ChatGPT Sovereignty Score: The TEE Method Assessment

The Question

How sovereign is your organisation when the AI platform it depends on can change its terms — and its model — without your consent? The TEE Method (Transparency, Ethics, Empowerment) was designed to answer exactly this question. It evaluates AI systems across five domains: Strategic Alignment, Technical Performance, Ethical Compliance, Sovereignty Impact, and Cultural Alignment. Each domain is scored from 1 (critical risk) to 5 (full sovereignty). The total out of 25 tells you, in concrete terms, how much of your institutional future you control — and how much you have surrendered to a single provider.

ChatGPT, specifically the GPT-4o and GPT-5 family of models, is the most widely deployed AI system in human history. As of mid-2026, OpenAI reports over 400 million weekly active users, with ChatGPT Enterprise deployed across more than 70% of Fortune 500 companies. This assessment applies the TEE Method rigorously to ChatGPT as a deployed system — not to the abstract capabilities of large language models, but to the specific product, terms of service, data practices, and geopolitical posture that institutions encounter when they integrate OpenAI’s offering. What follows is not a review of features. It is a sovereignty audit.

Part One: Training Data Opacity and the Black Box Problem

The first and most fundamental sovereignty question is simple: what was the model trained on? OpenAI’s answer, across every version of ChatGPT from GPT-3.5 through GPT-5, has been a studied ambiguity. The company publishes high-level summaries — “a diverse corpus of text from the internet,” “licensed data,” “publicly available data” — but refuses to disclose specific sources, dataset compositions, or the provenance of training materials. This opacity is not accidental; it is structural. In litigation ranging from The New York Times v. OpenAI to consolidated author class actions, discovery proceedings have revealed that OpenAI itself struggles to document the full contents of its training corpora. If the creator cannot trace its data, no user can either.

The consequences for institutional sovereignty are severe. When you cannot verify training data provenance, you cannot assess the model for embedded biases, copyrighted content, or private information. Every output carries latent risk: IP infringement claims, privacy violations, or regulatory penalties under frameworks like the EU AI Act, which imposes transparency obligations that ChatGPT’s current architecture cannot satisfy. The EU AI Act’s Article 28, for instance, requires providers of general-purpose AI models to disclose “a sufficiently detailed summary of the use of training data protected under copyright law.” OpenAI’s response has been to argue that compliance is impractical — an admission that is itself a sovereignty red flag.

Moreover, training data opacity is not static. As models are updated, fine-tuned, and replaced, the data composition shifts. GPT-5’s training corpus is not GPT-4o’s, and neither is what GPT-4 was trained on. Institutions that evaluated ChatGPT in 2023 on a given set of risk parameters cannot assume those parameters hold in 2026. The model you approved last year is not the model running today. Without versioned model cards, documented dataset manifests, and auditable training pipelines — none of which OpenAI offers — every deployment is an act of faith.

Part Two: Vendor Lock-In and the Economic Moats

Vendor lock-in in AI is fundamentally different from lock-in in traditional enterprise software. With a database or CRM, switching costs are primarily about data migration and retraining. With AI, the switching costs are cognitive, institutional, and structural. Your organisation does not just store data in ChatGPT — it develops practices around it. Prompts become workflows. Outputs become processes. The model’s particular reasoning style, its formatting conventions, its handling of domain-specific terminology, and its integration with your existing toolchain become embedded in how your teams operate. Replacing ChatGPT is not like replacing Salesforce. It is like replacing the way your organisation thinks.

OpenAI has built economic moats specifically designed to make exit impossible. ChatGPT’s plugin ecosystem, custom GPT store, and API integration patterns create network effects that deepen with every connected service. The company’s enterprise agreements typically include volume-based pricing that makes per-seat costs punitive for organisations that maintain parallel AI subscriptions. And critically, OpenAI’s proprietary fine-tuning API means that any domain-specific customisation you invest in — your legal review templates, your medical coding helpers, your customer service personas — are tied to a model family you do not control and cannot deploy elsewhere.

