Adrien Herbert, commercial partner at Excello, reflects on the rise of AI and why using it for ‘free’ legal advice is no substitute for a true expert’s help.
Along with crypto currencies and fungible tokens, AI and what it can do to enhance productivity, seems to be the talk of the world at present. But just like crypto and fungible tokens, AI, how it works and how to make best use of the technology remains a risk-ridden mystery to most dilettante users.
This is why those who believe that it provides a cost-effective alternative to paying for the advice of professionals, should just stop to consider a few less than comfortable “home truths” before risking their business on a hastily, or hopefully composed AI prompt.
“Chat” GPT says everything you need to know
AI learning is all important when it comes to what you get from one you “consult”. There is a very famous saying in computer programming parlance (which I won’t repeat verbatim here) referring to what you put into a computer (in terms of information in the form of code, data or prompts) directly correlating with what you get out (in terms of accurate results).
Early AI developers created what are referred to as LLM’s (Large Language Machines) by tapping into a huge reservoir of readily available and often largely licence-free data in the form of chat rooms and messaging services. This was a very quick way of achieving a “critical mass” of information upon which to launch AI agents. Owing to the global scale and ubiquity of chat, these LLM’s grew quickly and so the reservoir grew exponentially.
To sort through this morass of data therefore, those early developers created neural networks to do the thinking for the AI agents and natural language processors, to set all the information they were sifting through, out in a comprehensible form: at which point the whole process became deceptively satisfying and, therefore, appealing.
The operative word is however, “deceptively”. As anyone familiar with chat rooms and text messaging knows, for every accurate and useful suggestion, you’ll receive at least as many suggested solutions which either misinterpret your needs or simply offer up an answer which they think will help, with little empirical evidence to support that assumption.
It’s the thought that counts
The neural networks became the critical factor within early, chat-based AI agents and without stating the obvious that remains the case today. Just as we use our brains to evaluate evidence or select the best from all the options available to us, when it comes to reaching any decision, so the neural networks of an AI agent evaluate and select / deselect the information the agent needs and the “skills” it requires from all the stored information and skills-based processes available, in order to respond to the “prompt” it is given.
This is the first point at which the dangers of using an AI agent to provide an alternative to professional advice manifest themselves. The uninitiated user, without the appropriate professional training or expertise, has no way of knowing whether the answer they need will (in its entirety) be provided in response to the prompt which they themselves create.
A good professional adviser will ask the appropriate questions of their client in order to tease from them the essential elements of a complete instruction. They will then return to the client if, in the course of their work, they find that they need further details, or clarification of existing points, or if the instruction is in anyway incomplete, inconsistent or open to question.
An AI agent will not “interrogate the brief” in this way: AI agents assume that the person issuing the prompt provides something which is direct, comprehensive and has the meaning which the AI agent first applies to it.
Upon reflection
Often users know to query the first response that an agent provides, a good idea if the initial prompt is direct, comprehensive and has the meaning which the AI agent first gives it and if the user can accurately assess the response the agent provides.
Generative Engine Optimisation and similar techniques promote certain content online to increase the likelihood that it will appear in answers produced by AI agents … chatbots in particular; the affirmatory manner in which AI agents tend to respond, itself tends, at each stage, to provide an answer which supports the enquirer’s position; and so if you draft an imprecise prompt and then query the resulting response, the risk is that an AI agent will lean further towards the counterfactual, predicting that is the response that will best meet with acceptance.
The critical response
The dangers of using AI to replace professional advice should now be clear. An untrained “correspondent” doesn’t have the level of understanding of the professional to be able to assess the value of their prompts and so cannot be sure that they have asked the question which they need to ask, or that if asked, that they have achieved it with near perfect clarity and precision covering all aspects of their need.
When assessing the response given, an end user will also fail to encounter their “unknown unknowns” or to see the aspects in which the response provided fails to meet their actual requirements (as opposed to those they have expressed). If they simply hit “re-run” then the agent may respond with a solution which, if anything, may be further from a response which answers the user’s needs.
Nothing beats expert knowledge…
As the situation stands, professionals regularly receive instructions which have been created by prompting an AI. These are often imprecise and pages long, which can quite often require review and extraction of the essential details, whilst also making enquiries of the client to fill in gaps which result from the extent of the client’s comprehension of sources of the professional advice and how different sources might be required to address all the issues presented by the enquiry.
A real subject matter expert can swiftly review and assess the responses you yourself illicit from an AI. This isn’t to discount the part to be played by junior professionals, however: taking the opportunity to learn about their client’s case and business; developing their professional and consulting skills; taking some part of the workload of supervisors, who might otherwise have to prioritise one matter over another; and all whilst saving the client time they might spend doing those things which actually advance their business.
Most importantly, the pipeline for the development of senior professionals will inevitably dry up if all the work previously undertaken by junior professionals is provided by machines through AI legal advice… no matter how good and how well-trained the LLM’s become over time. Clearly this is not a consideration in the short term, but partners in practices can rise through the ranks and into partnership in less than ten years and so the effects of such a cultural change will be medium, not long term.
Where are we headed?
Some futurologists might suggest that over that timeframe, the quality of AI will develop to the extent that an AI can “interrogate the brief”, evaluate the circumstances and advise better than any relatively competent professional might. To what extent would insurers under these circumstances be prepared to indemnify a business which put all its money “on red” and lost? The idea that machines can replace professional advice in the short to medium term is not genuinely practicable.
As a final note, perhaps the societal and environmental impacts of regular AI use deserve more reflection than they currently receive. Those who determine the energy policy and water requirements of countries leading the development of data centres are already grappling with the seemingly insatiable demands of data centres. Plans including proposals to recommission nuclear energy at scale require careful monitoring, given the safety and environmental issues that prompted earlier decisions to decommission them.
For commercial legal advice which is not limited by AI, contact the team at info@excellolaw.co.uk or reach me personally.
