AI Liability in UK Legal Practice: SRA Duties, Court Evidence and Data Protection
JSH LAW / ARTIFICIAL INTELLIGENCE / LEGAL ACCOUNTABILITY
England & Wales · Legal and regulatory sources checked: 15 October 2026
AI errors are not just technical faults. In legal practice, they can misstate the law, distort evidence, expose confidential information or trigger unauthorised actions.
Whether an error gives rise to legal liability depends on the duties owed and the facts. This technical guide examines English law, SRA regulation, decided cases, data protection and the costs of AI-assisted legal work.
A discussion of artificial general intelligence, AI expenditure and the future of custom GPTs at MAICON 2026, featured in episode 246 of The Artificial Intelligence Show, provides a useful starting point for examining the legal consequences of this transition.
Who is responsible when an AI agent makes a mistake? What happens when confidential evidence enters an external AI system? Can AI-generated witness evidence be trusted? And should clients receive the financial benefit when artificial intelligence makes professional work faster?
These are not merely technological questions. They engage established principles of English law, professional regulation, procedural fairness, data protection and access to justice.
For the separate question of how organisations should control what an AI agent is authorised to do, see our earlier guide to agentic AI, delegated authority and meaningful human supervision. This article focuses instead on legal liability, court decisions and regulatory duties.
1. From Generative AI to Autonomous Agents
Three concepts are frequently confused.
Generative AI produces outputs such as written text, images, code and summaries in response to instructions.
Custom GPTs are configurations of underlying AI models, incorporating instructions, reference materials and selected capabilities for particular tasks.
Agentic AI describes systems capable of pursuing objectives through multiple steps, potentially using connected applications, retrieving information and undertaking actions.
Artificial general intelligence, or AGI, is a different and contested concept generally associated with broad capabilities across intellectual tasks. There is no universally agreed legal threshold at which a system becomes AGI.
The critical legal distinction is between producing information and taking action.
A practical legal example
A solicitor uses AI to summarise a court bundle. The resulting summary can be reviewed before any professional decision is taken.
Now consider an AI agent that searches the bundle, identifies allegedly missing evidence, prepares correspondence and sends a disclosure request automatically. The system could misinterpret an order, request privileged information, communicate with the wrong recipient or take an unauthorised procedural step.
The second system creates additional legal and operational risks because it is no longer simply generating text for review.
The closer AI moves towards decisions and external actions, the more important it becomes to establish permissions, accountability and effective human oversight.
2. Who Is Legally Responsible When an AI Agent Makes a Mistake?
Under the current law of England and Wales, an AI system does not have independent legal personality simply because it operates autonomously.
It cannot ordinarily be treated as a solicitor, contracting legal person or independently accountable professional.
That does not mean responsibility necessarily rests with one party. Depending upon the circumstances, liability may arise through contract, negligence, statutory duties, professional regulation or data protection legislation.
| Participant | Potential Legal Responsibility |
|---|---|
| Solicitor | Competence, professional judgment, supervision, confidentiality and duties to the court. |
| Law firm | Governance, risk management, supervision and regulatory compliance. |
| AI provider | Contractual obligations, data protection responsibilities and other applicable legal duties. |
| Deploying business | Authorisation, system configuration, monitoring and consequences of attributable actions. |
| Individual user | Their own instructions, decisions, representations and applicable duties. |
These responsibilities may overlap.
A business cannot simply assume that the supplier of an AI system bears every loss arising from its deployment. Equally, a supplier’s contractual disclaimer will not necessarily defeat every possible claim.
The law of agency and electronic actions
Traditional agency law provides useful concepts, including authority and attribution. But an AI system is not necessarily an agent in the conventional common-law sense.
Where an AI system communicates contractual acceptance, sends instructions or initiates transactions, the legal analysis must consider the parties’ agreements, the authority granted and how the relevant electronic act is attributed.
The central issue is not whether AI acted independently in a technological sense, but which legal duties and consequences attach to the humans and organisations involved.
3. Solicitors Cannot Delegate Professional Responsibility to AI
On 17 August 2026, the Solicitors Regulation Authority published a formal warning notice concerning the misuse of artificial intelligence.
The warning identifies two particular concerns: inaccurate AI-generated legal material and the improper disclosure of confidential client information.
The SRA Code of Conduct for Solicitors imposes relevant obligations including:
- Paragraph 1.4: not misleading clients, the court or others.
- Paragraph 2.2: not influencing the substance of evidence.
- Paragraph 2.4: advancing only properly arguable assertions and representations.
- Paragraph 3.2: providing competent service.
- Paragraph 3.5: accountability and effective supervision.