The deeper sovereignty concern is that OpenAI controls the model plane entirely. You do not get to decide when GPT-5 is retired or replaced. You do not get to veto a capability regression — models that were good at reasoning may become worse at creativity, or vice versa, with no recourse. You do not get to opt out of safety filters that might block legitimate institutional use cases. In early 2025, for example, OpenAI updated ChatGPT’s refusal behaviour, causing a wave of unexpected blocks on medical and legal queries that had previously worked without issue. Enterprise customers discovered the change when their workflows broke. No warning. No rollback option. No compensation for the disruption. That is the nature of full-stack AI vendor lock-in.

Part Three: Data Usage for Training and the Privacy Betrayal

Few issues have damaged institutional trust in OpenAI more than the company’s evolving position on using customer data for model training. When ChatGPT launched, the default setting allowed OpenAI to use all conversation data to train and improve its models. There was no opt-out for free users. Enterprise customers were told their data was excluded, but the boundary between enterprise and consumer data within OpenAI’s infrastructure has been repeatedly questioned — including in internal communications revealed during the New York Times discovery process, where employees discussed the difficulty of ensuring that enterprise data was genuinely isolated from training pipelines.

The April 2024 data policy update marked a watershed. OpenAI quietly changed its terms to allow using customer API data for training unless customers explicitly opted out via a portal that many did not know existed. The change was not proactively communicated. It was discovered by privacy researchers monitoring the company’s legal pages. For organisations handling sensitive data — healthcare under HIPAA, legal work under attorney-client privilege, financial data under SOX — this policy ambiguity is existential. If your client meeting notes, patient records, or trade secrets become part of a training corpus, the damage is not remediable. You cannot retract what a model has learned.

OpenAI has since introduced ChatGPT Enterprise with contractual guarantees that enterprise data is not used for training. However, sovereignty assessment requires evaluating the system as broadly deployed, not the premium tier alone. The majority of ChatGPT usage globally — including by employees at enterprise-licensed organisations who access the consumer product on personal devices or through non-enterprise accounts — falls outside these protections. Data governance in a ChatGPT-integrated organisation is a patchwork of policies, technical controls, and trust. The TEE Method finds that trust insufficient as a control mechanism.

Part Four: Geopolitical Alignment, API Dependency, and the Single Point of Failure

ChatGPT is a US-based system, operated by a US-based company, hosted on US-based infrastructure, subject to US law, and aligned — explicitly and by design — with US geopolitical interests. OpenAI’s own policy documents state that the company “aligns with democratic values” and will not make its models available in jurisdictions that conflict with US national security interests. For institutions outside the United States, or for institutions whose interests diverge from US foreign policy, this creates a fundamental sovereignty risk. Your AI infrastructure is a political actor — and it does not represent you.

The practical manifestations of this are already visible. OpenAI has restricted access to its services in regions including China, Russia, Iran, and parts of the Middle East, and has stated it will comply with US export controls. As of 2025–2026, the Biden and Trump administrations’ successive AI executive orders and chip export controls have created an environment where access to frontier AI capabilities is increasingly tied to geopolitical alignment. If your country or institution falls on the wrong side of a future sanctions regime — or even a trade dispute — your access to ChatGPT could be restricted or terminated with minimal notice. SaaS agreements typically do not protect against geopolitical service interruption.

API dependency compounds this risk. For organisations that have built applications on OpenAI’s API — and there are thousands — the API is a single point of failure. OpenAI experienced 14 publicly documented service outages in 2025, three of which exceeded four hours. For a customer service chatbot or internal knowledge tool, a four-hour outage is an inconvenience. For a clinical decision support system, a financial trading assistant, or a national security analysis pipeline, it is a crisis. OpenAI’s enterprise SLA offers service credits as compensation — not a guarantee of availability, and certainly not the ability to run the model on alternative infrastructure if OpenAI’s cloud providers go down.

