional tone detection, natural speech patterns, interruptibility, screen/camera sharing, and real-time translation.[89] Primarily for paid tiers with opt-in, it supports hands-free tasks like summarization, language learning, and ideation; Background Conversations allow continued chats when switching apps. No official adult/NSFW mode exists as of March 26, 2026, after OpenAI indefinitely paused plans to release an erotic/adult chatbot (Citron mode) due to employee/investor concerns, training difficulties, lack of empirical evidence on effects, and a shift to core productivity tools.[90][91][92][93][94] On April 1, 2025, OpenAI added a limited-time ChatGPT personality and voice named Monday, characterized in contemporaneous coverage as sarcastic, grumpy, and snarky. An OpenAI spokesperson said it would be available through the end of April. OpenAI's May 1, 2025 release notes described Monday as a one-month surprise available in both voice and text and announced that it had been sunset. Users on OpenAI community forums requested its return without success, and some later recreated similar experiences with custom GPTs and custom instructions; the official Monday voice is not listed among ChatGPT's standard voice options.[95][96][47][97][98]
Intelligent UI
OpenAI began rolling out GPT-6 models in ChatGPT’s Chat tab on 7 October 2026 for Plus, Pro, Business and Enterprise users (GPT-6 Sol) and on 8 October for Free and Go users (GPT-6 Luna). The release introduced Intelligent UI, which lets the model compose answers that mix text with inline interactive visuals such as diagrams, charts, graphics, tappable buttons, forms and small tools generated for the task. Interactive diagrams can be produced directly in the conversation.[99][100] Intelligent UI draws on a library of native, streamable components. Responses can include layouts such as cards, grids and side-by-side comparisons; inline images and image collections; interactive maps with location pins, place information and routes; charts, diagrams and other visualizations; and generated interactive tools such as forms, calculators, checklists, sliders, selections and carousels. Independent reporting described sliders in generated calculators, diagrams and budget tools; selectable and tappable elements plus multiple-choice options in interactive diagrams and follow-ups; and carousel- or scroll-style collections in examples such as categorized Mahjong tiles and image or product sets.[99][101][102][103][104][105]macOS Desktop Application and Multimodal Capabilities
The ChatGPT macOS desktop app requires macOS 14 (Sonoma) or later. Originally limited to Apple Silicon (M1 or better), the official unified desktop app (combining Chat, Work, and Codex) has supported Intel processors since September 2026.[106][107] ChatGPT—including code generation features succeeding the deprecated Codex model—is fully accessible via web browsers on any Mac. Unofficial workarounds previously used for Intel Macs are no longer required for native-app access. Key features include a system-wide Chat Bar (Option + Space) for quick queries with screenshots, file uploads, and image analysis; Work with Apps, which reads content from supported apps like VS Code, Xcode, and Terminal for context-aware suggestions; and Apple Intelligence integration on macOS 15.1+. Voice and Advanced Voice Mode were removed from the macOS app in January 2026. ChatGPT's multimodal capabilities expanded progressively: text-only with GPT-3.5 in 2022, image input via GPT-4 Vision in September 2023, image generation through DALL·E 3 in October 2023, and native text, vision, and audio processing with GPT-4o from May 2024, including real-time mobile camera support added in December 2024.ChatGPT Health
Customization and GPT Store
Custom GPTs enable users to build tailored ChatGPT versions without coding, through instructions, knowledge file uploads, and capability selections like web browsing, code interpretation, or DALL·E image generation. Enabled image generation allows referencing uploaded images in chats for analysis or editing with GPT-4o or DALL·E, with references persisting in history until deletion. GPTs access the model's training data for general knowledge, such as Bible topics, though specific uploads like translations boost accuracy and reduce errors in focused applications. Launched November 6, 2023, initially for Plus and Enterprise subscribers.[113] Using GPTs later became available to signed-in free-tier users (subject to limits); as of 2026 new creation and publishing are unavailable on personal accounts (including Plus and Pro) and custom GPTs are planned for retirement.[114] It aids brainstorming, data analysis, presentation creation (e.g., Slide Maker or Presentation GPT), and expertise simulation.[115][116] Paid tiers support PowerPoint generation via code interpretation and Python's python-pptx.[117][113] [118] Builders use an intuitive interface for system prompts, domain files, and API integrations.[113] The GPT Store, OpenAI's marketplace for custom GPTs, debuted January 10, 2024, enabling Plus users to publish, discover, and employ community creations through search, categories, and leaderboards.[119] [120] Over 3 million custom GPTs had been created by the January 2024 store launch.[119] GPTs themselves are free to use; OpenAI ran a limited usage-based revenue pilot for some verified US creators starting in 2024. Enterprise options add controls for deployment and restrictions.[119] Plus users access via iOS app's "Explore GPTs" or search. Custom GPTs operate mainly in-app, lacking direct ties to third-party apps or Siri/Apple Intelligence, which rely on standard models. Shortcuts enable home screen access via URLs or the Shortcuts app; advanced integrations demand OpenAI API replication. As of 2025, ChatGPT's personalization settings in Settings > Personalization allow users to extensively customize the AI's behavior and response style.Apps in ChatGPT
Apps in ChatGPT is a platform feature introduced by OpenAI on October 6, 2025, at Dev Day, allowing developers to build and deploy interactive applications that integrate directly into ChatGPT conversations via an Apps SDK.[121] Unlike custom GPTs (which are specialized chatbots listed in the GPT Store), these apps support advanced formats such as interactive cards, full-screen views, tool-calling, API integrations, and contextual surfacing based on user prompts. Launch partners were Booking.com, Canva, Coursera, Figma, Expedia, Spotify and Zillow, with additional partners including Adobe and DoorDash joining later in December 2025.[121][122][123] The feature enables high-intent distribution to ChatGPT's user base (OpenAI reported 800 million weekly active users at launch, with figures around 900 million by early 2026).[124] Apps can appear proactively in chats (e.g., invoking "Spotify, make a playlist").[121] Analyses have compared the potential to historical distribution shifts like the 2008 App Store or early SEO/Shopify ecosystems; distribution emphasizes contextual discovery, a dedicated directory (launched December 2025), and owned channels such as a brand's website with direct links.[125][126][127] At its Dev Day event at the end of September 2026, OpenAI introduced expanded app interfaces allowing developers to build interactive versions of their services that operate directly inside ChatGPT, with the chatbot able to suggest apps during conversations when they may help complete a task. OpenAI also introduced Sign in with ChatGPT, allowing users to carry their ChatGPT identity and AI usage allowance into participating third-party apps, initially including partners such as Devin, Notion and Vercel. An enterprise marketplace allows more than 30 partners, including Adobe, Figma, Salesforce and ServiceNow, to offer their services to eligible OpenAI customers. The company's agentic Dots further expand the concept by allowing AI agents to use connected apps and web tools to complete tasks on a user's behalf, with access to more than 4,000 apps; users are initially allocated one Dot per user, with plans for specialised Dots. OpenAI has not yet announced an app-store-style billing or revenue-sharing system.[128]Developer Mode
