AI in criminal defence: reviewing disclosure without losing the point
By the ChamberOne team 4 minute read
Modern criminal cases can arrive with phone downloads, thousands of messages, hours of footage and boxes of documents that the prosecution does not intend to rely on. A defence lawyer on a legal aid timetable cannot read all of it, and the defendant's interests may turn on a single message that nobody noticed. AI tools can help with the sorting. They can also make it easier to skip the one page that matters. This article looks at how to use them for triage while keeping the lawyer's attention where it belongs.
Key points
- Use AI to order and point at large material, not to decide what is irrelevant.
- Read first what contradicts your client's account, and sample the low-ranked material by hand.
- Treat summaries as an index, check messages in context and use human review for slang and other languages.
- Check the rules on use of disclosure and keep privileged notes out of outside tools unless you know how they are handled.
The problem is volume and the cost of a miss
In a criminal case, an overlooked item can matter more than in most civil work. A message that shows the defendant elsewhere, a witness who contradicts herself, a gap in a chain of custody: any of these may change the outcome. At the same time, the sheer quantity of material makes thorough reading impossible within the time and funding available.
That tension is exactly where software is tempting. The question is not whether to use it, but how to use it so that it narrows the pile without hiding what should be read.
Where AI can help
Think of AI as a way of ordering the pile, not as a reader that replaces you. The useful tasks are those that organise and point.
- Grouping messages and documents by participant, date range and topic.
- Searching by meaning as well as keyword, so that a message about 'the car' is found when you searched for a vehicle's make.
- Building a draft timeline from dates in messages, records and logs.
- Flagging references to people, places, times or events named in the charge.
- Producing short summaries of long threads, each linked to the source so that you can jump to the original.
- Spotting near-duplicates and repeated material so that you read it once.
Where the lawyer must keep reading
A tool ranks material by what it was asked about. It does not know the defence theory unless you tell it, and it cannot know what will matter once a witness has given evidence. The material most likely to help or hurt is often outside the obvious topics: a tone of voice, a throwaway remark, a change of pattern.
So keep a rule: nothing is set aside as irrelevant on the strength of an AI judgment alone. Sample the low-ranked material by hand, read everything involving the defendant and key witnesses, and read the material nearest to the dates and events that matter. Pay particular attention to anything that could undermine your client's account, because you need to know it before the other side points it out.
- Group by participant, date range and topic
- Search by meaning as well as keyword
- Draft a timeline from dates in the material
- Summarise long threads with links to the source
- Never set material aside on an AI judgment alone
- Sample the low-ranked material by hand
- Read first what contradicts your client's account
Working with a theory, and testing it
Start with a working account of what the defence will say, and ask the tool to find material that supports it, contradicts it and bears on it. The contradicting set is the most valuable output. Read it first.
Be careful with leading prompts. A model asked to find evidence for an alibi may produce plausible-sounding support from text that does not say what it claims. Always open the underlying message, image or page. Treat summaries as an index, not as evidence, and never quote a summary to the court.
Messages, slang and other languages
Chat messages are full of abbreviations, local slang, code words, emojis and mixed languages. A tool trained largely on formal text may misread them, and a mistaken reading of a message can change its meaning entirely. This is especially true where messages mix languages or use transliteration. Ask someone who understands the language and the community to review any message that matters, and consider whether a qualified interpreter or translator is needed for material that may be put before a court.
Do not assume that an AI transcription of audio is accurate. Check key passages against the recording, and note any uncertainty.
Client data, privilege and the disclosure regime
Disclosure material is sensitive. It may include personal data of third parties, intimate images, medical records and information subject to restrictions on use. Before any of it goes into an outside tool, check the rules that govern how material received in disclosure may be used and stored, any undertakings or orders in the case, and the guidance from your regulator or bar. In some matters the safest answer is that the material does not leave a controlled environment at all.
Your own notes, defence statements and instructions from the client are privileged and must not be fed into a tool unless you are sure how the provider handles the text. Consider keeping the defence theory out of prompts altogether and using neutral topic searches instead.
If the matter is funded by legal aid or a similar scheme, check whether the funder has rules on the use of technology or on the time claimed for reviewing material, and keep notes that support the time spent.
Keeping a record of the review
If your approach to disclosure is ever questioned, you will want to show what you did. Keep a short log: what material was received, which tools were used for sorting, which searches were run, what was reviewed by hand and why material was set aside. Note the date and the name of the person responsible.
Brief the client too. Explain that technology is helping to sort material, that a lawyer is still reading what matters, and what you need from them. Clients often know the people in the messages and can tell you quickly which threads are important.
Plan for the later stages as well. As the case develops, new issues emerge and material you set aside may become important. Keep the sorted set organised so that you can re-run searches against a new defence theory without starting again, and schedule a second look at the set after the prosecution case is served in full or after witness statements arrive.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
AI in Witness Bundle and Proof of Evidence Preparation: What to Automate and What Not To
A practical guide for litigators on which parts of witness bundle and proof of evidence preparation AI can handle, and which parts it must not touch.
6 minute readHallucinated citations: what court sanctions teach about checking AI output
The patterns behind fake authorities reaching courts, and the supervision and checking controls that would have caught them.
5 minute readE-discovery and technology-assisted review explained for non-specialists
How predictive coding and generative review differ, why validation matters and how to document the process.
5 minute read