ChatGPT needs no introduction. Since its launch in late 2022, it has become the most widely used AI tool in the workplace, used by employees across every sector to draft communications, summarise documents, write code, answer questions, and accelerate almost every knowledge work task imaginable. According to OpenAI, ChatGPT now has over 100 million weekly active users. A significant proportion of them are using it at work, often without any formal guidance from their employers.
That last point is the problem. ChatGPT is a powerful, genuinely useful tool. OpenAI has invested heavily in making it capable and, increasingly, in making it enterprise-ready. But the default version of ChatGPT, the free and Plus consumer tiers that most employees are accessing, has data handling characteristics that most organisations have not considered and would not be comfortable with if they had.
This article sets out exactly what the risks are, what OpenAI's data policies actually say, what a high-profile data incident looked like in practice, and what your organisation should be doing about it.
What Is ChatGPT, and Who Makes It?
ChatGPT is a conversational AI product built on OpenAI's GPT series of large language models. OpenAI was founded in 2015 as a non-profit AI research organisation and has since become one of the most commercially significant AI companies in the world, with major investment from Microsoft.
ChatGPT is available in several tiers: a free version, ChatGPT Plus (a paid consumer subscription), and ChatGPT Enterprise, a separately contracted business product with different data handling terms. It is also accessible via OpenAI's API, which developers and organisations use to build integrations and custom applications.
The distinction between these tiers matters enormously from a security and compliance perspective, and it is a distinction that most employees and many IT teams are not making.
What Does OpenAI Actually Do With Your Data?
The answer differs significantly depending on which tier of ChatGPT your employees are using. Most are using the wrong one.
ChatGPT Free and Plus (Consumer Tiers)
By default, conversations on the consumer tiers of ChatGPT are used by OpenAI to train and improve its models. OpenAI's privacy policy states that it collects content users provide, including what is typed into the chat interface, and may use this content for model training purposes.
Users can opt out of having their conversations used for training, but this must be done manually in account settings, and the default is opt-in. The majority of employees using ChatGPT at work have never changed this setting. They are, by default, contributing whatever they type into ChatGPT to OpenAI's training dataset.
OpenAI also states that conversations may be reviewed by human trainers as part of its safety and quality processes. This means that a human employee of OpenAI could, in principle, read the content of a conversation, including any sensitive business information contained within it.
High risk scenario: An employee pastes a draft board presentation into ChatGPT to ask for feedback on the executive summary. That presentation (containing unreleased financial projections, strategic priorities, and commercially sensitive detail), may now be retained by OpenAI, used in model training, and potentially reviewed by a human operator. The employee intended to get a writing tip. The result may be a material data breach.
ChatGPT Enterprise and the API
ChatGPT Enterprise is a different product with different terms. Under the enterprise agreement, OpenAI explicitly commits to not using customer data to train its models. Data is encrypted in transit and at rest, conversations are not stored beyond the session by default, and organisations sign a data processing agreement that provides a contractual basis for data handling under UK GDPR.
Similarly, data submitted via the OpenAI API is not used for training by default, and is subject to API-specific data handling terms that are materially different from the consumer product.
If your organisation wants to use ChatGPT in any context involving non-public information, ChatGPT Enterprise or the API, not the consumer product, is the appropriate tool.
Key distinction: ChatGPT Free / Plus = conversations used for training by default, may be reviewed by humans, no DPA. ChatGPT Enterprise / API = no training on your data, DPA available, organisational controls. These are fundamentally different products from a compliance perspective.
The Samsung Incident: A Real-World Warning
In April 2023, Samsung became one of the first high-profile victims of corporate AI data exposure. Within weeks of allowing employees to use ChatGPT for work purposes, Samsung engineers had shared sensitive internal information with the tool on at least three separate occasions, including proprietary source code and confidential meeting notes.
Samsung subsequently banned the use of ChatGPT on company devices and began developing internal AI tooling. But the damage, the disclosure of proprietary technical information to a third-party AI provider, could not be undone.
Samsung is a technology company with sophisticated IT governance. If it happened to them within weeks of permitting ChatGPT use, it can happen to any organisation. The lesson is not that ChatGPT is malicious, it is not. The lesson is that employees will use AI tools fully and naturally, and without clear boundaries they will share things that should not be shared.