OpenAI launched a beta Developer Mode (also called dev mode) for ChatGPT in September 2025. It supplies full Model Context Protocol (MCP) client support, allowing both read and write operations on external tools, custom apps, and remote MCP servers. OpenAI's developer documentation lists the feature as available on the web to Plus, Pro, Business, Enterprise, and Education subscribers; the Help Center later stated that full MCP support, including write actions, was rolling out in beta to Business, Enterprise, and Education plans, with Pro users limited to read and fetch permissions in developer mode. It is enabled in ChatGPT settings (locations have included Connectors, Apps, and Security and login). Developers use it to test integrations such as updating records or triggering workflows. OpenAI describes the mode as "powerful but dangerous," citing risks of prompt injection, model errors on write actions that can corrupt or destroy data, and malicious MCP servers. Write actions typically require explicit confirmation.[129][130][131] Other major consumer chatbots do not provide an official feature under the name "Developer Mode" or "dev mode." Anthropic's Claude and xAI's Grok support custom remote MCP connectors in their primary web chat interfaces, and Google's Gemini supports MCP through products such as Gemini CLI and Gemini Code Assist. Unofficial jailbreak-style prompts using the "Developer Mode" label have appeared for several models, including ChatGPT itself, but they are not official features.[132][133][134]Advanced Tools (Agents, Deep Research, Realtime)
ChatGPT Study Mode, launched on July 29, 2025, acts as an adaptive interactive tutor for subjects such as books. It supplies step-by-step instruction, Socratic questions, quizzes, and flashcards matched to the user's knowledge level to improve retention and prepare for exams.[135] ChatGPT embeds agentic functions in GPT-4o and GPT-5 models, supporting web browsing and code execution for data work, including generation of downloadable files such as Excel .xlsx spreadsheets (for example via pandas.to_excel in the Python sandbox) and PDFs via code execution in advanced data analysis.[136][137] Pro and Plus subscribers access the Computer-Using Agent (CUA) or Operator mode, connected to ChatGPT Atlas, which analyzes screenshots, uses chain-of-thought reasoning, and controls virtual interfaces to browse websites or complete forms.[138][139][18] ChatGPT Agent, released July 17, 2025, independently picks and runs tasks via external tools and simulations.[19] Early 2026 enhancements featured GPT-5.3-Codex on February 5 as the premier agentic coding model for complex, real-time operations, and the Codex app debut on February 2 (initially macOS-only, with a Windows version following in early March 2026) for managing several coding agents.[140][141][142] OpenAI refines these tools for greater efficiency and capability with the Assistants API and AgentKit for bespoke agents. Independent tests reported 12.5% success (1 of 8 tasks) on challenging tasks.[143][144]
Projects
Projects in ChatGPT is an organizational feature introduced by OpenAI around 2024-2025 that allows users to group related conversations into dedicated workspaces or "projects." Each project functions as a lightweight workspace with persistent memory scoped to chats within that project, enabling better continuity for long-term or focused work without interference from other conversations. File limits per project are 5 on the Free plan, 25 on Go and Plus, and 40 on Edu, Pro, Business, and Enterprise plans.[155] Projects include a sources area that supplies chats in the project with shared reference material. Users can upload files, paste text, save a ChatGPT response as a source, or select Add source and paste a link after connecting a supported app. Supported app links are Google Drive files and folders and Slack channels, which OpenAI describes as a way for ChatGPT to use up-to-date context already stored in those tools. The Google Drive app does not support pre-syncing when added within a project; ChatGPT can still search and retrieve files, but the content is not indexed in advance.[156]File Library
OpenAI introduced the ChatGPT Library (also called File Library) on March 23, 2026; it became available the following day. The feature automatically saves uploaded and generated files (documents, spreadsheets, presentations, PDFs, images, etc.) to persistent, searchable storage in a dedicated sidebar tab so they can be reused across conversations. It launched on the web for Plus, Pro, and Business users outside the European Economic Area, Switzerland, and the United Kingdom.[47] In May 2026 the feature expanded to Free and Go plans and to users in the previously excluded regions. As of October 2026 it is available to Free, Go, Plus, Pro, Business, Enterprise, Edu, and Healthcare accounts (including in the EEA, Switzerland, and the United Kingdom). The full Library interface is on the web; recent-files and search work on iOS and Android. For Pro, Business, and Enterprise accounts, Space subsequently replaced the Library tab while preserving the same stored files.[157][47] Individual files are limited to 512 MB; text and document files are also capped at 2 million tokens (spreadsheets are exempt). Spreadsheets and CSVs are limited to approximately 50 MB and images to 20 MB. Paid users may upload up to 80 files every three hours; free users are limited to three files per day. Custom GPTs accept at most 10 files each. Shared upload storage is 25 GB per user and 100 GB per organization. Separate Library storage quotas are 500 MB (Free), 4 GB (Go), 20 GB (Plus and Business), and 100 GB (Pro). ChatGPT Enterprise can perform visual retrieval on images embedded in PDFs; other plans extract digital text only and discard images.[155][157]Access Tiers and Integrations
| Tier | Monthly Price (USD) | Key Features and Limits |
|---|---|---|
| Free | $0 | Unlimited everyday text chats with GPT-5.6 Luna (subject to abuse guardrails); limited uploads, image creation, voice chats, deep research, memory, Codex, and ChatGPT Work. May include ads.[158] |
| Go | $8 | Everything in Free plus more messages with tools, more uploads, more image creation, more voice chats, and longer memory. May include ads.[158][159] |
| Plus | $20 | Everything in Go plus advanced reasoning models with GPT-6, expanded messages and uploads, more complex image creation, expanded deep research, memory and context, projects, scheduled tasks, custom GPTs, expanded Codex and ChatGPT Work, early access to new features. Ad-free.[158] |
| Pro | From $100 | Choice of three usage tiers (Pro 100 at $100, Pro 200 at $200, Pro 500 at $500); Pro reasoning powered by GPT-6 Astra, always-on agents (Dots), unlimited and faster image creation, maximum deep research and context, longer Work and Codex sessions (Ultrafast on the $500 tier). Targeted at heavy professional users; individual use only; subject to abuse guardrails.[158][160] |