- Paragraph 6.3: protecting client confidentiality.
- Paragraph 8.7: providing clients with appropriate information about pricing and costs.
SRA-regulated firms must also maintain effective governance arrangements, systems and controls.
A solicitor who delegates research to an AI system remains responsible for checking legal authorities and conclusions. A firm using AI-generated procedural correspondence must maintain suitable review and approval arrangements.
The SRA’s regulatory obligations apply to those it regulates. They should not be presented as though every litigant in person or non-SRA-authorised support service is directly subject to the SRA Code.
4. What the Courts Have Already Said About AI-Generated Legal Material
The consequences of presenting inaccurate AI-generated information to the court are no longer hypothetical.
Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin)
The Divisional Court considered serious issues arising from false authorities being presented to the court.
The judgment addressed the professional obligations of lawyers relying upon legal material that had not been adequately verified.
It also considered the court’s Hamid jurisdiction, which concerns the court’s powers to address professional failures affecting the administration of justice.
A lawyer presenting a legal proposition must take responsibility for its accuracy, irrespective of the technology used to prepare it.
Read the Ayinde and Al-Haroun judgment
Cork and another v Smith [2026] EWHC 1199 (Ch)
This May 2026 High Court judgment provides an important illustration of the dangers of uncritical reliance on artificial intelligence.
A junior solicitor used AI when researching an issue concerning insolvency legislation. Correspondence subsequently included purported statutory material that did not exist.
The court examined the failure to verify the original material, the response provided when the error was queried and the relevant supervisory arrangements.
The judgment reinforces the importance of checking legal sources and responding candidly when an error is identified.
The court’s examination of the conduct should not be confused with final disciplinary findings against every individual involved.
Re A, B, C, D (Extension of assessment; Use of AI: hallucinations) [2026] EWFC 71 (B)
The Family Court also confronted inaccurate AI-generated legal material in proceedings concerning children.
The case arose from proceedings involving BCP Council and included concerns about false legal authorities appearing in submissions prepared by a third party.
It demonstrates why inaccurate legal material can be especially consequential in family litigation, where the decisions under consideration may affect the welfare and future arrangements of children.
Read the Family Court judgment (PDF)
What these cases establish in practice
The problem is not the mere use of artificial intelligence. The problem is presenting unreliable legal or factual material as though it has been properly checked and can safely be relied upon.
5. AI and the Integrity of Witness Evidence
There is an important distinction between using artificial intelligence to organise evidence and allowing it to generate or change the substance of a witness’s account.
A witness statement should accurately reflect the witness’s own evidence.
AI may help with transcription, identifying dates and preparing documents. However, a system designed to improve fluency or persuasion may introduce greater certainty, alter emphasis, combine separate events or incorporate facts that the witness cannot independently confirm.
These changes may affect the evidential reliability of the statement and the witness’s ability to explain the account during cross-examination.
The Civil Justice Council’s 2026 Consultation
In June 2026, the Civil Justice Council published an update on its consultation concerning AI use in preparing court documents.
Respondents broadly considered that existing professional responsibility frameworks were generally sufficient for lawyers preparing pleadings, skeleton arguments and similar documents.
Witness statements, expert evidence and litigants in person were identified as requiring further consideration.
The Council recognised concerns that AI could reshape or embellish personal recollections without making those alterations immediately apparent.
The proposals under consideration should not be confused with a universal new requirement to disclose every use of AI in preparing a court document.
Preserving Authentic Evidence
As a practical safeguard, a reliable drafting record should identify:
- The witness’s original account.
- The relevant source documents.
- Material alterations made during drafting.
- Any substantive content suggested or generated by AI.
- The witness’s confirmation that the final statement is accurate.
These are practical evidential safeguards, not a statutory five-stage legal test.
For related practical guidance, see the JSH Law article on police disclosure and evidence in Family Court proceedings .
6. Confidentiality, Legal Professional Privilege and External AI Systems
The confidentiality problem may be more significant than the risk of inaccurate legal research.
Legal files frequently contain highly sensitive information, including medical records, financial disclosures, children’s information, domestic abuse allegations, confidential correspondence and privileged advice.
Uploading such information to an AI platform is not legally insignificant simply because the transfer takes place through a familiar chat interface.
UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC)
The Upper Tribunal raised concerns about client information being entered into open AI systems and the potential consequences for confidentiality and legal professional privilege.
The SRA subsequently highlighted the case in its August 2026 warning notice.
However, it would be inaccurate to assert that every use of an external AI platform automatically destroys legal professional privilege.
That question requires examination of the material concerned, the purpose and nature of any disclosure, the contractual and technical arrangements, and whether the necessary confidentiality has been preserved.
