Should We Ban ChatGPT, or Is That Pointless?

This page answers the questions UK executives and IT leads ask after a board member, an auditor or an incident forces the AI tool question. It covers when a ban is a rational control, why most bans move usage out of sight and what a declaration-based alternative looks like.

Should we ban ChatGPT at work?

Usually no, although a short dated ban can be the right call while you establish control. A permanent blanket ban rarely reduces AI use in a knowledge-work organisation. It moves that use onto personal devices and personal accounts, where you have no logging and no evidence for an auditor.

The decision most boards think they are making is allow or block. The decision they are actually making is visible or invisible. Both choices carry risk, but only one leaves you with a record of which client data touched which model and only one gives your security team something to assess. That is why we treat the ban question as a control-design question rather than a policy-tone question.

Does banning ChatGPT actually work?

Partially, and only inside a boundary that keeps shrinking. Egress blocking and browser policy will stop the named tool on a managed device on the corporate network. None of it stops the phone in someone’s pocket, the personal laptop at home, a browser extension or the AI feature quietly added to a platform you already licence.

So the observable effect of a blanket block is a drop in visible usage. Whether actual usage falls depends on why people started. Where AI use is driven by deadline pressure and volume, in bid teams, client reporting, first-draft legal and marketing copy, the incentive to keep using it survives the block. This is the mechanism behind Shadow AI: the tool disappears from your telemetry and stays in the workflow.

When is a ban the right decision?

Four situations warrant blocking first.

• Active incident containment, where you suspect data has already left.
• A specific tool whose terms permit training on inputs, when you process client-confidential or special category data.
• A contractual restriction you cannot interpret without legal advice.
• Regulated processing where no data protection impact assessment exists yet.

In each case the ban should be narrow, named and dated. Block the specific service rather than the category, state the review date in the notice, name the sanctioned alternative and name the person who can approve an exception. A ban with a defined end and a defined route through it is a control. A ban with neither is a statement of intent that your staff will quietly route around.

What does a blanket ban cost us that the board does not see?

Three things. You lose the inventory: nobody will tell you what they were using once telling you is an admission of breach. You lose the questions: staff who would have asked whether a tool is safe for a particular dataset stop asking, because the answer is predetermined. You lose the trail, so when a client asks whether their material was processed by a third-party model, your honest answer is that you cannot say.

The selection effect compounds all three. The people most likely to keep using a blocked tool are the ones under the most delivery pressure.

What should we do instead of banning AI tools?

Buy the information first. Our AI Amnesty framework is a time-boxed, no-blame window in which staff declare which AI tools they use, on what data and for which task, with a written commitment that declarations made inside the window carry no disciplinary consequence.

The amnesty runs alongside technical Shadow AI Discovery, because self-declaration alone under-reports and discovery alone misses the context. Together they produce something a ban cannot: a dated inventory of real usage, with named owners and identified data flows. You then decide with evidence which tools to approve, which to restrict and which to replace.

How does an AI Amnesty work in practice?

Four stages, typically across six to eight weeks:

1. Sponsor statement. A named executive, not the security team, states the no-blame terms and the window dates. Without this the declaration rate collapses.
2. Declaration window. Two to three weeks, one short form per tool: what it is, what data goes in, what the output is used for, whether a personal account is involved.
3. Corroboration. Technical discovery across identity and SaaS telemetry finds what was not declared, sized as a gap rather than used to identify individuals.
4. Triage and control. Each tool is approved, restricted to defined data classes or replaced with a sanctioned equivalent, and the policy is rewritten to match the work people actually do.

The deliverable is an inventory plus a control decision per tool, which is the same artefact your assurance obligations require. The evidence sources behind stage three sit in how to find out which AI tools employees use.

Does an amnesty not reward people who broke policy?

It reprices information that you cannot obtain any other way. Near-miss reporting works on the same logic: you accept that you will hear about mistakes you could have punished, because hearing about them early is worth more than the punishment.

The amnesty covers declaration, not conduct. Deliberate exfiltration, use of client data in breach of a contractual restriction after being told or falsifying a declaration all remain disciplinary matters, and we set that boundary out in writing before the window opens. The volume of declarations is itself the finding. When forty people declare the same unsanctioned tool, you have a process problem to solve rather than forty people to discipline.

What do we tell our auditor or regulator?

Tell them what you know, and be able to show how you know it. ISO 42001 expects an organisation to know which AI systems it operates, who owns them and what risk each carries, and its Annex A controls are the level at which an auditor tests that. EU AI Act obligations may also reach UK organisations through supply chains, typically where a system, its output or its deployment touches the EU, and those obligations assume you can identify the systems you deploy. An amnesty plus discovery produces exactly that, with dates.

A blanket ban produces a policy document and an unverified assumption of compliance. We would rather your inventory answered the question first.

Related reading: should I be worried about employees using ChatGPT at work and our AI Security Gap Analysis, which is where most amnesty findings are turned into a prioritised control plan.

This page is general information on AI Security practice, not legal or compliance advice. Confirm how ISO 42001, the EU AI Act and data protection obligations apply to your own circumstances with your legal and compliance advisers.

Reviewed by Jason Holloway, Director, QL Security, May 2026.