AI Reliance, CPS practice and public law risk

Overview

In the case of Tobosaru v Court of Law Craiova, Romania [2026] EWHC 1720 (Admin), the High Court upheld extradition orders for two Romanian nationals. While grounded in established public law principles, the case has attracted attention for a more contemporary concern: the use of AI-assisted outputs and the submission of non-existent legal authorities by the Crown Prosecution Service, commonly described as “AI hallucinations”.

A notable feature of the case was the scrutiny applied to how material relied upon by the CPS had been generated, summarised or analysed, including concerns raised in the proceedings about insufficient interrogation of AI-assisted content forming part of the evidential picture.

The CPS accepted that generative AI was the likely source of the fictitious authorities cited in the extradition appeal. However, it maintained that the true cause of the error was not the technology itself, but the reviewing lawyer's failure to verify the accuracy of the citations before the material was filed and served.

The Court’s Core Findings

The Administrative Court reaffirmed two practical principles:

• Decision-making responsibility is non-delegable
  The court emphasised that neither foreign judicial reasoning nor internally generated summaries, however produced, can displace the obligation on UK authorities to exercise independent judgment.
• Fairness requires intelligible reasoning
  Where an affected party cannot properly understand the basis of the decision, or where reasoning appears formulaic, compressed or inadequately explained, procedural fairness may be compromised.

These principles framed the court’s approach to the material said to underpin the decision.

CPS Reliance on AI-Assisted Material

A key practical issue arising from the case was the treatment of material relied upon by the CPS, where it was suggested that:

• Elements of the evidential summary or analysis may have been produced or streamlined using AI tools; and
• That output was relied upon without sufficiently testing its accuracy against primary source material.

The court’s concern was not the use of technology itself, but the absence of demonstrable critical engagement. In particular:

• There was limited evidence of line-by-line verification against underlying documents.
• Certain conclusions appeared over-simplified or asserted without clear evidential linkage.
• The reasoning lacked transparency as to how conclusions had been reached.

This created a risk that the CPS had, in effect, adopted a generated narrative rather than independently constructing one.

Why This Matters: AI and Automation Bias

The case highlights a practical risk increasingly seen across public bodies:

• Automation bias: A tendency to treat AI-generated outputs as inherently reliable or authoritative.
• False coherence: AI summaries can appear structured and persuasive while omitting nuance or contrary evidence.
• Traceability gaps: Without clear audit trails, it becomes difficult to demonstrate the reasoning behind a conclusion.

In Tobosaru, the issue was not simply technical, it went directly to whether the decision-making process met public law standards.

Implications for AI Users

• Verification is Essential

Where AI tools are used to:

• Summarise evidence
• Identify patterns
• Draft analytical material

there must be systematic verification against primary sources. This cannot be assumed or inferred; it must be demonstrable.

• Maintain Evidential Ownership

Avoid adopting AI outputs as a substitute for reasoning. The final analysis must be clearly attributable to the decision-maker, not the tool.

• Ensure Explainability

Users should be able to demonstrate to the Court:

• How was the material produced?
• What checks were applied?
• Where in the primary evidence do conclusions derive from?

Failure to do so risks decisions appearing opaque or mechanistic.

• Document the Process

A defensible record should include:

• Whether AI tools were used
• The scope of their use
• The validation steps undertaken

This is particularly important in cross-border cases where courts must assess the reliability of external material.

• Training and Guidance 

The case underscores the need for:

• Clear internal AI usage policies
• Training to identify and mitigate automation bias
• Defined standards for evidential checking and sign-off

What This Means

Tobosaru illustrates that the legal risk lies not in using AI, but in relying on it uncritically. For the CPS and other public bodies, the practical lessons are clear:

• Interrogate every output especially where it influences evidential conclusions.
• Anchor reasoning in primary material, not generated summaries.
• Be transparent about process, particularly where technology plays a role.

The decision signals that courts will expect visible, human-led reasoning, even in increasingly digital workflows. Where authorities cannot demonstrate that AI-assisted material has been properly interrogated, they risk findings of procedural unfairness or irrationality.

The priority is to move from informal or ad hoc AI use to structured, auditable and accountable practices ensuring that technology enhances, rather than undermines, the integrity of decision-making.

More broadly, the decision underscores the principle that accountability for prosecutorial and public authority decision-making cannot be delegated to technology. AI may assist in the preparation of legal work, but responsibility for the final product remains firmly with the human decision-maker who adopts and relies upon it.