Push Button Access to Justice: Too Much of a Good Thing?

Ease of access to justice is a good thing but can there be too much access? The Employment Tribunals in the United Kingdom have discovered the downsides of generative Artificial Intelligence (AI) with an inundation of interim applications for relief. The tsunami of recent applications prompted the issuing of new guidance.

Before we get to the tools the employment tribunals have devised to attempt to address the issue, it is worth talking about friction in dispute system design. “Friction” is the effort, steps or complexity involved in completing an action: “Every extra step between intention and action is a moment at which people fall away”. Generally, in access to justice too much friction is not a good thing, because it prevents people from accessing benefits or exercising rights (referred to as “sludge”). As behavioural psychologist Tom de Bruyne states, “understanding which friction to add, which to remove, and which is sludge, is one of the most practically important skills in behavioural design”.

Understandably, ideas on reducing friction have received prominence in the access to justice community – we generally want to eliminate barriers to access to justice. What has received less attention is when some friction might improve access to justice for everyone: “deliberately adding friction to undesired behaviour to make it harder and less automatic”. In digital design, de Bruyne calls these “speed bumps” – slowing down access enough to prompt “a small amount of deliberation”.

Good friction serves the person experiencing it by creating space for reflection, preventing irreversible mistakes, and slowing down automatic choices that do not reflect considered preferences. … The question is not how much friction, but what it is for.

One of the frictions in daily life is effort. Amazon discovered a way to almost totally eliminate that friction when it introduced “one click” shopping. The result was an increase in impulse buying. In the case of consumer purchases, some friction is beneficial as a way of forcing minimal reflection by the potential purchaser. Similarly, when something is “free”, people rarely consider the downsides of obtaining it. This is described as the “zero-price effect”. People are instinctively afraid of loss, so when something is “free” those fears disappear. We usually decide if something is worth obtaining by weighing the cost, the effort to get it, and the benefits it offers. However, when something is “free” we tend to ignore that normal decision-making process, thereby ignoring any potential drawbacks. Putting a price on something – even a non-monetary price of time – can result in a better decision-making process that balances benefits against costs.

AI is a tool that eliminates beneficial friction. A complaint or claim can be generated by typing a question and hitting “return”. A recent study noted the difference between previous technologies that reduce effort (such as washing machines, power steering and spell checkers) and AI:

… First, AI targets intellectual rather than physical or merely clerical work. In our lifetime, we can think of no other developments that so directly target the creative process. And second, AI’s removal of friction is extreme. While washing machines and power steering remove excess friction—tedious or insurmountable obstacles that offer little benefit for learning or meaning—AI also strips away beneficial friction. Working with a chatbot, people can move from ideation to evaluation without exerting meaningful effort, without questioning the output, and without engaging the cognitive processes that foster ownership, retention, or critical thought.

AI also reduces the time involved in articulating a position. This not only means that a complaint might be poorly articulated, but it also feeds the sentiment of “what do I have to lose by submitting this claim?”. This can lead to meritless applications clogging up the tribunal process. This brings us back to the issues facing employment tribunals in the UK, where the number of interim applications has become overwhelming, even where the chances of success are low (although not zero).

The Employment Rights Act in the UK provides interim relief in cases of unjust dismissal complaints. The granting of interim relief prevents a dismissal from taking place (either by way of reinstatement or suspension with full pay) until the tribunal has determined the outcome (after a full hearing). The guidance document notes that because of these effects, there is a high threshold for complainants to meet to convince the tribunal that it should be ordered – most applications are unsuccessful. The test is whether the complainant is likely to win on the merits. The guidance document states that the test is set comparatively high for reasons of policy: Dandpat v University of Bath EAT/0408/09: “It is something nearer to certainty than mere probability”.

There has been a “significant increase” in interim relief applications to the tribunals largely in cases with whistleblower allegations, “often indicating use of artificial intelligence (“AI”)”. The guidance document does not provide total numbers of interim relief applications but does state that in previous years the tribunals would receive about 20 per year, whereas now most tribunal offices are receiving 20 a month. In addition, there has been a significant increase in the amount of documentation accompanying each application. Although not explicitly stated, AI may be behind this as well.