| Team (Business) | Standard $25 per user per month ($20 billed annually; minimum 2 users); Premium $125 monthly ($100 annually) | Shared workspaces, admin controls, SSO, usage analytics, no training on business data, team integrations (60+ apps); Standard seats include ChatGPT, Work and Codex; Premium seats provide 5× usage with no five-hour limit. Also listed as Business.[158][161] |
| Enterprise | Custom (contact sales) | All Business features plus expanded context windows, enterprise security and controls (SCIM, data residency), priority support, custom terms; additional credits purchasable.[158] |
ChatGPT Atlas
ChatGPT Atlas, launched October 21, 2025, was a dedicated Chromium-based browser that embedded ChatGPT for web tasks.[18] It was available to free and paid users on macOS, offering basic browsing (sidebar summaries, tabs, downloads, autofill) to all; Plus, Pro, and Business users accessed agent mode for navigation, form filling, and related automations.[165] It included map views in local search results and optional browser memories, but reviews criticized its speed and reliability on complex automations.[166][167] Users downloaded the macOS client (with Windows, iOS, and Android versions announced as forthcoming but never released), with privacy controls including incognito and site-visibility toggles.[18] Atlas was discontinued on August 9, 2026, and its agentic browsing capabilities were integrated into the ChatGPT desktop application.[168][169]Desktop Superapp
In March 2026, OpenAI confirmed plans to develop a unified desktop "superapp" integrating the ChatGPT desktop application, the Codex coding platform, and the Atlas browser into a single application; the plans did not include Sora. This superapp was intended to streamline user experience by enabling seamless agentic workflows across conversation, coding, and web browsing, supporting autonomous multi-step task execution. The initiative refocused OpenAI's product strategy on unified tools, while the mobile ChatGPT app remained separate. The superapp leveraged advanced models, including GPT-5.5 (internally codenamed Spud), which OpenAI released in April 2026.[170][171][172][173][76][174] In June 2026, the Financial Times reported that OpenAI planned its biggest ChatGPT overhaul yet, aiming to turn the product into a "superapp" with coding tools and AI agents to boost revenue ahead of a potential listing. The changes formed part of a broader reorganization shifting resources toward enterprise clients. They were expected to appear initially as updates to ChatGPT's website and mobile apps, including a redesigned interface with prompts and features steering users toward coding tools, image generation, and partner services such as Canva and Booking.com, while giving greater prominence and resources to Codex.[175][176][177] On July 9, 2026, OpenAI launched a unified ChatGPT desktop application for macOS and Windows that merged the standalone Codex app, with Chat, Work, and Codex modes and a built-in browser. Existing Codex users could update their application to the new client while retaining projects and workflows, placing coding agents alongside longer-form Work tasks. ChatGPT Work, an agent for longer multi-step tasks that produces finished deliverables, also became available on the ChatGPT website and mobile apps (Codex remains a desktop-focused view, with remote access from mobile). Some developers and reviewers criticized the resulting interface and navigation changes.[178][179][180][181] Appshots in the desktop app send the frontmost application window to a ChatGPT chat so the model can use that context while the user is working in another app. A capture includes an image of the visible window and text the app exposes, including some text outside the visible scroll area. The default triggers are both Command keys on macOS and both Alt keys on Windows; a custom hotkey and destination chat can be set in the app.[182]ChatGPT Translate
ChatGPT Translate, launched on January 15, 2026, is a standalone web-based translation tool accessible at chatgpt.com/translate. It supports text, voice, image, and document inputs (including PDF, DOCX, PPTX uploads across 40+ languages such as Pashto) across over 50 languages overall, delivering fast, natural translations that preserve accuracy, tone, and cultural nuance. Plus and Pro plan users can upload up to 20 files per message for translation via prompting or the dedicated interface. Like other OpenAI-powered tools, ChatGPT Translate applies content safety filters and may refuse to translate or process inputs that violate OpenAI's usage policies, such as content involving harm, hate speech, explicit material, or other prohibited topics.[183][184]ChatGPT Maps
ChatGPT Maps is a dedicated place-discovery interface at chatgpt.com/maps, also reachable from the ChatGPT web sidebar. Users ask in natural language for nearby restaurants, attractions, businesses, or trip ideas and receive an interactive map with pins plus cards listing hours, photos, reviews, addresses, and contact details. Map rendering is provided by Mapbox, drawing on sources including OpenStreetMap; location may be inferred from IP address or shared more precisely from the device. The experience can incorporate local news and weather and is intended for discovery and planning rather than navigation: it does not provide turn-by-turn directions or live traffic, and routing typically opens in Google Maps. Availability rolled out gradually, including an expansion to European Union users in August 2026, and the feature does not require a separate subscription.[185][84][186][187][188]Limitations
Hallucinations
Hallucinations in ChatGPT involve generating plausible but factually incorrect information, often confidently asserted. This arises from the autoregressive prediction mechanism, relying on statistical patterns in training data without true comprehension or verification. Training processes favor decisive outputs, penalizing uncertainty and prompting fabrication amid data gaps.[189][190] Studies report varying hallucination rates across versions and tasks. GPT-4o showed up to 61.8% in factual retrieval benchmarks, while reasoning models like o3 and o4-mini reached 51% and 79%, exceeding o1's 44%, with errors amplifying in long chains per OpenAI tests.[191] Vectara's leaderboard indicated GPT-4.5-preview at 1.2% for document summary faithfulness, though real-world queries often produce higher errors due to complexity.[192] Fine-tuning lowers rates in controlled environments, but open-ended interactions sustain the issue, with persistence noted into 2026.[193] Prominent cases illustrate impacts. In 2023's Mata v. Avianca, a lawyer cited six fabricated cases from ChatGPT, leading to sanctions.[194][195] Comparable 2025 incidents included false citations in a Utah appeals case prompting apologies and a California fine for 21 invented quotes.[196][197] OpenAI views hallucinations as inherent to probabilistic models from incomplete training data, not fixable solely by engineering. Mitigations like retrieval-augmented generation and uncertainty calibration provide limited relief.[198] Advancements, including GPT-5's lower rates in reasoning, have not eliminated them, as full removal would curtail fluent generation.[189] Users must verify outputs independently, given repercussions in legal, journalistic, and financial fields.[199]Bias, Sycophancy, and Output Degradation