Client confidentiality, legal professional privilege and data protection are related but legally distinct concepts.
What Organisations Should Check
- Whether information is retained or used for model training.
- Who can access submitted files and generated outputs.
- What contractual confidentiality protections apply.
- Whether international data transfers occur.
- What subprocessors are involved.
- How information is deleted or retrieved.
- Whether appropriate security and audit arrangements exist.
A paid subscription does not, by itself, establish that a platform is suitable for confidential legal work.
7. UK GDPR, Automated Decisions and the Data (Use and Access) Act 2025
Where an organisation processes personal data through AI, the UK GDPR and Data Protection Act 2018 remain relevant.
Important provisions include:
- Article 5: data protection principles.
- Article 6: lawful bases for processing.
- Article 9: additional conditions for special-category data.
- Article 28: controller-processor arrangements.
- Article 32: security of processing.
- Article 35: data protection impact assessments for processing likely to result in high risk.
The Data (Use and Access) Act 2025 changed the legal framework governing certain significant decisions made solely through automated processing.
The reforms permit a wider range of automated decisions subject to applicable conditions and safeguards, with particular restrictions and requirements where special-category personal data is involved.
Relevant safeguards concern information about significant decisions, representations, human intervention and the ability to contest outcomes.
This matters where AI systems are used to assess eligibility, evaluate individuals or make decisions producing legal or similarly significant effects.
Not every use of AI requires a data protection impact assessment. The obligation depends upon whether the particular processing is likely to result in high risk to individuals.
Meaningful human oversight matters. Placing an approval button at the end of an automated process does not necessarily demonstrate that a human has genuinely evaluated the decision.
The Information Commissioner’s Office publishes guidance on artificial intelligence and data protection . Some of its earlier guidance remains under review following the 2025 reforms.
8. The Judiciary’s October 2026 Position: AI May Assist, but Judges Must Decide
On 9 October 2026, the judiciary of England and Wales published a position statement on the judicial use of artificial intelligence.
It adopts a cautiously positive approach, recognising that AI may improve the administration of justice while preserving the essential role of human judicial decision-making.
Judicial independence, fairness, security and public confidence remain central.
AI may assist with appropriate supporting tasks, but judicial reasoning, the evaluation of evidence and the exercise of judicial discretion must remain with judicial office holders.
The statement addresses the governance and safeguards applying to judicial use. It is not a universal technical specification for every commercial AI system.
Nevertheless, it illustrates an important principle for responsible legal technology:
9. AI Costs: Should Clients Receive the Financial Benefit of Automation?
Artificial intelligence is changing the economics of professional services.
Tasks previously requiring several hours may, in some circumstances, be completed considerably faster using AI-assisted document review, drafting or information extraction.
That creates an important question: how should clients be charged?
The answer depends upon the retainer, the nature of the work and any applicable regulatory or costs-assessment requirements.
Fixed Fees
A properly agreed fixed fee does not automatically become unlawful merely because technology improves efficiency.
A fee may legitimately reflect expertise, professional responsibility, the complexity of the service, technology costs and the work undertaken.
Hourly Billing
Where a client is charged on a time basis, the records must accurately reflect the chargeable work undertaken under the retainer.
Time that was not actually worked must not be represented as though it had been expended.
Solicitors remain subject to applicable pricing and costs-information obligations.
Recoverable Litigation Costs
Where costs are recoverable from another party in civil proceedings, CPR Part 44 may engage questions of reasonableness and proportionality.
These issues must be distinguished from the contractual fee payable by a client to their own adviser.
AI efficiency may also involve additional expenditure on verification, security, supervision and professional quality assurance.
There is no defensible universal rule that AI-assisted services must become cheaper by a fixed percentage.
The proper focus is accurate charging, transparent scope and a defensible relationship between the work undertaken and the fee claimed.
10. The End of Custom GPTs and the Problem of Platform Dependence
OpenAI has announced plans to retire custom GPTs and migrate relevant workflows towards plugins.
Its published guidance identifies 11 December 2026 as the standard scheduled retirement date, subject to relevant account or workspace arrangements.
This does not mean that GPT models themselves are disappearing. It concerns the future of a particular reusable assistant format.
However, the announcement illustrates a wider commercial and legal issue.
Businesses may invest significant time developing specialist AI workflows containing instructions, reference documents, integrations and internal business procedures.
Their continued operation may depend on decisions taken by an external technology provider.
The Business-Continuity Questions
- Who controls the underlying information and reference files?
- Can the business export or reconstruct its workflows?
- Will essential integrations continue functioning?
