Artificial Intelligence’s ‘Oh Really’ Moment

Most don’t know that GPT stands for ‘generative pre-trained transformers’, let alone that a main rival is called BERT (Bidirectional Encoder Representations from Transformers). You’ll be pleased to know that CHAT doesn’t stand for Cheery Histrionics About Twaddle; it’s just ‘chat’; and it’s really not twaddle.

Even the term ‘A.I.’ is becoming passe, apparently; the cool kids refer to M.L. now, but that’ll probably change next week too. So is there really more than hype to this latest galloping herd of unicorns?

Yes, But.

Old wine and new bottles

There’s a big difference between good old-fashioned ‘A.I.’ and the new fun/dangerous stuff called ‘generative A.I.’ (aka ‘ML’). A convergence of technologies from chips to architecture has accelerated (in the past ten years especially) to deliver something powerful. It all borrows heavily from past ‘bubbles’ though. Decision trees and expert systems are positively scribbles on the cave wall by comparison to what’s emerging now. Equally much of what masquerades as A.I. is actually just chatbot front-ends on some well thought out business process automation. Adding machine learning and neural networks, however, is a different game – one where chess is really the play, not just some more crazy golf.

Intelligence starts with data, and data structure work goes back to the bad old days of Dotcom’s DTDs/XML even. Then there’s all that Big Data work on image libraries, and eDiscovery work on data strings, tags, and taxonomies. Add the enablers of gaming chip power leaps and data warehousing capacity, and out of this ‘soup’ has come some pretty useful stuff. It’s also ‘stuff’ that really only mega firms like Amazon, Microsoft and Google have the heft to manage. MS’s investment in OpenAI is estimated to be $10bn+ already. It runs from proprietary chips through the full stack. The Legal Market leader Thomson Reuters just spent $650m on a start-up here too; small change by comparison perhaps, but pulling out of law firm enterprise software (Elite) to then dive in here is instructive. A.I. currently is a bit like Turing’s ACE computer that filled a room ending up in a smartphone (eventually). It’s just they’ve done this in 8 years, not 80.

The Oh Really? test

The key features of ‘generative AI’ are NLP, neural networks, and large language models (LLM). NLP (forget neurolinguistic programming for now) means ‘natural language programming’ here and covers issues such as: information retrieval, named entity recognition, semantic relationship extraction, text/document classification and ranking, annotation, topic modelling, keywords, machine translation, speech tagging, semantic role modelling, word sense disambiguation, grammatical error correction, textual similarity, summarisation, Q&A, question generation, image captioning, and sentiment analysis. (Yup, you read that right – ‘sentiment’.)

In short, this is the ‘Oh really’ test. Sentiment analysis, for example, is a key component in LLMs that makes all the difference. It’s robotic otherwise. But a neural network is really useful when it can be trained to recognise when ‘Oh really’ is used as (a) a good surprise; (b) a bad surprise, (c) a request for confirmation, (d) a cynical insult, and not just (e) a Tennessee surname. This is not linear programming on Boolean rails. It has exponential reach and potential. It genuinely is clever stuff.

M’learned Friend?

Why is it attractive to lawyers? Lawyers and lawmakers like to assume they are setting rules that will be obeyed (‘laws’). They are essentially aiming for a predictive approach to human behaviour. Programmers have entrenched that in what they call the ‘backpropagation algorithm’ used in ‘deep learning’. The programmers aim to adjust weights and biases among network neurons to get the actual result of actions as close to the intended result as possible. As lawyers would say, they are aiming for being ‘ad idem’; being of one mind on whatever ‘it’ is. The fun starts when mathematical generative approaches meet real life when creative minds start gaming systems and mathematical ones rely on stochastic plays.