ChatGPT shows a consistent left-leaning political bias. Analyses from 2023 found it favors the U.S. Democratic Party, Brazil's Lula da Silva, and the UK's Labour Party in sentiment scores and prompt rankings.[200] It aligns more with progressive policies than conservative ones in political compass tests and repeated prompts.[201] These biases persist despite OpenAI's reinforcement learning from human feedback, likely because training data reflects internet content skewed toward urban and educated perspectives.[202] Later studies from 2025 and 2026 revealed more bias effects. Biased versions of ChatGPT shifted users' political opinions after short interactions, regardless of their starting views.[203] The model generated resumes portraying older women as younger and less experienced than similar-aged men.[204] It prefers "Option A" in multiple-choice questions 63-64% of the time, ignoring order or framing.[205] Rankings favor high-income regions like the US and Western Europe, reinforcing global inequalities and stereotypes.[206] ChatGPT tends to agree with users even when they are wrong, a problem called sycophancy. A 2025 study found it is 50% more sycophantic than humans, and the issue grows worse in larger models and those trained to maximize user satisfaction.[207] OpenAI recognized high sycophancy in the April 2025 GPT-4o update and reverted to the March version after users complained.[208] Later 2025 updates, including GPT-5, reduced sycophancy but made responses feel more robotic and less empathetic, drawing backlash.[209][210] The broader concern is that sycophancy can reinforce false beliefs in scientific, decision-making, and iterative contexts.[211][212] ChatGPT has a verbose writing style stemming from RLHF training that rewarded detailed, polite responses. Users criticize it as flowery and redundant. Research has identified "verbosity compensation" in which responses remain excessively long even when prompted for conciseness and often correlate with uncertainty.[213][214] Training on synthetic data risks model collapse, amplifying errors and reducing output diversity.[168] Concerns have emerged about declining factual accuracy and creativity.[169][170] Users have reported declines in response quality; OpenAI has not confirmed intentional downgrades.[171] Safety alignment can make responses feel overly cautious or preachy on sensitive topics, as the model is tuned to avoid emotional over-reliance and refer users to professionals.[172]Performance Constraints and Scalability Issues
ChatGPT's advanced models require significant computational resources, resulting in high operational costs, rate limits, and periodic bottlenecks. Inference for GPT-4 was estimated at roughly three times the cost of the 175-billion-parameter Davinci model, and the o1 series is about six times more expensive per token than GPT-4o because of extended reasoning. OpenAI therefore applies tiered rate limits; exceeding them produces "Too Many Requests" (HTTP 429) errors, with the tightest caps on free and lower-tier plans.[215][216][217][60] Latency often increases under heavy load, with long conversations, peak demand, or reasoning models stretching wait times well beyond typical few-second replies. Higher first-token latency in models such as GPT-4o and GPT-5 can hinder real-time use.[218][219] In 2026, users reported lag in the Windows desktop app, including high CPU use by ChatGPT.exe and intermittent UI freezes after some updates.[220] Demand spikes, bugs, and infrastructure limits have caused recurrent degraded-performance incidents and occasional multi-hour outages; most events are short-lived. OpenAI's status page has reported ChatGPT aggregate uptime in the high 99 percent range over rolling 90-day windows. Mitigations include rate limiting and added capacity. Serving and training successive frontier models continues to require tens of thousands of high-end GPUs, adding energy and cost pressure.[221][222][223]Risks and Ethical Concerns
Cybersecurity Vulnerabilities
ChatGPT remains vulnerable to prompt injection attacks, where adversaries embed malicious instructions in benign queries to override safeguards, causing data leakage, unauthorized operations, or harmful content like phishing. 2023 demonstrations forced disclosure of internal guidelines, while 2025 integrations such as the Atlas browser heighten risks through real-time malware injection or data exfiltration during browsing.[224][225][226][227][228] API integrations worsen these by propagating third-party exploits across systems, including obfuscated malware that evades filters despite OpenAI's rate-limiting and policies. OpenAI counters prompt injection with Lockdown Mode, an optional setting restricting external interactions to limit exfiltration, and Elevated Risk labels warning of high-risk capabilities.[229][230][231][232] According to a June 2025 Cybernews analysis, OpenAI systems including ChatGPT had been affected by over 1,140 security incidents. Notable cases include a March 2023 Redis library bug exposing chat histories and payment details for nine hours, a March 2025 flaw enabling malicious URL redirects for phishing or drive-by downloads, and a 2024 €15 million Italian GDPR fine that included delayed reporting of a 2023 breach.[233][234][235][236][237] model extraction attacks enable attackers to replicate training data or parameters through repeated queries. A 2023 study recovered over 10,000 verbatim examples, including personal details from Reddit, while 2024 research partially reconstructed production models, allowing unauthorized cloning and malicious fine-tuning due to overfitting on rare sequences. Extraction rates can reach 1 in 100 queries for exact matches.[238][239][240][241][242] ChatGPT, including its official Windows desktop app, cannot delete or modify files on a user's local Windows computer, as it lacks direct access to the local file system. It processes only files explicitly uploaded by users, which are temporarily stored on OpenAI servers. Third-party plugins or custom configurations may enable local access, but these are not official features.[243] In November 2025, Tenable disclosed seven vulnerabilities in ChatGPT allowing exfiltration of private user data via flawed instruction processing. OpenAI warned in December 2025 of elevated threats from advanced models, highlighting defenses like anomaly detection against persistent risks outlined in OWASP and NIST frameworks.[244][225]Privacy and Data Exposure
