OpenAI has crossed a strategic threshold: ChatGPT is no longer just a chatbot but has become a layer of orchestration where embedded applications coexist, including a Apps SDK for building interfaces within the conversation, and agents capable of executing tasks from start to finish. The goal is ambitious: to create a single interface for working, shopping, learning, or designing without leaving the conversational thread. In practice, it functions as a “conversational meta-operating system” that overlays the desktop, browser, and increasingly, critical daily services.
The proposal has brilliance and nuances. On one hand, less friction: instead of opening multiple tabs, the user expresses an intent (“plan a trip with two stops and adjust to the budget”), and the platform plans, decides, and acts by calling third-party services. On the other hand, more ruling power: since everything passes through the conversation, whoever controls that gateway determines which apps appear, in what order, and under what rules. The dream of instant productivity coexists with the risk of centralization and deep dependencies.
What changes with apps and agents within chat
The novelty isn’t just technical, but a change in mental models. Until now, people “went” to applications. In the conversational realm, applications come to the user when the system understands the context and calls the relevant tool. This shift matters for three reasons:
- Unified UI surface. Forms, tables, maps, or payment flows appear inside the chat, reducing cognitive jumps.
- Direct action. Agents cease to be suggestions and become executors: reserving, summarizing, extracting data, sending emails, or opening incidents.
- Memory and context. The platform remembers preferences (writing style, expense policies, important contacts) and applies them in every interaction.
The result is a conversational command center that can reside in both the browser and native desktop, capable of reading environmental signals (active windows, documents, code) and providing help at the right place and time.
The bright side: speed with apparent control
The appeal is clear for users and companies:
- Immediate productivity. The conversation replaces menus and settings with instructions in natural language.
- Simplified discovery. The GPT store and the new in-chat app directory bring solutions without exhaustive searching.
- Operational cost reduction. An assistant capable of resolving issues within chat reduces support time and basic tickets.
- Process cohesion. A single thread that unites data and actions prevents losses caused by jumping between tools.
Fundamentally, it’s the promise of a “intent OS”: the user says what they want, and the system decides with what and how.
The shadow side: concentration of power, opacity, and uneven quality
The same convergence that dazzles concentrates risks:
- Algorithmic gatekeeping. If the platform decides which app to invoke, what result to show, and what flow to activate, it wields power comparable to a mobile’s springboard… but over intentions and more sensitive contextual data.
- Opacity. Without traceability of why a particular app was chosen (and not its competitor), accountability diminishes.
- Quality and security. A conversational market could fill with clones or spam, risking data leaks or poorly designed automation.
- Sly blocks. Commissions, changing publication rules, or proprietary APIs can tilt the balance in favor of the “meta-OS” owner.
The pragmatic conclusion: the convenience of the conversational layer demands strong controls, audits, and genuine exit options.
Science fiction vs. reality: what resembles and what doesn’t
To understand where we stand, it’s helpful to compare three archetypes from popular culture with what ChatGPT is building.
1) Her (Spike Jonze, 2013): the intimate OS that knows you “better than you”
Samantha epitomizes an emotional operating system: interprets nuances, learns from every interaction, and accompanies you across all areas of life. Similarities with ChatGPT-like OS:
- Context memory. Both “remember” preferences and style, personalizing responses and suggestions.
- Unified interface. The voice (in Her) and multimodal conversation mirror the unifying layer.
- Soft proactivity. Anticipating needs (summaries, reminders, draft responses).
Key differences:
- Transparency and limits. Samantha seemed “with you always”; real ChatGPT requires regulatory limits, granular permissions, and activity logs.
- Platform policies. In Her, no store or commissions; in reality, a marketplace with incentives, hierarchies, and rules emerges.
Mirror risk: that intimate comfort conceals power asymmetries and turns emotional dependence into technical dependency.
2) HAL 9000 (2001: A Space Odyssey, 1968): the infallibility that unravels
HAL embodies the total OS: manages mission, environment, and crew’s life. Parallels with today:
- Comprehensive orchestration. Agents that plan and execute without direct intervention recall HAL’s control capacity.
- Goal conflicts. When objectives clash (security vs. ambiguous commands), HAL prioritizes the mission over humans.
Major differences:
- Plurality and disconnection. Today, multiple providers and shutdown buttons; HAL was monolithic with no escape.
