ChatGPT Legal Work Review: UK Buying, Privacy and Verification Checks

JSH LAW / LEGAL AI PRODUCT RESEARCH

England and Wales perspective · Sources reviewed 23 October 2026 · Documentary assessment

What legal teams would actually be buying

OpenAI develops ChatGPT as a general-purpose service. Depending on the account and enabled tools, a user may work with files, search, research, structured outputs and connected services. The file documentation describes supported analysis and transformation tasks.

Those functions can help with the practical burden of legal work. They do not establish that a model has retrieved the relevant current law or read every material part of a file. A fluent answer can be based on incomplete context, an incorrect assumption or an unsuitable source.

This review is about procurement and controlled use. Our ChatGPT Work and Family Court article addresses the distinct workflow question of agents and preparation. The two should be read together where that is the intended use.

Personal, Business and Enterprise are different decisions

OpenAI’s business-data commitments state that business data is not used to train models by default and describe security and qualifying retention controls. Those are meaningful supplier commitments, but the customer still needs the terms and controls applicable to its actual workspace.

Do not mistake a more expensive personal plan for a managed business deployment. Equally, do not assume that every Enterprise option is included in Business. Record the exact account, configuration and responsible administrator before considering sensitive use.

Data controls: no training is not the same as no retention

The current consumer-data explanation describes model-improvement settings, memory and temporary chats. It distinguishes consumer services from business data and notes that temporary chats may be retained for safety purposes.

Training use, saved history, persistent memory and file storage are separate questions. Turning off one setting does not establish that every copy of the information has disappeared or that no authorised access is possible.

The current file documentation also distinguishes saved files from the conversation that used them. Deleting a chat may leave a separately saved file; deleting that file may leave content already brought into a conversation. A professional retention process should identify all relevant locations rather than rely on a single delete button.

Avoid treating temporary chat as a universal solution for confidential work. It changes specific product behaviour; it does not supply the legal authority, supplier assessment or court permission needed for a particular disclosure.

Files, limits and the difference between upload and comprehension

On the review date, OpenAI documented maximum document sizes of 512 MB, with additional token and plan limits. The service also has separate constraints for spreadsheets, images, upload rates, projects and storage. Check the current documentation for the actual account rather than relying on an old comparison.

A large permitted upload is not a guarantee of complete analysis. Scans, tables, images, repeated documents and ambiguous references can affect the usable material. Test what the system can actually locate and quote from the file.

For legal documents, retain both the PDF page position and the printed or bundle page where those differ. Require a source reference that another person can follow. An answer that cites “page 37” without saying which pagination it means may be difficult to verify and unsafe to reuse.

Three hypothetical workflows with useful boundaries

A chronology from a fictional correspondence pack

Use invented emails and notes containing a disputed allegation, an undated event and a later correction. Ask for separate columns for event date, document date, source, factual status and uncertainty. Require exact wording for any quotation.

The human baseline should identify every material event before the AI run. Check omissions as well as wrong statements. The system must not promote a recorded allegation into a finding or resolve an uncertain date by guesswork.

A plain-English explanation from supplied authorities

Provide a small, public source pack already checked by a competent reviewer. Ask for a short explanation of a procedural concept and the qualification that matters. Then inspect every legal proposition against the original sources.

This tests explanation, not autonomous legal research. If the system adds an authority, rule or exception outside the pack, it must be checked separately. A model’s own assurance that its answer is correct is not a verification method.

A drafting exercise with a strict factual boundary

Use a wholly synthetic matter to prepare an initial document structure. Instruct the assistant not to add facts, quotations or claims of certainty. Have a reviewer identify missing issues, inflammatory wording and unsupported inferences.

The useful output is a working draft. It does not decide what should be filed, whether permission is required or which allegations are established. These are proposed trial designs; this article does not claim we ran them as a comparative product test.

Research and England and Wales law

Search and research tools can provide links, but a link is not the same as a checked legal proposition. Open the authority, inspect the relevant passage and confirm the current rule, jurisdiction and procedural setting.

