Artificial intelligence presents the law of evidence with a familiar problem in an unfamiliar form. Courts have always had to decide what can be trusted, what must be proved, and how confidence in a conclusion is justified. AI does not change those questions. What it changes is the ease with which convincing material can be produced, and the difficulty of knowing how it came into being.
For most of legal history, the production of evidence involved friction. Documents had to be written, copied, stored, and retrieved. Witnesses had to remember, recount, and be tested. Even dishonest evidence usually bore the marks of effort. Artificial intelligence reduces that friction dramatically. Text, images, audio, and video can now be generated quickly, cheaply, and at scale. Much of it looks plausible. Some of it looks indistinguishable from the real thing.
This creates a problem that is not primarily technical, but epistemic. The question is no longer simply whether a document is authentic, but what “authentic” means when material is generated rather than recorded. An AI-produced document may not be forged in the traditional sense. It may not purport to be a copy of anything that ever existed. Yet it may be relied upon as if it were the product of genuine human observation or reasoning.
The law of evidence is not well adapted to this distinction. Traditional doctrines tend to assume a human source. A witness saw something. An expert formed an opinion. A document records an event or transaction. Artificial intelligence disrupts that model. It produces outputs that look like testimony, analysis, or records, but which are in fact statistical artefacts of training data and prompt design.
One response is to insist on transparency. Parties should disclose when AI has been used, how it has been used, and for what purpose. That is a sensible starting point, but it is not a complete solution. Disclosure tells the court that AI was involved. It does not tell the court whether the output is reliable, appropriate, or probative. Nor does it answer the more difficult question of how much weight should be attached to it.
Another response is to treat AI-generated material as a form of expert evidence. After all, it often purports to analyse information, draw inferences, or summarise complex material. But this analogy is also imperfect. An expert can be cross-examined. Their assumptions can be exposed. Their methodology can be tested. An AI system cannot explain itself in any meaningful sense. At best, one can interrogate the people who designed or deployed it, but they may not be able to account for a particular output.
There is also the question of contamination. AI systems trained on vast quantities of publicly available material inevitably absorb errors, biases, and falsehoods. When such a system produces a confident statement, it may be repeating a mistake that has already circulated widely. The authority of the output derives not from its correctness, but from its fluency. That is a dangerous foundation on which to build evidential weight.
For judges, the difficulty is acute. Courts are accustomed to assessing credibility and reliability in human witnesses. They are less well equipped to assess the provenance of machine-generated material. The risk is not that judges will be deceived in every case, but that they will be required to make decisions on the basis of material whose epistemic status is unclear. That uncertainty undermines confidence, even when the outcome is sound.
For practitioners, the risks are practical as well as theoretical. There is a growing temptation to use AI tools to generate chronologies, summaries of evidence, or even draft witness statements. Used carefully, such tools may assist organisation and clarity. Used carelessly, they may introduce errors that are difficult to detect and embarrassing to correct. More seriously, they may cross the line into manufacturing evidence, even where no dishonesty was intended.
The problem is compounded by professional incentives. Time pressure, cost constraints, and client expectations all push in the direction of speed. AI tools promise speed. The discipline required to check, verify, and contextualise their outputs runs against that pressure. The danger is not that lawyers will deliberately mislead the court, but that they will come to rely on material whose origins they do not fully understand.
It is sometimes suggested that technological solutions will resolve these difficulties. Watermarking, provenance tracking, and cryptographic verification may all play a role. But technical fixes cannot eliminate the need for judgment. Even a perfectly verified AI-generated document remains what it is: an artefact produced by a system optimised for plausibility, not truth.
The deeper issue is one of trust. Courts operate on a fragile but essential assumption: that the material placed before them is offered in good faith, and that professional obligations act as a brake on excess. Artificial intelligence strains that assumption by making it easier to produce material that looks trustworthy without being so. The response cannot be to abandon trust altogether, but to reinforce the norms that sustain it.
That points towards a renewed emphasis on professional responsibility. Lawyers will need to be clear, both with clients and with courts, about how AI has been used and what reliance can properly be placed on its outputs. Judges will need to be willing to ask direct questions about provenance and verification. Procedural rules may need to evolve, but cultural change will matter at least as much.
Artificial intelligence does not make evidence meaningless. It does, however, make the evaluation of evidence more demanding. The law’s task is not to keep pace with every technological development, but to preserve the conditions under which reasoned judgment remains possible. In the evidential sphere, that means insisting on clarity about source, method, and responsibility, even when the tools involved are opaque.
The credibility of the justice system depends not on novelty, but on trust. Artificial intelligence tests that trust by making deception easier and verification harder. The answer lies not in panic or prohibition, but in discipline: intellectual, professional, and institutional. That, at least, is a problem the law has faced before.