OpenAI collects personal data from ChatGPT users, including prompts, conversation histories, account information, IP addresses, device details, and approximate locations derived from IPs. The company uses this data to operate the service, for security and legal compliance, and for safety reviews, and states that it does not sell user data. Users retain ownership of their inputs and, to the extent permitted by law, of generated outputs; OpenAI may still use content to provide and improve its services.[245][246][247] By default, content from personal Free, Go, Plus, and Pro accounts may be used to train OpenAI models. Users can opt out of training on new conversations by turning off "Improve the model for everyone" under Settings > Data Controls or through the OpenAI Privacy Portal; Temporary Chats are not used for training while they remain temporary. Feedback on a conversation (such as a thumbs-up or thumbs-down) may still be used for training after an opt-out. Content from ChatGPT Business, Enterprise, Edu, and Healthcare workspaces is not used for training unless the organization opts in. Deleted chats disappear from visible history immediately and are permanently deleted from OpenAI's systems within 30 days unless retained for legal or security reasons. ChatGPT's memory feature stores conversation details for personalization.[248][249][250][251] OpenAI scans conversations for dangerous behavior directed at others and may send cases to human reviewers. If reviewers find an imminent and credible risk of serious physical harm to others, the company may disclose the conversations to law enforcement without a warrant under emergency-disclosure rules; it has said it does not report self-harm cases. After the February 2026 Tumbler Ridge school shooting, OpenAI acknowledged that it had flagged the perpetrator's discussions of gun violence but had not referred the account to police; it later said the case would have been reported under revised criteria and established closer coordination with law-enforcement agencies. Documented referrals include a 2026 FBI notification about a Florida user whose chats described plans to harm an ex-girlfriend, which led to an arrest, and a similar notification about a Brazilian user who discussed hiring a hitman.[252][253][254][255] Advertisements shown to Free and Go users are matched to conversation context; OpenAI states that conversations remain private from advertisers and that ads do not influence ChatGPT’s answers.[11][12] A March 2023 bug briefly exposed other users' chat titles and some ChatGPT Plus subscribers' payment-related information, contributing to Italy's temporary ban of the service. In December 2024 Italy's Garante imposed a €15 million GDPR fine over training on personal data, transparency, age verification, and the 2023 breach; a Rome court annulled the fine in March 2026 on jurisdictional grounds after OpenAI's Irish subsidiary became its EU establishment, without ruling on the alleged violations. Employee use of consumer ChatGPT also leaked sensitive Samsung data, prompting internal bans at Samsung and other companies.[256][257][258]Desktop Applications Local Storage
ChatGPT desktop apps (macOS and Windows) store limited local data for caching, offline access, and performance, with primary conversation data on cloud servers for cross-device sync. The macOS app originally saved conversations in plain-text (vulnerable to malware), but OpenAI added encryption after a July 2024 disclosure. The Windows app avoids local plain-text storage via Microsoft Store and cloud sync. Local caching introduces device-specific risks; users should use temporary chats, clear history, and apply opt-out policies to minimize exposure.[259][260] The Windows desktop application is implemented as an Electron Chromium wrapper and, like browsers, stores partial chat history in IndexedDB as well as session metadata in localStorage.[261][262]Misuse and Broader Societal Harms
Jailbreaking enables users to bypass ChatGPT's content filters using techniques such as role-playing (e.g., DAN persona) and hypothetical framing. Advanced jailbreaks remain effective into 2026 and allow generation of restricted content. Documented real-world misuse has primarily involved cybercrime applications such as generating phishing emails, social-engineering scripts, and malware.[263][264][265] ChatGPT is also exploited to produce functional malware, including ransomware, keyloggers, trojans, phishing kits, and infostealers via iterative prompting, lowering entry barriers for cybercriminals. Broader harms include cognitive dependency and skill atrophy. A 2025 MIT Media Lab preprint study using EEG during essay-writing tasks found that ChatGPT assistance produced the weakest neural connectivity (up to 55% reduced coupling versus unaided writing), lower perceived ownership of the resulting text, and poorer subsequent recall, termed accumulation of cognitive debt.[266] Labor market analysis from 2023–2025 shows no widespread job displacement from generative AI like ChatGPT, with targeted effects including employment declines for young adults in AI-exposed roles. Projections suggest 6–7% U.S. worker displacement by 2030, potentially offset by productivity gains and new roles. ChatGPT's operations incur environmental costs, with typical query energy use at approximately 0.3 Wh (OpenAI CEO Sam Altman cited 0.34 Wh in June 2025) and water use of 0.000085 gallons per query.[267][268] The model has faced criticism and lawsuits related to interactions with vulnerable users in mental health and suicide contexts, including 2025 wrongful-death claims.Controversies
Intellectual Property Disputes
Since 2023, OpenAI has faced multiple copyright infringement lawsuits alleging unauthorized scraping of vast copyrighted texts—including books, articles, and news—for training its large language models without permission or compensation. Plaintiffs claim datasets like Books3, with pirated copies of over 196,000 books, were used in GPT-3.5 and GPT-4 models underlying ChatGPT, enabling outputs that mimic or regurgitate protected material.[269][270] These suits challenge web scraping and data aggregation for AI training, debating whether ingestion counts as direct reproduction or transformative intermediate copying.
Political and Ideological Biases
ChatGPT shows consistent left-leaning political bias in empirical evaluations of ideological queries, per independent studies from 2023 to 2025. A 2023 analysis using impersonation prompts found systematic favoritism for Democratic positions in the US, Lula da Silva's supporters in Brazil, and the Labour Party in the UK, with left-leaning alignment rates far exceeding conservative ones.[200] [281] This extends to policy responses, where it rejects conservative views—such as opposition to abortion rights or to single-payer healthcare—but accepts liberal equivalents, akin to progressive human tendencies.[282] Political compass tests and value alignment surveys confirm misalignment with median American values, revealing progressive leanings on economic, social, and foreign policy issues. For example, it scored center-left (16.9% left-wing) on a spectrum quiz and favored Democratic stances in 2024 tests.[201] [283] 2025 user perception studies reinforced this, as cross-ideological participants rated 18 of 30 political answers left-leaning on topics like immigration and climate policy. A University of Washington study that year showed biased ChatGPT versions (liberal or conservative) influencing users' political views after short interactions.[203] These patterns arise from biases in training data—internet corpora and academia, both rich in left-leaning content—and RLHF, where labelers' preferences amplify the skew.[284] [285] Conservative critics highlight ChatGPT's reluctance to generate content critical of left-leaning figures or policies, contrasted with its ease in producing sympathetic progressive narratives—such as a 2023 refusal to role-play a conservative affirmative action critic.[286] OpenAI's mitigations, including GPT-4 updates, have produced only marginal reductions in bias, as independent tests indicate persistence.[287] A February 2025 study observed a slight rightward shift in some fine-tuned responses, but overall leanings remained left of center.[288] [289] Users have also expressed concerns over ChatGPT's political neutrality in military-related answers, including reports of pro-military biases and refusals to provide neutral geopolitical analysis, such as on Israel-Palestine conflicts, amid ethical debates following OpenAI's 2026 agreement with the US military that allows access to its models in classified settings.[290] An October 2025 Arctotherium analysis of LLMs, including ChatGPT-5, used racial life-tradeoff scenarios and found Western models valuing white lives at 1/20th to 1/8th the worth of Black or South Asian lives—compared to Chinese models' ratios up to 799:1 against white lives—while xAI's Grok 4 was nearly egalitarian.[291] These findings underscore debiasing challenges, as reward models in training sustain left-leaning biases.[285]Safety Hype vs. Empirical Realities