- How will confidential information be retained or deleted?
- Can previous outputs and decisions be audited?
- What happens if a supplier changes its product or pricing?
The OpenAI migration guidance also notes that certain custom actions do not transfer automatically and may need rebuilding.
For legal organisations, a platform change should not result in the loss of records needed to establish how work was undertaken or advice reached.
11. The JSH Law Six-Question Check for AI-Assisted Legal Work
Before relying on an AI-generated output or permitting an AI agent to undertake a legal task, six questions provide a practical starting point.
This is a practical governance framework, not a statutory legal test.
-
Authority — Who authorised the task?
Was the system permitted to undertake the particular activity, and were limits placed on consequential actions?
-
Confidentiality — What information was disclosed?
Did the task involve personal, privileged or confidential information, and were appropriate safeguards in place?
-
Verification — Can the output be checked?
Are legal authorities genuine, quotations accurate and factual assertions supported by identifiable evidence?
-
Accountability — Who is responsible?
Is there an identifiable person with the competence and authority to review and approve the work?
-
Evidential Integrity — Has the evidence changed?
Has AI altered a witness’s account, introduced assumptions or confused allegations with established findings?
-
Auditability — Can the process be explained?
Are there appropriate records of instructions, source material, generated outputs, checks and approvals?
The questions do not replace professional duties or applicable legal requirements.
They provide a structured approach to identifying risks before AI-generated work is relied upon.
12. Frequently Asked Questions
Is it illegal to use ChatGPT to prepare court documents?
There is no general rule making every instance of AI-assisted court-document preparation unlawful. However, the document must satisfy the applicable procedural and evidential requirements. Legal authorities must be checked, evidence must be authentic and confidentiality restrictions must be respected.
Can I use AI to help write my witness statement?
AI may assist with administrative drafting tasks, but the final statement must genuinely reflect your own evidence. AI should not invent facts, alter recollections or introduce certainty that you cannot support.
Is an AI provider automatically liable when an agent makes a mistake?
No. The answer depends upon the contractual arrangements, the nature of the failure, the duties owed and whether the conditions for legal liability are established.
Does uploading information to AI automatically waive privilege?
Not necessarily. The answer depends upon the nature of the information, disclosure and confidentiality arrangements. However, using inadequately protected external services can create serious risks to both confidentiality and legal professional privilege.
Are all GPT-based AI models disappearing in December 2026?
No. OpenAI’s announced retirement concerns its custom GPT product format and related migration arrangements. It does not mean the underlying GPT models are being discontinued.
Conclusion: The Future of AI Depends on Accountability
Artificial intelligence will continue changing legal practice.
Some processes will become faster. Others will become more complex as AI systems acquire the ability to interact with software, process larger quantities of information and undertake increasingly sophisticated tasks.
But technological advancement does not remove the fundamental requirements of professional competence, reliable evidence, confidentiality, fairness and accountability.
The decisions in Ayinde and Cork, the SRA’s 2026 warning notice, the Civil Justice Council’s work and the judiciary’s October 2026 position statement all point towards a developing legal response.
It is neither a wholesale rejection of artificial intelligence nor an unconditional acceptance of automation.
It is an insistence that technological capability operates within an effective framework of responsibility.
English law has long dealt with the consequences of people acting on behalf of others, the reliability of evidence and the duties owed by professionals entrusted with significant responsibilities.
AI introduces new forms of technological delegation. It does not make those longstanding legal concerns obsolete.
If anything, it makes them more urgent.
The central challenge is not whether machines will become capable of carrying out legal work. It is whether the organisations relying upon them can demonstrate that the resulting work is lawful, reliable, secure and properly accountable.
Related JSH Law Resources
Primary Sources and Further Reading
- SRA — Misuse of AI: Warning Notice, 17 August 2026
- SRA Code of Conduct for Solicitors
- Ayinde; Al-Haroun [2025] EWHC 1383 (Admin)
- Cork and another v Smith [2026] EWHC 1199 (Ch)
- Re A, B, C, D [2026] EWFC 71 (B)
- UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC)
- Civil Justice Council — Use of AI in Preparing Court Documents
- Judiciary — Position Statement on Judicial Use of AI, 9 October 2026
- Information Commissioner’s Office — AI and Data Protection Guidance
- OpenAI — Custom GPT Retirement and Migration FAQ
- The Artificial Intelligence Show — Episode 246, MAICON 2026
For further analysis of AI-related professional duties, evidence and legal accountability, explore our Artificial Intelligence and Family Justice resource centre.
© 2026 JSH Law Ltd. All rights reserved.
Leave a Reply
Want to join the discussion?Feel free to contribute!