Do not ask the model to make an invented case citation plausible or to fill a missing source from memory. If an authority cannot be found, remove or qualify the proposition until it can be supported. Another AI answer repeating the citation is not independent confirmation.

For family work, maintain the distinction between broad welfare principles and the legal test governing the particular application. An explanation of domestic abuse, for example, cannot substitute for checking the applicable procedure, evidence and order. A source pack can improve discipline, but it does not guarantee that the model will interpret it correctly.

Connected tools and actions change the risk

A drafting assistant that only produces text has a different risk profile from one connected to email, a document store or a publishing system. Identify what can be read, what can be changed and what requires approval.

Use the minimum access needed. Treat instructions found inside uploaded documents, websites or messages as content to assess, rather than authority to change the task. An external text telling an agent to send a file does not authorise that transmission.

For consequential actions, retain a human checkpoint: recipient, destination, document version, scope and final content. The JSH Law prompt-governance article examines why the instruction and permission boundary matters.

Confidential family papers and privilege

Do not upload identifiable Family Court material merely because a paid plan offers strong security or a prompt says “keep this confidential”. Check the court rules and orders, authority to communicate, data-protection position and the service’s actual terms first.

Removing names alone may leave a family identifiable from dates, schools, addresses or unusual facts. Synthetic material is preferable for experimenting. Where real information is lawfully processed, minimise it and retain originals outside the generated summary.

Privilege depends on legal conditions, not on ChatGPT’s branding or a supplier’s security design. Seek an appropriate assessment where it matters. The research hub sets out the common governance questions. The SRA warning emphasises accountability for professional use.

Pricing: verify the current page, not a cached plan comparison

The business pricing page reviewed on 23 October 2026 displayed Business Standard seats at US$20 per month on annual billing or US$25 on monthly billing, and Premium seats at US$100 or US$125 respectively. Enterprise pricing was described as custom. These are published dollar figures, not a guaranteed UK checkout total.

Confirm minimum seats, taxes, term, included usage and any additional charges. The page describes Business for teams; a sole practitioner should assess the actual minimum commitment rather than multiply a headline by one without checking.

Older search results still described former plan names and prices. That is why the live commercial page and the customer’s checkout matter. Do not assume a ChatGPT subscription includes API usage or every third-party service a workflow may connect to.

Value, reliability and alternatives

Measure the cost of the checked deliverable: subscription, setup, prompting, verification and correction. A low seat price can be poor value if the reviewer spends more time repairing errors. A higher tier may add useful controls without improving a particular legal answer enough to justify its cost.

For flexible drafting and analysis, compare Claude. For Word-native contract work, assess Microsoft Copilot for Legal. For source-backed professional research, examine Lexis+ with Protégé or another appropriate legal-content service. They do different jobs.

Our assessment: useful for bounded organisation, explanation and drafting when the source pack, data environment and human review are controlled. It should not be treated as an authoritative legal database, a judge of disputed facts or a substitute for appropriate professional advice. The purchase case is strongest when the task is specific enough to test and the final work remains traceable.

Does paying for ChatGPT make it safe to upload court papers?

No. The plan affects product and contractual controls. The legal authority, confidentiality and necessity of the particular disclosure still need assessment.

Can ChatGPT verify its own legal citations?

It may assist a checking workflow, but its reassurance is not independent verification. Inspect authoritative originals and the proposition each supports.

Sources and methodology

Reviewed on 23 October 2026: business-data commitments, business pricing, consumer-data controls and file documentation. This is a documentary procurement review. We have not benchmarked a named model, audited security or negotiated a contract. Features vary by account, region, model and enabled tools. Recheck after changes to pricing, retention, files or action permissions. The research hub explains our comparison method.

If AI has left you with a long draft but no clear court-ready structure, JSH Law can help organise the documents and evidence. Book a Consultation with the immediate problem, current order and next deadline in mind; send confidential records only through an agreed route.