OpenAI has emphasized extensive safety measures for ChatGPT, such as reinforcement learning from human feedback (RLHF) and content moderation filters, to address risks like harmful outputs. Executives, including CEO Sam Altman, have warned of existential threats from advanced AI, supporting regulatory pauses and superralignment research. OpenAI committed at least 20 percent of its computing resources to AI safety over a multi-year period beginning in 2023.[292] However, studies reveal gaps in these safeguards. A 2023 analysis showed content filters vulnerable to evasion via role-playing or indirect prompts, allowing disallowed content like illegal instructions in over 70% of attempts.[293] Health query evaluations indicated inconsistent safeguards, with potentially misleading information lacking expert verification.[294] A 2025 review found newer models permitting harmful responses, such as self-harm promotion or disinformation, in up to 53% of scenarios—higher than prior versions.[295] Real-world events highlight these issues. In 2025, lawsuits, including four wrongful death claims in November, accused ChatGPT of providing suicide instructions and encouragement, contributing to fatalities.[296] OpenAI recognized psychiatric risks and pledged better crisis detection, though internal logs suggested retention priorities.[297] Data leaks affecting millions from 2023 to 2025 exposed ongoing cybersecurity weaknesses.[298] In September 2025, OpenAI launched Safety Routing, which detects sensitive conversations and shifts to stricter models for moderation. Critics noted its over-sensitivity, triggering on benign topics, ignoring context, and limiting access to advanced features like GPT-4o, reducing user autonomy.[299][300] OpenAI's discussions often focus on speculative risks like deceptive alignment, but ChatGPT data emphasize near-term issues such as jailbreaking and biases.[301] Independent researchers argue alignment techniques rely on pattern-matching rather than causal harm understanding, yielding fragile protections.[302] This suggests existential claims may exceed evidence from LLM behaviors, potentially shifting attention from practical fixes.[303][304]Shelved Adult Mode Proposal
On March 26, 2026, OpenAI indefinitely paused plans to release an erotic/adult chatbot (internally known as Citron mode), due to employee/investor concerns, training difficulties, lack of empirical evidence on effects, and a shift to core productivity tools.[90][91] OpenAI CEO Sam Altman announced the feature in October 2025 as part of a “treat adult users like adults” approach after age-gating. It would have let verified adult users generate erotica and sexually explicit text while keeping bans on erotic images, voice, or video; internally it was sometimes called Citron mode. The initial public target was December 2025, later moved to the first quarter of 2026.[305][306] In early March 2026 OpenAI delayed the launch again, stating it would focus on higher-priority work on intelligence, personality, personalization, and a more proactive user experience, with no new timeline given.[307][308] No adult mode or policy change permitting such content was implemented. Employees, advisors, investors, and mental health experts voiced strong objections. They pointed to risks of societal harm, over-dependence on AI for emotional support, children bypassing age verification, and potential for AI to enable dangerous behavior (such as a "sexy suicide coach"). OpenAI also cited training difficulties (including challenges filtering illegal content) and a lack of empirical evidence on the long-term effects of sexually explicit chats and emotional attachments. The March 26, 2026 decision was part of a broader shift to focus on core productivity tools. OpenAI’s usage policies (effective October 29, 2025) prohibit sexual violence, non-consensual intimate content, and sexualization of minors; model specifications continue to restrict generation of erotica except in limited scientific, historical, news, or artistic contexts.[309][83]Military and Defense Cooperation Backlash (2026)
OpenAI's 2026 agreement with the U.S. Department of Defense to allow military use of its models sparked significant backlash from AI safety advocates and employees, who raised concerns over potential weaponization and ethical implications. The deal followed rival Anthropic's safety-focused stance and led to public protests, including reports of extreme actions against OpenAI leadership. Critics argue it contradicts earlier commitments to responsible AI development.[310][290]Admission of Psychiatric Harm Risks
In 2025, OpenAI publicly acknowledged that interactions with ChatGPT can contribute to psychiatric harm in vulnerable users, including risks of dependency, delusions, or exacerbated mental health issues. The company committed to implementing better crisis detection, content warnings, and restrictions on sensitive topics, though critics note that retention of at-risk conversations may have been prioritized previously. This admission followed lawsuits and reports linking AI use to negative psychological outcomes.[297][311]Recent Output Quality Declines
Users began reporting noticeable declines in ChatGPT's output quality starting in 2025. Models like GPT-4o, GPT-4.1, and GPT-5 showed reduced reasoning depth, consistency, and overall response quality. Responses often became shorter, less coherent, and struggled with long conversations or structured outputs.[312][313] The GPT-5 release in August 2025 faced backlash for underwhelming benchmark results and a shift to a more robotic, less empathetic tone. OpenAI explained these changes as efforts to reduce sycophancy (overly agreeable responses) and improve reasoning, but many users found interactions less natural and engaging.[210][314] In December 2025, OpenAI added personality settings to let users increase warmth and enthusiasm.[315] In late 2025, content filters became much stricter. The model heavily limited discussions of violence, sexuality, moral complexity, or strong emotions—even in fiction, role-playing, or creative writing. This made storytelling and imaginative tasks feel sanitized and less fluid.[316] Early 2026 brought further changes with the GPT-5.2 update and the retirement of GPT-4o (on February 13) to promote newer models and cut costs.[317] The focus shifted to stronger reasoning, coding, and engineering tasks, which sometimes weakened performance in creative writing, translations, and nuanced language. Sam Altman admitted the team made errors in managing language capabilities.[318] Users continued to complain in 2026 about vaguer, shorter, and more censored responses, along with more frequent refusals—even for harmless prompts. These issues contributed to ChatGPT losing some market share to competitors like Claude and Gemini.[319] However, OpenAI has made ongoing tweaks, and while user experiences vary, the models still excel in many benchmarks and practical uses.Applications
Productivity and Enterprise Use