Model availability is a further concern. OpenAI maintains sole discretion over which models are offered, when they are deprecated, and what capabilities they have. GPT-4, still the most widely used model for enterprise applications as of late 2025, was partially superseded by GPT-4o, which trades some reasoning depth for speed. Organisations that built workflows optimised for GPT-4’s deliberative reasoning found their outputs changing — sometimes subtly, sometimes catastrophically — when the underlying model was swapped without a version bump. Even within a single model name, OpenAI uses dynamic routing to different model instances based on capacity and cost optimisation. Your consistent output is not guaranteed because your consistent model is not guaranteed.

Enterprise Controls: What Exists, What Is Missing

It would be inaccurate to suggest that OpenAI offers no enterprise controls. ChatGPT Enterprise includes data encryption at rest and in transit, SOC 2 Type II certification, and contractual commitments not to use enterprise data for training. The enterprise API offers granular usage controls, rate limiting, and usage dashboards. These are genuine improvements over the consumer product and represent meaningful steps toward institutional readiness.

However, from a sovereignty perspective, what is absent is more revealing than what is present. There is no on-premises deployment option — every query, every prompt, every output traverses OpenAI’s cloud infrastructure. There is no local model variant, no air-gapped deployment for classified or sensitive work. There is no guaranteed model version — enterprise customers receive the current production model, whatever it happens to be. There is no audit log of model changes that affect output behaviour. There is no public vulnerability disclosure program for model-specific security issues. There is no independent model evaluation board with institutional customer representation. And critically, there is no contractual exit mechanism that ensures data deletion across all OpenAI systems, including dark data in training pipelines, when the relationship ends.

For European organisations subject to GDPR, this last point is potentially fatal. GDPR Article 17 — the right to erasure — requires that data controllers ensure deletion across all systems, including backup and training datasets. OpenAI’s position that data once used in training cannot be removed from the model is in direct tension with this requirement. The tension remains unresolved as of mid-2026, with multiple European data protection authorities investigating, but no definitive regulatory ruling. Institutions adopting ChatGPT in this regulatory environment are assuming a risk that has not been litigated to conclusion.

Sovereignty Test Matrix: ChatGPT (GPT-4o / GPT-5)

Applying the TEE Method’s five-domain framework yields the following scores. Each domain is evaluated on a 1–5 scale where 1 represents critical sovereignty risk and 5 represents full sovereignty. The scores reflect ChatGPT as a deployed system in its mid-2026 state, considering both the enterprise and consumer tiers where relevant.

Total Sovereignty Score: 14/25 — Category: Limited Sovereignty

A score of 14/25 places ChatGPT in the “Limited Sovereignty” band of the TEE Method classification system. The system is powerful and useful — possibly the most capable AI system currently available — but it is structurally incapable of meeting the sovereignty requirements of critical institutional infrastructure. It can be used as a tool. It cannot be trusted as a foundation.

Red Flag Checklist

• Training data provenance: OpenAI cannot provide a verifiable inventory of training data sources. This is a material risk for IP and privacy compliance.
• Retroactive policy changes: Data usage terms have been changed after deployment without proactive customer notification. Trust once broken cannot be fully restored.
• No on-premises deployment: Every query traverses OpenAI-controlled infrastructure. For sensitive industries, this is a non-starter.
• No model versioning guarantee: The same model name can serve different underlying instances. Output consistency is not contractually assured.
• Single-provider API dependency: No fallback, no multi-provider routing, no portability layer. Outages are single points of failure.
• Geopolitical service restriction: Access can be terminated or restricted based on US foreign policy. Institutional AI infrastructure should not be a political instrument.
• GDPR erasure conflict: Open tension between the right to erasure and OpenAI’s claim that training data cannot be removed from models. Regulatory risk is unresolved.
• No independent audit: No external model evaluation board, no public red-teaming results, no verifiable safety claims. Self-assessment is not assessment.
• Prohibitive switching costs: Custom GPTs, fine-tuned models, and embedded workflows are structurally tied to OpenAI’s ecosystem. Exit is designed to be painful.
• Capability regression without recourse: Model updates can degrade performance on specific tasks with no rollback path. Your workflows are not in your control.