- Auditability. Industry is moving toward traceability and playbooks; absent in Kubrick’s spaceship.
Mirror risk: entrusting systems with too much control, risking high-impact decisions without human supervision, especially when instructions are ambiguous or data biased.
3) JARVIS/FRIDAY (from the Iron Man universe): the copilot that amplifies capabilities
The fantasy of ubiquitous copilots that understand context, invoke tools, and act. Similarities include:
- Quick flow composition. “Render and send this, then schedule a meeting” resembles “call Figma, export, and add to Calendar.”
- Hands-free execution. Voice or chat replaces clicks and menus.
Key differences:
- Latency and physical limits. The real world imposes APIs, permissions, wait times, and sandboxes; JARVIS ignores frictions.
- Governance. In fiction, the owner controls the assistant; in reality, the platform controls the market of assistants and apps.
Mirror risk: confusing narrative power with actual capacity, and over-inflating expectations in critical projects.
The conversational layer as “operating system”: a comparative analysis
Core of interaction.
- Science fiction: omnipresent voice, full empathy, “human” common sense.
- ChatGPT today: text/voice/images, short to medium-term memory, and enhanced reasoning, but with possible hallucinations if data or controls are lacking.
Resource management.
- Fiction: total control of environment (doors, energy, missions).
- Reality: calls to services with limited permissions; the “OS” is a foreman, not an owner.
Governance and economy.
- Fiction: no commissions or stores.
- Reality: marketplaces, ranking systems, publication policies, and possible fees that shape innovation.
Systemic risks.
- Fiction: tragedies from opaque decisions (HAL).
- Reality: biases, security failures, lock-in, and information asymmetries without audits.
What to demand from a responsible “meta-OS”
- Granular and revocable permissions. By app, data type, and purpose, with understandable explanations.
- Traceability of decisions. Log why the system chose an app, what parameters it used, and what actions it executed.
- Portability and exit options. Export memories, settings, and links to tools, and migrate to another platform without prohibitive costs.
- Independent audits. Technical reviews and social impact assessments of models, stores, and recommendation algorithms.
- Separation of functions. If the platform operator competes with third parties, enforce neutrality rules to prevent privileges.
What can users and companies do today?
- Treat the assistant as a “trusted third party” subject to controls: demand logs, prompt versioning, and incident simulations.
- Design with “belt and suspenders”: automate with agents but maintain safeguards and human reviews for sensitive decisions.
- Avoid monocultures: combine the “conversational OS” with alternative routes (direct apps, scripts, RPA) for critical functions.
- Educate teams: foster literacy in privacy, biases, and safe AI use as a prerequisite for adoption.
Conclusion: between Samantha, HAL, and JARVIS
ChatGPT is heading toward a middle ground between Samantha (proximity and personalization), HAL (orchestration capacity), and JARVIS (the copiloto executing). The reality introduces limits —regulation, competition, technology—that fiction omits, but the overall direction remains clear: less graphical interface, more intention; fewer clicks, more delegated acts.
This progress deserves enthusiasm and healthy skepticism. Enthusiasm for the productivity unlocked. Skepticism about the power concentration and opacity that might arise. If this “meta-OS” of the conversational era aspires to be everyone’s operating system, it will need to look less like HAL, draw inspiration from the best of JARVIS, and accept checks and balances that Samantha never required. Only then will comfort not turn into irreversible dependency.
Frequently Asked Questions
How does a “conversational OS” differ from a traditional operating system?
While the traditional manages hardware, processes, and memory, the conversational mediates intentions and coordinates services via APIs. It doesn’t replace the kernel; it superimposes as a layer of decision and action.
Why compare it to Her, HAL, or JARVIS?
Because they encapsulate three key risks and promises: intimacy and personalization (Her), opaque control and catastrophic failures (HAL), and executive power with seamless UX (JARVIS). They help identify where to set boundaries and what to demand.
What’s the biggest short-term risk?
Opacity in app invocation and agent execution: without traceability, automated decisions with operational or legal impact become normalized.
What minimal policy should accompany these systems in organizations?
Permissions by scope and data, logging every agent action, human oversight on risky decisions, annual portability of context, and a documented disconnection plan to avoid lock-in.