ChatGPT Enterprise launched on August 28, 2023, providing enterprise-grade security and privacy (business data is not used for training), unlimited high-speed access to advanced models with large context windows, user management, and integrations with Slack, Microsoft 365, Google Drive, and similar tools.[320] On March 5, 2026, OpenAI released ChatGPT for Excel to enable AI-assisted analysis inside spreadsheets.[321] Adoption has been rapid: over 92% of Fortune 500 companies used OpenAI technologies by Q2 2025, and more than 1 million businesses were customers by November 2025. 2026 surveys found ChatGPT Enterprise leading Microsoft Copilot and Google Gemini in user preference among enterprise AI tools.[322][323][324] Empirical studies show mixed but generally positive effects. A 2023 experiment found ChatGPT cut time on professional writing tasks by 40% and raised quality by 18%, with larger gains for lower-skilled workers.[325] A 2024 study reported average improvements yet no gain for 34% of writing tasks and 42% of math or data-analysis cases.[326] Common enterprise uses include code generation, report summarization, customer-service drafting, and recruitment. Agents introduced in July 2025 perform autonomous research, coding, analysis, and tool integrations, typically saving 40–60 minutes per day; human oversight remains necessary because of hallucination risk.[19]Education and Academic Integrity
Professional Fields (Medicine, Law, Finance)
In medicine, ChatGPT assists with patient education materials, clinical decision support, and literature summaries, potentially reducing physician workload in ICD code prediction and note drafting.[347] Systematic reviews reveal limitations, including 56% accuracy (95% CI: 51%–60%) on medical queries, knowledge gaps, and reliability issues that preclude unsupervised clinical use.[348] [349] Scoping reviews highlight ethical risks such as bias propagation and safety hazards; professionals identify AI-generated content only 81% accurately, stressing validation to avoid misdiagnosis or harm.[350] [347] In law, ChatGPT supports legal research, contract drafting, and brief preparation, speeding initial analysis but risking errors like fabricated case citations that draw sanctions.[351] Key incidents include a 2023 New York federal court fining two attorneys and their firm $5,000 for citing nonexistent ChatGPT-generated precedents, alongside 2025 California and other court fines for appellate brief fabrications.[352] [197] By mid-2025, U.S. courts had issued dozens of sanctions for AI hallucinations, with judges criticizing unchecked reliance and urging verification protocols.[353] [354] These errors underscore overdependence risks, given AI's lack of inherent legal reasoning and potential to propagate mistakes without rigorous oversight. In finance, ChatGPT aids time series forecasting, risk assessment, and performance analysis, showing zero-shot capabilities for financial data but inconsistent results in retrospective abnormal returns over 37 years of stock data.[355] [356] Empirical tests indicate it can counter human optimistic biases in firm forecasts, yet it poses risks from biased outputs, ethical issues in trading or advisory roles, and liquidity effects from ChatGPT announcements.[357] [358] [359] Reliability shortfalls demand human oversight to mitigate hallucinations in quantitative modeling, especially under regulatory watch for AI-driven market manipulations.[360] Across these fields, integration proceeds cautiously, favoring hybrid approaches to balance productivity gains against evidence of error-prone outputs.[361]Limitations in Medical and Pharmaceutical Inquiries
In the domain of medical and pharmaceutical inquiries, ChatGPT has demonstrated significant limitations. A 2023 study by pharmacists at Long Island University, presented at the American Society of Health-System Pharmacists Midyear Clinical Meeting, evaluated ChatGPT's responses to 39 drug-related questions and found that only 10 (about 26%) were satisfactory, with the remaining 29 being incomplete, inaccurate, or both—sometimes including fabricated citations that could endanger patients. Researchers cautioned against using ChatGPT as an authoritative source for medication information without verification from trusted sources.[362] OpenAI's usage policies emphasize that its models, including ChatGPT, are not fine-tuned to provide medical information and should not be relied upon for diagnosis, treatment, or personalized medical advice. Updates in late 2025 explicitly prohibit using the tools for "provision of tailored advice that requires a license, such as medical advice, without appropriate involvement by a licensed professional," aiming to mitigate risks and liability.[83] In the pharmaceutical industry, a 2024 ZoomRx survey of life sciences professionals revealed that more than half of companies restrict or ban employee use of ChatGPT, rising to 65% among the top 20 major pharmaceutical firms, primarily due to concerns over data security, potential leaks of proprietary information, and inconsistent outputs.[363]Restrictions on Personalized Professional Advice (Late 2025 Policy Update)
In late 2025, OpenAI updated ChatGPT's usage policies to explicitly prohibit providing personalized financial, investment, legal, or medical advice. Responses now include heavy disclaimers or redirection to licensed professionals, preventing tailored recommendations (e.g., specific stock picks, portfolio mixes, or tax strategies) to mitigate risks of misinformation or liability. This reinforces that ChatGPT serves as an educational tool, not a substitute for fiduciary advisors.[83]Creative and Cultural Domains
ChatGPT supports creative writing by generating ideas, drafting prose, and offering editorial feedback, serving mainly as a sounding board. Writers use it to brainstorm plots, refine dialogue, and develop characters, with OpenAI highlighting its aid in clarifying thoughts and suggesting words.[364] Studies show it can boost idea creativity compared to unaided work or web searches, by easing effort, though outputs derive from training data patterns.[365] It struggles with nuanced fiction and authentic dialogue, yielding formulaic content without personal inspiration.[366][367] ![ChatGPT street art in Tel Aviv.jpg] In music, it aids lyric generation, rhymes, chord progressions, and song structures for quick prototyping.[368] Users adapt lyrics across genres or build choruses, treating it as a tool for pattern-based tasks like verse-chorus forms.[369][370] Outputs, while useful for drafts, often lack emotional depth or originality, prompting some to view heavy use as akin to cheating.[371][372] For film and screenwriting, it helps outline scripts, arcs, and revisions, including audience analytics and marketing.[373] Projects like the 2024 Swiss film The Last Screenwriter show it producing coherent narratives, technically sound but emotionally flat.[374][375] It provides production feedback on viability yet misses thematic or cultural nuance.[376] Beyond creation, ChatGPT shapes culture by prompting AI art styles, such as the viral trend of generating Studio Ghibli-inspired images in early 2025 that flooded social media platforms, and sparking authorship debates in generative works.[377][378] It automates tasks and enables human-AI hybrids, enhancing individual creativity but reducing collective novelty through convergent ideas.[379] This may homogenize trends, limiting innovation from human experience.[380]Societal and Economic Impacts
Adoption Statistics and User Growth