Most of what is currently trumpeting ‘Chat-GPT’ A.I. credentials is not that exciting really, with limitations on the scope of the data available for its learning or simply not enough time yet to get out of kindergarten on the networks. The investment in real commercial innovation is both prescient and significant, however, with deep pockets roaming around looking for the protein to make the beast roar. In legal, regulatory and compliance worlds, the main use cases for A.I. are becoming fairly obvious:

Prediction: Insurers love the fact that you can automate ‘likelihood of success’ assessments in bulk cases/actions. It resolves the low value/high cost and vice versa case conundrum that bedevils litigation, especially in consumer law. There will be versions that work for litigation funders too, especially the mid-market commercial court actions. Several complex processes can be automated with time and care, and it can unlock litigation potential. Fletcher’s PI and negligence practice, for example, started with IBM’s Watson and is migrating to ChatGPT to build decision support systems in medical and structured data environments. That’s a £2.5m investment so far from a law firm with £35m in sales. These are potentially powerful extensions of consumer legal power when they work, but for now, they are also efficiency tools for the law firm.

Clever Chatbots: B2C service firms generally love the ability to get genuinely clever ‘chatbots’ handling an ever wider (and seemingly less structured) range of end-user interactions, and the ability to deploy these in processes from CRM and low-value-high-frequency issues is attractive. Keeping the fourth estate at bay is a hardy perennial, especially for insurers and their law firms.

PI Protection and research: GCs love the fact that they can ensure their underfunded juniors with limited legal exposure can get up to speed quickly on legal research with an AI bot holding their hand (it also reduces recruitment pressures at least in theory). In the US the research issue is much more important to law firms as litigators have to cover all the angles, so it has a special role there too, and this explains much of the Silicon Valley interest currently.

‘Virtual associates’ are becoming a class of their own now, some trusted to even be client-facing rather than associate back-ups, and private practice is finding them useful. A big step up from the chatbots they have a business process ancestry predominantly and a deeper approach to particular technical processes. Some are already growing ‘wings’ in technical niches such as patent litigation, for example. While the white papers will show fantastic ROIs for some law firms and GCs, in reality, they’re also still mostly professional indemnity premium protection tools, and effective in fairly constricted environments/processes. Some specific use cases are emerging, however, notably:

Contract risk: ‘Playbooks’ are the new gizmo in contract management and led by GCs the CLM industry is grabbing the potential of contract screening, creation, and management with both hands. The bastard child of CLM, Big Data, and eDiscovery tech has enormous disruptive impact. Trawling Outlook has become routine recently, but integrating this into A.I.’s large language modelling is new and powerful, especially when it takes distributed document automation and creation through all of the CLM stages. In compliance roles, running every contract, including drafts and versions in real-time against established standards, precedents, and ‘playbooks’ to generate escalations and actionable corrections or alerts is as the saying goes: ‘already here, just not well distributed yet’.

Not so fast?

There are downsides too:

These guys don’t speak like humans any more. They’ll say they apply asymptotic tests to the kinetics of your ontology rather than admitting they haven’t a Scooby’s about what the hell your business actually does. That looks like it’s only going to get worse I’m afraid as they keep purloining TLAs like NLP and LLM at will.

There is already blood on the carpet, too. Microsoft’s OpenAI faces class action lawsuits for copyright and privacy law violations. The power and accuracy of A.I., and especially generative A.I., need data, good data, and plenty of it. That’s why the likes of Thomson Reuters can’t avoid getting involved here. Their very position in the legal industry is at stake.

It’s an arcane lesson, but early iterations of large language modelling were, well, large. So could it be agile too? Especially in legal argument, could it do a 180o switch on limited data? Legal argument is not always linear or even probabilistic, but this is what most digital LLM models work on. Well, they’ve thought of that too, no doubt from the chess, journalistic, and gaming worlds. GANs (generative adversarial networks) pitch a ‘generator’ against a ‘discriminator’ (sometimes more than one) with the latter unconstrained by the former and able to take a view on whether it is real or fake. In short, it can argue, a lot. Its gaming DNA sneaked out, however, when in the Avianca case in the US recently a litigator, having double-checked the cases he was citing from a Chat-GPT search, found in front of the judge that FastCases’ ChatGPT had actually just made them up. Chat GPT, it seems, wants to win at all costs.