Action Plan: Using ChatGPT Sovereigntily

Based on this assessment, the TEE Method recommends the following actions for institutions that currently deploy — or are considering deploying — ChatGPT:

1. Conduct a deployment audit: Inventory every ChatGPT integration across your organisation. Document what data flows through each integration and what prompts are being used. You cannot govern what you have not catalogued.
2. Isolate sensitive workloads: Do not use ChatGPT — even the enterprise tier — for genuinely sensitive data. Classified information, health records, privileged legal communications, and trade secrets should be handled by systems with verifiable data isolation, not contractual promises.
3. Implement a multi-provider AI strategy: No institution should be dependent on a single AI provider. Invest in alternative models — open-weight models like Llama 3, Mistral, or Qwen; sovereign AI platforms; or self-hosted solutions — for different risk tiers of your workload portfolio.
4. Build an exit plan: Document exactly what it would take to migrate from ChatGPT to an alternative provider. Identify each custom integration, each fine-tuned model, each embedded workflow. Create a migration timeline. Update it quarterly. The best time to plan the exit is before you need it.
5. Negotiate contractual safeguards: Enterprise agreements with OpenAI should include: model version guarantees, 30-day advance notice of policy changes, the right to audit training data exclusion, contractual data deletion upon termination (with verification), and SLAs that compensate meaningfully — not with service credits — for capability regressions or outages.
6. Establish an AI governance board: Create a cross-functional body that oversees AI procurement, deployment, and risk assessment. This board should have the authority to halt AI deployments that fail sovereignty criteria. It should include legal, IT, security, and business representation — and a dissenting voice with no vested interest in the deployment.
7. Demand transparency: Join institutional calls for OpenAI and other AI providers to publish verifiable model cards, dataset manifests, and audit results. If your institution has leverage — through purchasing power or regulatory influence — use it. Sovereignty standards will not emerge from market forces alone.

The Sovereignty Verdict

ChatGPT is a remarkable achievement. It is also, from a sovereignty perspective, the most consequential AI system for institutions to evaluate — because it is the most embedded, the most capable, and the most structurally locking. OpenAI has built a system that is extraordinarily useful and extraordinarily difficult to leave. The two facts are not unrelated.

The TEE Method’s assessment of ChatGPT is not a recommendation against its use. Many institutions will — and should — deploy ChatGPT for appropriate use cases. But deployment must be intentional, informed, and bounded. The institution that adopts ChatGPT without a sovereignty assessment, without an exit plan, without multi-provider diversification, and without contractual safeguards is not adopting a tool. It is adopting a dependency. And dependencies, in the absence of sovereignty, become vulnerabilities.

This assessment was conducted using the TEE Method (Transparency, Ethics, Empowerment) framework, developed for the book SOVEREIGN: Who Owns the Future? The TEE Method evaluates AI systems across five domains — Strategic Alignment, Technical Performance, Ethical Compliance, Sovereignty Impact, and Cultural Alignment — scoring each from 1 (critical risk) to 5 (full sovereignty). Total scores classify systems as Full Sovereignty (21–25), Strong Sovereignty (16–20), Limited Sovereignty (11–15), or Critical Risk (5–10).

Assessment date: June 2026. AI systems and their terms of service evolve rapidly. This assessment reflects the state of ChatGPT (GPT-4o and GPT-5 families) as of the date above. Re-assessment is recommended quarterly for any AI system classified below Strong Sovereignty.

SOVEREIGN — Who Owns the Future? Your AI. Your Data. Your Choice.