The backlog at the employment tribunals continues to grow. At the end of March this year, there were 64,000 cases open, up from 45,000 the previous March. The impact of this increase in interim applications has had a cascading effect on the ability of the tribunals to manage their caseloads. Because interim applications are an emergency measure, the tribunal must give them priority. As a result, hearings on the merits of other cases are often postponed or delayed. The guidance document explains:

…Because the success rate of applications for interim relief remains low, these trends have an adverse effect on the administration of justice, including causing unnecessary delay to other users. Further complexity is added when, in response to applications for interim relief, parties then make consequential applications (and counter-applications) for costs/expenses.

The guidance document sets out the steps the tribunal will take to manage interim applications (both current and prospective steps):

• Since the tribunal is doing an assessment only of the prospects of success it will not hear oral evidence (unless it directs otherwise)
• The parties must ensure that the material placed before the tribunal is proportionate
• Tribunals will make case management orders that limit the amount of material submitted to them, including page limits for documents or word limits for witness statements or submissions
• There is an allocation of time (only extended in exceptional circumstances): one hour for the judge to read the materials provided, 30 minutes for each side for oral submissions, and one hour for the judge to reach a decision and provide a short oral judgment with reasons.
• If the parties provide the tribunal with material exceeding what can be read and understood in the time allocated, the judge will require them to identify the most important documents on each side and then restrict consideration to those documents.

The guidance document also notes the tendency of AI generated legal submissions to be “too long and complex, contain irrelevant material and fail to focus on the key points in the case”. The guidance states that parties using AI have the responsibility to ensure that what is submitted is “concise, relevant and accurate”.

One is reminded of the adage about bailing a rowboat with a thimble. Although these steps may chip away at the onslaught of applications, it will not significantly address the problems.

Courts in Canada are facing similar challenges with the use of AI by self-represented litigants. In a recent article in the Canadian Lawyer Magazine, Justice Marie-Anne Paquette of the Superior Court of Quebec noted that because AI tools dramatically lower the cost and effort involved in producing documents, the length of filings by many self-represented litigants has increased: “this places an additional pressure on judicial resources, a trend that we can already observe very, very precisely and concretely in my office”. British Columbia Court of Appeal legal counsel Shirley Smiley noted the same concerns:

Because genAI platforms make it easier for a litigant to produce more material, faster, and at little or no expense, litigants may file more applications or apply to review or vary previous decisions even with very little chance of success.

This burdens the court and responding litigants without improving access to justice outcomes.

Tribunals cannot arbitrarily reject applications. They can devise triage mechanisms for applications – but that does not allow them to simply ignore applications.

The solutions to the inundation of AI-assisted claims are, unfortunately, beyond the authority of tribunals to address. The solution lies with AI companies or, if they are not sufficiently motivated, with government regulation.

AI tools need to be trained to be:

• accurate;
• honest (not sycophantic); and
• concise.

We have laws about truth in advertising as well as laws around negligence. If a legal professional can be sued or disciplined for providing inaccurate or misleading legal advice, the same rule should apply to AI bots providing legal advice.

Sycophancy is one of the biggest flaws in AI design. The ideal advisor is one who will give you an honest answer on your likelihood of success in a dispute. A good advisor will tell you if pursuing a dispute is a good use of your time and money – AI is rarely that honest.

AI bots are also notoriously prolix, giving you four paragraphs when one will do. This prolixity adds to the burdens facing tribunals, as decision makers must read all the documents in the file (even if some of the material is just skimmed).

Some AI evangelists will say that you can get results from AI bots that are accurate, honest and concise if you use the proper (and detailed) prompts. First of all, why should it be necessary for consumers to fix the problems of a product? We don’t expect such detailed interactions with advisors. In fact, most advisors would be insulted if you kept going back to them with more and more detailed questions. Secondly, most AI users in the general public do not have the training or the inclination to draft proper (and lengthy) prompts.

AI does have the potential to improve access to justice. Justices Hinckley and Paquette both noted that the quality of submissions from self-represented litigants has improved. Paquette noted that many filings are “clearer, better organized, and closer to the expected structure of legal proceedings than what we would have seen a few years ago.” Hinkley stated that “the format gap has closed. … Materials… now look like what they should look like, and that’s a real change.” He also said that that does not mean that the arguments in those submissions are sound.

Until we fix the underlying flaws of generative AI design, AI is more likely to increase meritless legal filings, thereby decreasing access to justice for all.