GIGO still applies too and LLMs rely on scouring large amounts of data, some (much?) of which can be polluting. There’s an inherent problem with the quality of data available from ‘scouring’ (hence Thomson guarding its proprietary content), but in some cases there can be bad actors introducing poisons deliberately too. Open.ai has already encountered this and had to shut down to cope with it; data privacy is not as simple as it looks. The fields of AI safety and ethics are behind the curve by some way currently.

Amazon, Google, and MS are all coming at this from their respective hills-to-die-on. MS is offering copilots to work in every conceivable way around their desktop apps. Contract drafting from within Word is already a reality, for example, and the (impressive) translation tools are being given away. Amazon is getting close to the programmer community with integrations for developer tools like Visual Studio Code. They’re in effect doing what MS is doing, just for the black-screen-flashing-cursor guys. Google has a growing list of plug-ins covering every facet of its search operations, and all three are arm wrestling over chip power and development.

A Beginner’s Guide

Well, try this. It’s a simple example of a complex legalistic issue which A.I. can accelerate and help with.

Shareholder Agreements do not easily ‘template’ as each company has different issues. For a seed capital player, you can have pages of dilution protection provisions, while for a family firm the big issue may be death in service insurance. The costly bit is getting a group of shareholders (the majority shareholder being first among equals) to agree on what needs covering at all, and then how, preferably in plain English.

So here are five ChatGPT prompts, that should get (a) the issue fleshed out, and (b) the shareholders in a position to get things finalised with a ‘proper’ lawyer for a reasonable fee. The main reason they typically remain unsigned (and largely unenforceable) is that lawyers’ fees escalate while negotiations falter or dithering delays things. If you can get the ducks in a row and the contentious issues down to two or three, any lawyer worth their salt can sort that easily. They also hate having clients who can’t be billed because the damn thing’s dragging on and not resolved or executed.

Anyone with shares in an SME could use this, and it is in effect using ChatGPT to automate a non-executive director process, getting shareholders on a level playing field (or at least an agreed one). The queries are numbered according to the frequency a non-exec has to deal with them, not the typical order an SA tackles them.

A simple prompt engineering discipline is to use CIDI (context/instruction/details/input), so start with, and you could even go straight to asking for a draft document, but as a non-exec we’re looking at the process a stage earlier. So for now say:

You are the majority shareholder in a limited company looking to regularise decision-making between shareholders. Show explanatory examples for each issue in compiling a Shareholders Agreement including the implications of not addressing it. Finally, provide draft clauses to cover each point in a draft Shareholders Agreement based on the commercial law of England & Wales:

Explain how drag-and-tag, put-and-call, anti-dilution, and right of first refusal/pre-emptive rights work (with examples), and where waivers can apply to these, what consequences they will have; give examples of clauses covering each of these issues.

Show some typical clauses setting out a capitalisation table, shareholder rights regarding voting, receiving dividends, share classes and definitions, preferential rights, application of insurance proceeds, rights to information, directors’ powers, appointment and removal of directors and the roles defined among board and non-executive positions.

Show how good-leaver and bad-leaver provisions work and within draft clauses list the conditions that define or trigger each in the context of minority protection provisions including breach of contract, misconduct, bad faith, breach of the shareholders agreement provisions, unapproved share transfer, non-participation, competition with the company or specific performance issues.

Suggest clauses for use in shareholders agreements showing how DCF and ebitda-based valuations work, the role of ‘willing-buyer-willing-seller’ approaches, how zero net assets definitions work alongside those, as well as how control premiums should be treated, and optional versus mandatory buy-back provisions.

List the clauses most commonly needed in a shareholders agreement with sample clauses for dispute resolution and arbitration provisions; explain the significance of having a shareholders agreement executed as a deed, and why you need to include defined jurisdiction (in this case, the laws of England & Wales) and entire agreement provisions (with examples for clauses for each).

It’s a small example, but a key one. Online ‘quickie’ templates simply don’t work here. The value is in the issues pre-empted, and typically it is fear of legal fees that stops them being tackled and/or completed.

AI here is not doing a bad job, and this is only scratching the surface…