The judgment of Master Matthews in Pyrrho Investments Ltd v MWB Property Ltd [2016] EWHC 256 (Ch) is the first reported decision of the English High Court endorsing the use of the technology commonly referred to as "predictive coding" or "technology assisted review" in place of "manual" review in disclosure. In appropriate cases the technology can save substantial costs by vastly reducing the number of documents that need to be reviewed by a lawyer.
What is this technology?
Technology assisted review (TAR) comes in various forms but essentially relies on algorithms similar to those utilised by Amazon, Google and other websites to show you products and services of potential interest based on previous browsing and purchasing history. Many readers may be unaware that this technology is probably already filtering their own emails – most spam filters utilise it to pre-emptively divert emails that they deem not to merit human review (even by the recipient) to a spam folder. At the cutting edge of the technology, a learning algorithm recently allowed Google's DeepMind to learn how to beat a grandmaster at the strategy game Go, an achievement recognised as surpassing that of IBM's Deep Blue in beating then world champion Gary Kasparov at chess in 1997).
The TAR software "learns" from the decisions a human reviewer has made on a small sample and applies that learning to predict the choices that it should make in respect of the rest. Best practice is for "training" to be undertaken by a senior lawyer immersed in the detail of the case. This approach aims to ensure the software is given the best and most consistent training on what is in reality a subjective exercise, recognising that even two senior lawyers with equal knowledge of the case can reach conflicting decisions on a particular document, or reach the same decision for different reasons. The human reviewer reviews a representative sample of the population of documents (after any agreed filters, including date ranges and keywords, have been applied).
The TAR software then reviews the concepts and phrasing of the documents or parts of documents categorised as "responsive" and "non-responsive" and, by applying what it has "learned", categorises the un-reviewed balance of the documents. A representative sample of those document categorised by the software are then reviewed by the human reviewer to validate and QC check the decisions until the statistical confidence and margin of error for the software's categorisation are within pre-agreed or at least pre-determined levels.
Authorities in support of TAR
The parties in Pyrrho, having agreed between themselves the use of predictive coding (TAR), sought express judicial approval of this approach, no doubt mindful that there were no English judicial authorities specifically approving their agreed approach. That approval was given on 2 February 2016 and the reasoned judgment was handed down two weeks later.
Master Matthews identified the fundamental importance to the disclosure process of the quality of the search and referred to the concept of "reasonable search" as set out in CPR 31.7 and developed in paragraphs 20 and 21 of Practice Direction 31B. Master Matthews observed that there was no guidance in the CPR or Practice Direction as to how the search should be conducted. He noted that paragraphs 25 to 27 of Practice Direction 31B all refer, in terms, to the use of "other automated search" as distinguished from "Keyword Searching" and the judges of the Technology and Construction Court support an eDisclosure Protocol (which he acknowledged has no normative force).
It is also worth noting that, since April 2013, the sixth of the "menu" options provided by the amended CPR Rule 31.5 (7) has given parties free rein to agree (and the court to endorse) "any other order in relation to disclosure that the court considers appropriate".
Why use TAR?
In making his decision Master Matthews noted 10 factors in favour of approving the use of TAR in this case (adding that there were no factors of any weight pointing in the other direction). The following factors were of general application:
- Experience in other jurisdictions, whilst so far limited, has been that predictive coding software can be useful in appropriate cases.
- There is no evidence to show that itsuse of TAR software leads to less accurate disclosure being given than manual review, and indeed there is some evidence to the contrary
- Moreover, greater consistency may be gained from the computer applying the approach of a senior lawyer towards the initial sample (as refined) to the whole document set, than in using dozens, perhaps hundreds, of more junior fee-earners, each seeking independently to apply the relevant criteria in relation to individual documents
- There is nothing in the CPR or Practice Directions to prohibit the use of such software.
The factors that weighed in favour of the use of TAR in this particular case included the agreement of the parties to its use, the number of documents to be considered for disclosure, the estimated cost of the review and, by contrast, the cost of a manual review of all of the documents being "unreasonable" within the meaning of paragraph 25 of Practice Direction 31B.
What does this mean for future practice?
As might be expected from a case management decision (particularly given the agreement of the parties to the proposed course of action), there is little more of general application and future cases will turn on their facts. There is, for example, little discussion of any factors that might weigh against the use of TAR and no test proposed for when it might be appropriate.
We have been using TAR in some cases "behind the scenes", using the technology to prioritise documents for review, to assist in quality control and to identify sets of near-duplicate and similar documents.
The first decision in the USA (in Da Silva Moore v Publicis Group [2012] saw (self-confessed technology advocate) Judge Andrew Peck approve the use of TAR in a situation where the parties agreed in principle to its use, although they disagreed as to how best to implement such a review. In the landmark case in the Republic of Ireland (referred to by Master Matthews in his judgment) of Irish Bank Resolution Corporation Ltd v. Sean Quinn & Ors [2015], the 2015 first instance decision in this claim was upheld by the Irish Court Of Appeal on Friday 26 February 2016.
The Irish Court Of Appeal upheld the first instance decision and confirmed that there was no obligation to provide "full and frank disclosure" of all relevant documents, and that the use of TAR approved by the court at first instance was "fair, proportionate and took account of the increasing need for the courts to ensure that [disclosure] is complied with in a cost effective manner". This decision will remind English lawyers of the dicta of Mr Justice Morgan in Digicel v Cable & Wireless & Ors [2008] which is as relevant to this new technology as it was in 2008:
"[I]t must be remembered that what is generally required by an order for standard disclosure is "a reasonable search" for relevant documents. Thus, the rules do not require that no stone should be left unturned. This may mean that a relevant document, even "a smoking gun" is not found. This attitude is justified by considerations of proportionality. This point is well made by Jacob LJ in Nichia Corporation v Argos Limited [2007] EWCA Civ 741 at [50] to [52]."
Process, not outcome, counts
The reasonable search, including the review of potentially disclosable documents, is a process, and it is that process, and not the outcome that must be assessed. Whether the review process is undertaken by man or machine, if it satisfies the requirement of a reasonable search, it will meet the requirement of an order for disclosure. Compliance before TAR was not disclosure of 100% of all discloseable documents, nor will it be in future.
This decision represents the first step in a move towards utilising technology to further reduce the number of documents reviewed by lawyers for disclosure. Once again, the US and other jurisdictions have a head-start on the English courts, which will have to develop a body of case law refining the rules and practice of using TAR. Battles will no doubt be fought, as they are now in the US, over the TAR protocol to be used, the confidence levels and margin for error which are acceptable and the extent to which an opposing party may look behind or audit the process by which the TAR software was trained. Thus Pyrrho is likely to be just the first in a long line of cases touching on the use of disruptive "AI" technology, which is on the verge of affecting our lives more generally, and not just disclosure.