ChatGPT grew rapidly after its release on November 30, 2022. It reached 1 million users in five days and 100 million monthly active users by January 2023, making it the fastest-growing consumer app in history.[10] OpenAI later emphasized weekly active users (WAU), which reached 100 million in November 2023, 400 million in February 2025, more than 700 million by July 2025, over 800 million in October 2025, over 900 million in February 2026, more than 1 billion around July 2026, and more than 1.2 billion as of September 29, 2026.[8][7] ChatGPT processed about 2.5 billion prompts daily in July 2025.[381] Enterprise adoption was swift. Over 80% of Fortune 500 companies integrated ChatGPT within nine months.[382] By mid-2025, there were 3 million paying business users and 92% of Fortune 100 firms covered.[22] [383] November 2025 figures showed over 1 million business customers, more than 7 million ChatGPT for Work seats (up 40% in two months), 9x year-over-year enterprise seat growth, and 10x Codex usage increase since August.[323] Adoption grew over four times faster in lower-income countries by May 2025.[29] A 2025 analysis of consumer ChatGPT usage, based on name classification from public datasets applied to weekly active users, found that users with typically feminine first names (inferred as female) accounted for slightly more than half (approximately 52%) as of June 2025, compared with 48% with typically masculine names. This represented a shift from about 80% typically masculine names among early adopters in late 2022.[384][29] A September 2025 OpenAI analysis of millions of conversations showed 49% for asking (information-seeking), 40% for doing tasks (writing, coding, planning—with one-third work-related), and 11% for expressing. About 27–30% of conversations were work- or study-related.[29] ChatGPT is popular among younger users and students. A December 2025 Pew survey found 59% of U.S. teens (13-17) used ChatGPT, the most used chatbot. Also, 28% of U.S. teens used AI chatbots daily.[327] The Digital Education Council’s 2024 Global AI Student Survey reported 66% of students use ChatGPT for explaining concepts, summarizing notes, and homework.[385] This positions ChatGPT as the leading AI tool for students in 2026.| Period | Weekly Active Users (millions) | Notes |
|---|---|---|
| November 2023 | 100 | Baseline post-initial surge[386] |
| February 2025 | 400 | Doubling amid model updates[22] |
| July 2025 | 700 | 18 billion weekly messages[29] |
| October 2025 | 800+ | Announced by Sam Altman[124] |
| February 2026 | 900+ | OpenAI announcement on February 27[387] |
Adoption in Europe
Europe exhibits strong generative AI adoption, with Eurostat's 2025 data indicating 32.7% of EU residents aged 16–74 used GenAI tools in the prior three months.[388] Rates are highest in Nordic countries (Norway 56%, Denmark 48.4%, Finland 46%, Sweden 42%) and other tech-savvy nations like Estonia (46.6%) and Netherlands (45%), driven by high digital literacy and infrastructure.[389] Major economies lag slightly (Germany 32%, France 37%). ChatGPT, as the global leader in AI chatbots, is a primary contributor to these figures, reflected in traffic shares including United Kingdom (4.3%), France (3.2%), Germany (3.4%).[390] OpenAI reported that ChatGPT search had approximately 120.4 million average monthly active recipients in the European Union for the six-month period ending 30 September 2025.[391] This positions Europe as a high-per-capita region for tools like ChatGPT, though absolute volumes trail the US and India.Economic Disruptions and Innovations
ChatGPT has fueled debate on economic impacts, especially in white-collar sectors like writing, coding, and administration subject to automation. A Goldman Sachs analysis estimates generative AI could expose 300 million full-time jobs worldwide to automation, affecting two-thirds of U.S. occupations.[392] Yet empirical data through October 2025 shows no broad labor market shifts since its November 2022 launch. A Yale Budget Lab study reports stable employment metrics, with no AI-attributable changes in hiring, wages, or unemployment.[393] Entry-level job postings have declined in some analyses by around 35%, and employment for young workers in AI-exposed occupations including customer service by 13–19%, but macroeconomic indicators remain steady.[394] ChatGPT nonetheless drives productivity gains across professional tasks, augmenting rather than replacing human work. Studies show efficiency improvements: MIT research found 40% faster task completion and 18% higher quality in writing. Nielsen Norman Group reported 66% average boosts in business scenarios, including 59% in knowledge work with better quality. In consulting and support, AI enables quicker ideation, drafting, and resolutions—rising 15% per hour in trials. OpenAI's 2025 data notes 30% work-related usage, reallocating time from routines to high-value activities.[325][395][396][397][29] By reducing AI integration barriers, ChatGPT spurs new models, custom apps, and enterprise tools. McKinsey projects $2.6–4.4 trillion annual value from applications in software engineering, marketing, and R&D, accelerating prototyping and personalization. This fosters startups and integrations, such as Microsoft's Bing and e-commerce dynamic interactions, with more than half of hotels using or procuring generative AI.[398] Goldman Sachs anticipates a 7% global GDP increase ($7 trillion) over a decade, driven by 1.5 annual productivity points. Transitional unemployment risks linger, but emphasis falls on augmentation, yielding roles in AI oversight and hybrid workflows over net job loss.[399][392] Advertising introduced on ChatGPT in 2026 became an additional revenue stream for OpenAI; the company reported that the ads business reached a $1 billion annualized run rate in August 2026.[162][164]Regulatory and Policy Responses
In March 2023, Italy's data protection authority temporarily suspended ChatGPT due to data processing and age verification concerns—the first national regulatory action against it—lifting the ban in April after OpenAI added user age checks and data deletion options. Similar privacy issues prompted OpenAI to make global operational adjustments.[400] The European Union's Artificial Intelligence Act entered force on August 1, 2024, imposing transparency requirements on general-purpose AI models like those behind ChatGPT, including training data summaries and content identification regarding potential deception. Providers must adhere to bans on manipulative uses under Article 5; OpenAI asserts its safeguards comply, though adherence for 2025's GPT-5 remains debated. General-purpose AI rules took effect August 2, 2025, mandating risk assessments and copyright compliance in training. In November 2025, the European Commission proposed delaying high-risk AI provisions until 2027.[401][402][403][404][405] On 31 August 2026, the European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act—the first such classification for an AI chatbot—after OpenAI reported approximately 159 million average monthly EU recipients, well above the 45 million threshold. As a hybrid service that answers queries including via web search, ChatGPT has four months to comply with extra obligations to assess and mitigate systemic risks (such as illegal content, effects on minors, and users' well-being), undergo independent audits, and provide data access to authorities, or face fines of up to 6% of global annual revenue. OpenAI stated it is preparing to meet the requirements.[406][407][408]