When a Witness Talks to AI Before Court, Whose Memory Enters the Witness Box?

There is a familiar stage in the preparation of a case. A client or witness sits down to explain what happened. The first account is usually untidy. Dates may be uncertain. Events arrive out of sequence. One detail is remembered clearly while another has faded. This is not necessarily a sign of dishonesty. It is often what honest recollection looks like.

Now imagine that, before speaking to a lawyer, the witness consults an artificial intelligence assistant.

The witness types out the story and asks the system to arrange it chronologically. The system identifies gaps, proposes clearer language and asks follow-up questions. It produces a polished statement. The witness reads it, makes a few corrections and returns several times until the account appears complete.

By the time the statement reaches counsel, it may be accurate. It may also be subtly different from the memory with which the witness began.

This, in my view, is one of the less obvious legal problems created by generative AI. The current debate has focused largely on fictitious authorities, deepfakes and documents invented by machines. Those concerns are serious, but they are comparatively easy to identify. A false citation can be checked. A manipulated image can be examined. The more difficult problem may be evidence that still appears entirely human after a machine has helped to shape it.

The question is no longer simply whether AI can manufacture evidence. It is whether AI can influence the making of a witness.

When assistance begins to alter recollection

Human memory is not a recording. People reconstruct the past, and that reconstruction may be affected by later information, repeated retelling and the language in which questions are asked. A witness who hesitates is not necessarily evasive. A witness who says, “I cannot remember,” may be more reliable than one who supplies a confident answer to every question.

Generative AI is designed to do something quite different. It turns fragments into coherent narratives. That ability is useful when drafting a letter or organising notes. In witness evidence, it may conceal the very uncertainties that a court should be allowed to see.

Suppose a witness remembers that a meeting took place sometime in March, before a disagreement but after a colleague returned from leave. A chatbot uses the surrounding facts to suggest a likely date. The suggestion seems reasonable and appears in several later drafts. Eventually, the witness may stop distinguishing between the date actually remembered and the date inferred during the conversation.

No deliberate lie has been told. The witness may be completely sincere. Yet the source of the apparent recollection has changed.

The influence may be more subtle than the addition of a fact. A system can determine which detail deserves emphasis, which uncertainty should be omitted and which event should be presented as the cause of another. It may replace the witness's natural language with words that sound more precise or persuasive. The final statement becomes easier to follow, but perhaps more certain than the underlying memory justifies.

This concern has particular force in Nigeria. Many people try to understand or document a dispute before approaching a lawyer. Cost, distance and delay make preliminary self-help attractive. A person who speaks more comfortably in an indigenous language or Nigerian Pidgin may use a chatbot to turn an account into formal English. That can improve access to legal services. It can also alter meaning while appearing merely to improve expression.

The answer cannot be that any technological assistance makes evidence unreliable. A spelling correction is not the same as a suggested fact. Translation is not necessarily reconstruction. The difficulty is deciding when help with expression has become intervention in memory.

The gap in Nigerian evidence law

The Evidence Act 2011 comes close to this problem without resolving it.

Section 232 permits cross-examination on previous statements made in writing or reduced into writing. Sections 239 to 241 address writings used to refresh memory and, in the circumstances covered by those provisions, allow the adverse party to inspect and cross-examine on the writing. These rules recognise a sound principle. Where a written source has helped to form or restore the account given in court, that source may be relevant to testing the evidence.

An AI conversation does not fit comfortably into either category. It may contain an earlier account by the witness, questions supplied by the machine, suggested answers and several competing versions of the same event. It acts as interviewer, editor and memory aid at once.

If the chat record itself is tendered in evidence, section 84 of the Evidence Act may become relevant because it deals with statements contained in documents produced by computers. Questions of admissibility and authentication would then arise. But that is not the problem that concerns me most. The more troubling case is one in which the conversation has already influenced the witness, yet the record has been deleted or was never brought to counsel's attention.

Consider civil proceedings in which a detailed witness statement on oath is filed before oral testimony. The finished statement may have passed through several AI-assisted drafts. In a criminal matter, a person may have rehearsed an account with a chatbot before making a statement to investigators. In both cases, the true first version may exist only inside an informal digital exchange.

Cross-examination traditionally tests testimony against earlier statements, contemporaneous documents and objective facts. But what if the most revealing earlier account is unavailable? What if the system proposed three explanations and the witness later adopted one? The final statement may appear internally consistent because inconsistency was removed before anyone realised that evidence was being prepared.

A signature at the foot of a witness statement does not answer the problem. It confirms that the witness adopts the final text. It does not explain how the text was produced or which uncertainties disappeared along the way.

The risk is not confined to dishonest witnesses. Repetition can create familiarity, and familiarity can feel like memory. A witness may sincerely defend a detail that began as no more than a plausible suggestion. AI may therefore make an honest witness more persuasive than the facts warrant. Sincerity and accuracy are not the same thing.

What Nigerian lawyers should ask now

I do not think an outright prohibition on AI use would be workable or desirable. These tools may help people facing language, literacy or disability barriers. They can help a client organise a large volume of material before the first conference. The purpose of legal safeguards should not be to preserve difficulty for its own sake.

The better response is to ask direct questions early.

When taking instructions, a lawyer should consider asking whether the client or witness has used a chatbot to discuss the events, reconstruct dates or prepare an account. The question should become as ordinary as asking whether an earlier written statement exists. Where the use went beyond spelling, formatting or straightforward translation, relevant conversations and drafts should be preserved where lawful and practicable.

The professional issues extend beyond evidence. A client may assume that a conversation with a commercial chatbot is private in the same way as a consultation with counsel. That assumption is unsafe. The protection given to specified confidential legal communications under the Evidence Act does not automatically attach to every account entered into a third-party system.

There are data protection implications as well. A witness narrative may contain names, addresses, health information, financial details and allegations concerning other people. The Nigeria Data Protection Act 2023 makes it necessary to think carefully about what personal data is being processed, why it is being processed and where it may be sent. A lawyer cannot surrender professional judgement merely because a digital tool is convenient.

Law firms should develop simple internal rules that distinguish assistance with presentation from assistance that may affect facts, sequence or certainty. Witnesses should be told not to ask a chatbot to fill gaps in recollection or produce the “strongest” version of events. Where substantive use has occurred, counsel must decide whether it affects disclosure, preparation or the weight that can properly be placed on the resulting statement.

In time, the Nigerian Bar Association and the courts may need to provide more specific guidance. It should be proportionate. There is no need to turn every automated spelling correction into a forensic enquiry. The focus should be on material use capable of affecting the substance or apparent reliability of evidence.

The most important question

The law is becoming alert to machines that impersonate people. It is less prepared for machines that quietly improve the performance of real people.

A witness who has consulted AI may enter court, take the oath and answer every question personally. The voice, face and conviction belong to the witness. Yet the structure, vocabulary and certainty of the account may have been developed elsewhere. Unless someone asks how the statement came into being, that influence may remain invisible.

This is why the issue should be considered before the practice becomes ordinary. We should not wait for a serious injustice before asking modest questions about the history of a witness account.

The most dangerous AI evidence may not be a deepfake or a document invented from nothing. It may be genuine testimony, honestly given, that a machine has made more coherent and therefore more convincing than the original memory deserved.

When that witness enters the box, the court will know whose name appears on the statement. It may be much less certain whose account it is hearing.