The Real Risks of Using ChatGPT at Work
1. Unsanctioned Data Disclosure
The Samsung incident is not an outlier, it is a preview of what happens at organisations that permit AI use without governance. Employees are not being reckless. They are being productive. When a tool helps them do their job better, they use it more, and they give it more context. That context (customer data, internal documents, financial information, personal records), ends up in a system their organisation has no control over.
The absence of visibility is as much a problem as the disclosure itself. Most organisations have no mechanism to detect that this is happening. There is no alert when an employee pastes a client contract into ChatGPT. There is no log. There is no audit trail. The data simply leaves the organisation silently.
2. UK GDPR and Data Protection Act 2018
Under UK GDPR, organisations must have a lawful basis for processing personal data and must ensure that any third-party processor they use offers adequate data protection guarantees, typically formalised through a data processing agreement.
The consumer tiers of ChatGPT do not provide a DPA. Using them to process personal data (customer names, employee records, contact details, health information, financial data), is almost certainly a breach of your data protection obligations. The ICO has been clear that AI tools are not exempt from data protection requirements, and that organisations are responsible for data shared with AI providers just as they are for data shared with any other processor.
A data subject access request, a regulatory investigation, or a client complaint could expose this gap very quickly.
3. Confidentiality and Professional Obligations
Many organisations, and virtually all professional services firms, operate under confidentiality obligations to clients. Sharing client information with a third-party AI provider without the client's knowledge or consent may constitute a breach of those obligations, regardless of whether the information is subsequently misused.
This is particularly acute in legal, financial services, healthcare, accountancy, and any sector where client confidentiality is both a professional and contractual requirement. Bar standards bodies, the FCA, the ICO, and sector-specific regulators have all begun issuing guidance on AI use, and enforcement is increasing.
4. Memory and Persistent Data Features
OpenAI has introduced a memory feature to ChatGPT, which allows the model to remember information about users across conversations. This feature is enabled by default for many users. It means that information shared in one conversation, including sensitive professional context, may be recalled and referenced in future conversations, including those initiated by the same user in a different context.
For organisations, this creates an additional layer of risk: sensitive information shared in one session may persist and resurface in ways that are difficult to predict or control.
Memory feature risk: An employee mentions their employer's name, role, and client sector while asking ChatGPT for help drafting a proposal. ChatGPT stores this in memory. In a future session, the same employee asks an unrelated question, and ChatGPT references the stored professional context in its response. The employee is now interacting with a system that has accumulated a profile of their professional life, stored on OpenAI's servers, with no organisational oversight.
5. Plugin and Integration Risks
ChatGPT supports a growing ecosystem of plugins and integrations that allow the model to interact with external services, browsing the web, accessing files, connecting to third-party platforms. Each integration extends the attack surface and the data exposure risk. Organisations that allow employees to use ChatGPT with plugins enabled have significantly less visibility into what data is being accessed, transmitted, or processed.
6. Hallucination and Over-Reliance
As with all large language models, ChatGPT can produce confident, well-structured responses that are factually wrong. This is particularly dangerous in professional contexts, legal advice, compliance guidance, financial analysis, medical information, technical documentation. ChatGPT does not flag uncertainty reliably. It will produce a coherent, authoritative-sounding answer even when the underlying information is incorrect or outdated.
Organisations that allow employees to act on ChatGPT's output without independent verification are accepting risk that may materialise as errors, liability, or reputational damage.
Risk Summary
| Risk | Applies To | Level |
|---|---|---|
| Consumer ChatGPT used with personal or client data, no DPA | All organisations | High |
| No AI usage policy, employees using ChatGPT unsupervised | All organisations | High |
| Confidential client data shared via consumer ChatGPT | Legal, finance, professional services | High |
| Memory feature enabled, persistent professional context stored | All consumer tier users | Medium |
| Plugin and integration use expanding data exposure | Organisations allowing plugins | Medium |
| Hallucination risk, over-reliance on unverified output | All organisations | Medium |
| ChatGPT Enterprise or API with DPA, controlled deployment | Organisations on enterprise plan | Low |
How to Use ChatGPT Safely in Your Organisation
ChatGPT can absolutely be used safely in a business context. The risks above are real but manageable. Here is what responsible deployment looks like.
