1) Introduction
In R. v FGD [2026] EWCA Crim 918, the Court of Appeal (Criminal Division)
(Lady Justice May, Mr Justice Griffiths and His Honour Judge Picton) considered a contemporary problem:
a complainant in a rape trial had used AI tools to “prepare” for cross-examination, generating documents that
looked like structured witness coaching (a summary of their recollection plus anticipated defence questions with
“suggested answers”).
Mid-trial, after these AI-generated materials were found on a phone download ordered for disclosure purposes,
the trial judge stayed the proceedings as an abuse of process on the basis that a fair trial had become impossible.
The Crown appealed under s.58 Criminal Justice Act 2003.
The central issues were: (i) whether AI-assisted “coaching” necessarily makes a fair trial impossible, and
(ii) whether the trial process (directions, disclosure to the jury, further cross-examination, adjournment, etc.)
could adequately neutralise any prejudice to the defendant.
2) Summary of the Judgment
The Court of Appeal allowed the prosecution appeal, holding that the Recorder’s decision to stay the
indictment was wrong in law/principle or was not a reasonable ruling within s.67 Criminal Justice Act 2003.
While accepting that witness coaching is impermissible (and that the AI documents amounted to coaching),
the court held that a stay is a last resort and, on the facts, the alleged unfairness was
capable of being addressed by the ordinary trial process.
The ruling staying proceedings was reversed and directions were given for prompt listing before the Resident Judge,
with the retrial to be fixed at the earliest opportunity.
3) Analysis
A. Precedents Cited
Maxwell [2010] UKSC 48 (abuse of process framework)
The defence application was brought under the first limb identified in Maxwell [2010] UKSC 48:
that a fair trial is impossible. The Court of Appeal treated the question as a practical one:
whether the trial process could realistically cure the prejudice said to arise from the AI coaching materials.
R v Momodou and Limani [2005] EWCA Crim 177 (prohibition on coaching; curative role of trial)
The judgment’s anchor authority was R v Momodou and Limani [2005] EWCA Crim 177.
It was relied upon in two distinct ways:
-
Substantive rule: “witness training or coaching” is prohibited, because it risks tailoring,
contamination, and (even without dishonesty) altered emphasis and alignment with perceived needs of the case.
The court repeated the warning that the evidence may no longer be the witness’s own.
(The judgment also referenced the lineage of authority cited in Momodou:
Richardson [1971] CAR 244; Arif, unreported, 22nd June 1993;
Skinner [1994] 99 CAR 212; Shaw [2002] EWCA Crim 3004.)
-
Remedial approach: Momodou emphasises that even serious unfairness is “almost always”
manageable within the trial process by placing the problem before the jury, coupled with firm directions.
A stay is “salutary, but rarely exercised”.
The Court of Appeal used Momodou to show that the existence of coaching does not automatically
make a fair trial impossible; rather, the court must ask what can be done within the trial to restore fairness.
R v HGF [2026] EWCA Crim 570 and R v Ng and O'Reilly [2024] EWCA Crim 493 (stay as exceptional)
Reinforcing the exceptional nature of a stay, the court cited R v HGF [2026] EWCA Crim 570
(Lady Carr, LCJ), which in turn cited R v Ng and O'Reilly [2024] EWCA Crim 493.
These cases were deployed to underline that stopping a prosecution for abuse is a remedy of last resort.
R v Bater-James [2020] EWCA Crim 790 (phone examination; reasonable line of enquiry; proportionality)
Although not the appeal issue, the judgment flagged concerns about the breadth of the order requiring a full
phone download and referenced R v Bater-James [2020] EWCA Crim 790:
complainants’ phones are not routinely subject to wholesale examination; there must be a
“reasonable line of enquiry” and a proportionate search.
The court nevertheless acknowledged that the download yielded highly relevant material bearing on reliability.
B. Legal Reasoning
The Recorder treated the AI documents as coaching and concluded that no remedy could restore fairness.
The Court of Appeal identified the error as a failure to grapple with how the trial process could
address the prejudice and why it could not be left safely to a properly directed jury.
1. The correct question: “Is a fair trial impossible?”
The court accepted that coaching is impermissible, including when delivered via AI.
But the dispositive question under the first limb of abuse is whether the unfairness is
incurable by the trial process. The judgment emphasised that the Recorder’s reasons did not reveal
any structured assessment of alternative measures (adjournment, recall, agreed facts, directions, restarting the trial).
2. Why this case was particularly amenable to trial management
Two features mattered:
-
Uncontaminated evidence-in-chief: the complainant’s evidence-in-chief was given by pre-recorded
ABE interview shortly after complaint, and (on the information before the appellate court) long before the AI use.
This meant the core account could be treated as independent of the later “coaching”, focusing the problem on
cross-examination answers and credibility assessment rather than rewriting the entire narrative.
-
The coaching was recorded: unlike many coaching scenarios conducted “behind closed doors”,
here the AI output existed in documentary form and could be put before the jury in full (or by agreed summary),
enabling targeted cross-examination and careful jury evaluation.
3. The court’s implicit roadmap of proportionate remedies
The Court of Appeal endorsed the Crown’s submission that the following tools were available (depending on how
matters developed): adjournment; discharge and restart; recall of the witness to confront the AI documents;
placing the AI documents before the jury; robust directions; and (if justified) evidential consequences such as
exclusion or no-case submissions. The key point was not that any specific step was mandatory, but that the Recorder
had not adequately analysed why such steps would fail to secure fairness.
4. Guidance for future abuse applications
The court encouraged counsel to “explicitly” explore all trial-process solutions when abuse is alleged, so that the
judge can address each option in a reasoned ruling—reducing the risk of premature stays driven by concession or assumption.
C. Impact
1. AI does not change the legal test—but changes the evidential terrain
The judgment does not create a new category of abuse for AI. Instead, it applies established abuse principles to a new
factual pattern: AI-assisted “preparation” may be prohibited coaching, but it does not automatically compel a stay.
Courts should default to managing the problem through disclosure, cross-examination, and directions unless fairness is truly impossible.
2. Likely practical consequences in criminal trials
-
Early warnings to witnesses: the judgment states that all witnesses should be firmly discouraged from using AI to
prepare evidence. This is likely to be operationalised through police/caseworker guidance and witness care practices.
-
Expanded disclosure disputes: AI use may generate new lines of enquiry and may justify more intrusive scrutiny of devices
and online activity—though still bounded by R v Bater-James [2020] EWCA Crim 790 proportionality.
-
New “contamination” cross-examination: parties can be expected to explore prompts used, the scope of AI interaction,
whether suggested answers were adopted, and whether the witness’s recall evolved as a result.
3. Institutional and rule-making implications
The court expressly suggested that the National Police Chiefs Council, the Crown Prosecution Service and the Criminal Procedure Rules Committee
consider how to address and regulate AI use in criminal proceedings, echoing developing work on AI transparency in the civil jurisdiction.
This points toward future protocols (for example, disclosure of AI use, preservation of prompts/outputs, and standard witness warnings).
5) Conclusion
R. v FGD confirms that AI-generated witness coaching is treated as coaching in principle, but it also reasserts a key limit:
a stay for abuse is a remedy of last resort. Even where AI material is troubling and goes to credibility, the normal response is
to deploy the trial process—disclosure of the material, focused cross-examination, and robust directions—particularly where the core account
(such as an ABE interview) predates the AI interaction and where the “coaching” is fully documented.
The judgment also signals a wider system challenge: AI use by witnesses may drive more intrusive disclosure disputes and requires clear guidance
from policing, prosecution, and procedural rule-makers to protect fairness while avoiding unnecessary invasions of privacy.