1. Establish a Written AI Usage Policy
This is the single most important action any organisation can take. A clear, written AI usage policy sets out which tools are approved, what data must never be entered into any AI tool, how AI-generated output should be verified, and how to report concerns. Without a policy, employees have no guidance and organisations have no defensible position.
2. Move to ChatGPT Enterprise or the API
If ChatGPT is being used for anything beyond publicly available information, your organisation needs to be on an enterprise agreement or accessing the model via the API with a DPA in place. The consumer product is not appropriate for business use involving non-public data, full stop.
3. Consider Microsoft Copilot as an Alternative
For organisations already running Microsoft 365, Microsoft Copilot, which is built on OpenAI's models but deployed within your Microsoft 365 tenancy under your existing data governance controls, may be a more appropriate path than standalone ChatGPT. Your data stays within your Microsoft environment, your existing DPA with Microsoft applies, and your IT team has administrative oversight. It is not a perfect solution, but it is a materially safer default than consumer ChatGPT for M365 users.
4. Disable the Memory Feature for Business Accounts
If employees are using ChatGPT with any work-related context, the memory feature should be disabled. This can be done in ChatGPT settings and should be part of your standard onboarding guidance for any approved AI tool use.
5. Classify Your Data
Data classification, categorising information by sensitivity, is the foundational control that makes AI governance possible. If employees know what constitutes confidential data and understand that it must not be shared with external AI tools, you have significantly reduced your exposure. Without data classification, any AI usage policy is difficult to enforce in practice.
6. Train Your Staff
Technology controls are only as effective as the people operating them. A short, practical briefing on AI risks (what these tools are, what they do with data, and what must never be shared), should be part of your existing security awareness programme. It does not need to be complex. It needs to be clear.
The short version: ChatGPT is safe for business use if you are on an enterprise agreement or using the API, have a DPA in place, have a written AI usage policy, have disabled the memory feature, have classified your data, and have trained your staff. If none of that is in place, and for most UK firms it isn't, you are running an unmanaged risk right now.
What Should You Do Right Now?
If this article has surfaced gaps in your organisation's AI governance, you are not alone. Most UK businesses have not yet caught up with the pace at which their employees are adopting AI tools. The steps to close the gap are straightforward:
- Find out what your employees are actually using. Conduct a quick audit, a simple survey or a conversation with team leads, to understand which AI tools are in use and on which tier.
- Issue interim guidance immediately. Tell staff not to enter client data, personal data, or confidential business information into any AI tool until a formal policy is in place. This costs nothing and reduces your exposure immediately.
- Check your DPA position. For any AI tool your organisation uses commercially, verify whether you have a DPA in place. If you don't, you are likely processing data without adequate safeguards.
- Get a formal AI policy written. This should be drafted with input from both cybersecurity and data protection expertise. It does not need to be lengthy, it needs to be clear, practical, and communicated to staff.
- Review your broader AI posture. ChatGPT is the most visible AI tool, but it is unlikely to be the only one in use. Grammarly, Notion AI, Copilot, Gemini, and dozens of other tools all have data handling implications that warrant review.
Not sure if your organisation is using ChatGPT safely?
KAVIX works with London businesses to audit AI tool usage, identify data protection gaps, and put the right policies and controls in place. If you'd like a free conversation about your current AI posture (including ChatGPT, Copilot, and any other tools your team is using), get in touch.
Talk to Us About AI SecuritySummary
ChatGPT is a genuinely useful business tool, and OpenAI has made real progress in building enterprise-ready products with appropriate data protection controls. The problem is the gap between the product that is available and the product that most employees are actually using.
The consumer version of ChatGPT, which is what the vast majority of employees are accessing, processes data in ways that are incompatible with most organisations' data protection obligations, client confidentiality commitments, and basic security posture. The Samsung incident was a public and expensive lesson. Most organisations that have had similar incidents simply haven't publicised them.
The good news is that the fix is not technically complex. It requires policy, governance, and a clear conversation with your staff. The organisations that get this right now will be significantly better positioned as AI regulation tightens and enforcement increases.