Should lawyers in Bulgaria self-regulate their use of AI?
Artificial intelligence (AI) is no longer a hypothetical future, nor an exotic topic for technology conferences. It is here - transformative, powerful and inevitable. Three years have passed since ChatGPT appeared at the end of 2022, and during that time the adoption of various AI applications in business and public life has been growing exponentially. The legal profession is no exception to this process.
American law firms are leading this technological revolution, and the integration of language models into their workflows is already far advanced. The big players have gone as far as developing their own agents for solving legal tasks, in cooperation with legal-tech startups[1]. And Perplexity, a well-known online search engine in the form of a chatbot, has recently started offering built-in legal services in cooperation with the LegalZoom platform[2].
In Europe, including Bulgaria, things are happening somewhat more slowly, of course, because of a more cautious approach to the technology and the specific traditions in the regulation of the legal profession. But here too, more and more law firms are investing in the deployment of AI solutions, specialised cloud solutions are gradually making their way in, and the use of general-purpose applications such as ChatGPT, Gemini, Claude and others is widespread practice.
None of this is a coincidence. AI-based applications have the potential to ease the work of lawyers drastically and to improve access to justice, a potential comparable only to the arrival of the computer or of legal information systems in the past. This technology, however, comes with very serious risks that affect the interests of the legal profession and its clients, as well as of the justice system as a whole.
Against this background, the shortage of clear rules to guide lawyers through the stormy sea of change stands out. While colleagues across Europe already work with guidelines on safety and ethics, in Bulgaria we operate in a complete regulatory vacuum. Last year’s conference of the Sofia Bar Association on the occasion of Lawyers’ Day was devoted to artificial intelligence, but the only result of the otherwise interesting discussion was the presentation of the Guidelines of the International Association of Lawyers (UIA)[3], and nothing binding.
In these circumstances, I believe it is time at least to ask ourselves the question: is regulation of the use of AI by the Bulgarian legal profession needed or not? And if so, let us at least begin the conversation about what it should look like.
Thesis
I argue that AI systems have the potential to transform the legal profession for the better, and for good. However, the absence of any binding rules on their use by the Bulgarian legal profession creates uncertainty and distrust, which could either undermine the foundations of the profession or block the adoption of an otherwise useful technology. With this essay I call for us to begin the debate on future regulation, and to that end I will review the benefits of AI systems and the risks that accompany them.
The Benefits
I. Applications
In just two or three years since they became publicly available, AI systems have found their way into legal work in all kinds of ways. Of course, the initial ideas of an autonomously functioning “robot lawyer” are far from the truth and will probably stay there with the current technology. Instead, we are seeing a quiet and pragmatic introduction of language models into separate functional areas of the lawyer’s practice, such as:
1. Legal informatics
This is the obvious and immediate application of the technology[4], whose language models can read unstructured data (text) and extract structure (meaning) from it. Immediately with ChatGPT, and later with specialised tools, they began to be used for searching case law and legislation, or simply as a more detailed Google Search. There is no doubt that, despite the imperfections of language models, they can save every lawyer hours of work digging through text.
2. Administration and management
AI models are increasingly built into resource management systems (ERP), data analysis tools and other specialised law-practice management software. As a result, it is much easier to track hours worked, to book financial transactions, to analyse inefficiencies and to carry out other management tasks.
3. Client service
Here the technology fills the gap between the lawyer’s working hours and the client’s need for immediate attention. Intelligent assistants can provide round-the-clock “initial intake” on the law firm’s website, gathering the facts of the case before they reach a lawyer. Language models can be used for taking minutes of conversations, managing e-mail messages and more.
4. Document analysis
Even before the explosion of language models, machine learning and statistical models were used to analyse documents and to find evidence in particularly large data sets[5]. Through the process of “predictive analysis”, the system can be trained to indicate which documents are relevant to a legal dispute or contain the information sought. Today this technology is publicly available or built into document storage software services[6].
5. Marketing and business development
Generative models are becoming a powerful engine for maintaining lawyers’ digital presence. They can create outlines and drafts for blog articles, newsletters and social media posts, saving time on research and editing. In addition, AI applications have long been used to target audiences in advertising (e.g. Facebook), but they are now also being built into CRM systems to identify patterns in demand and to suggest new services that would be useful to clients (cross-selling)[7].
Despite this impressive scope, it is important to realise that we are only at the dawn of this transformation. All the applications listed are still in their “experimental phase”, and the real innovations - those specific solutions that will define the unique shape of the legal profession of the future - have yet to be devised and adapted by ourselves in the course of our work.
II. Will AI take our jobs?
Given all the “wonders” that AI systems are capable of, the question is reasonably raised: won’t they take lawyers’ work away?
Let us look at the nature of the work that is currently being “taken away” by AI systems: reading documents, compiling repetitive text, editing and applying grammatical and stylistic rules, searching databases, gathering information, calculating numbers, tracking events and behaviour, and so on. Yes, this is work that usually takes up a significant part of a lawyer’s working day, but it is not the essence of our mission. On the contrary, it distracts us from it.
A lawyer is needed by the client and by the justice system above all for the ability to solve a legal case: to gather and assess the facts, to subsume them under the applicable rule and to formulate a solution[8]. On the basis of such reasoning, combined with experience, knowledge of the context and of the client’s interests, knowledge of the subject matter of the regulated field, general culture and many specific human qualities, the lawyer produces the useful result. Namely: adequate, secure, confidential, ethical, sometimes long-running and durable services such as advice, representation in court proceedings, contracts, negotiations and so on. For now, the available AI technology does not cope with these tasks, and it is not certain whether it will ever be able to do so in a sufficiently reliable way.
Consequently, objectively, a large part of the hours that lawyers currently put in will be “eaten up” by AI, but there is nothing sad about that, because in practice we will free up time and energy to deal with the more essential part of our work. Thus, according to a study by Harvard University conducted among representatives of American law firms, the expectation is that the current ratio of lawyers’ work between gathering information and legal analysis will be reversed, from 80/20% to 20/80%[9]. In the same study, colleagues from the US state that the increase in productivity resulting from the introduction of AI systems will not reduce the number of billed hours, but will significantly increase their quality, and hence the demand for the services. The end result is expected to be more revenue for lawyers.
In fact, with the arrival of AI systems in legal work we will observe phenomena that are not without precedent. Until the introduction of the computer, a large part of a lawyer’s work consisted of typing or writing by hand. Statements of claim were much shorter than they are now, and so were court judgments. Did the computer take lawyers’ work away? No: it became easier to bring a case and to write a contract, and there was more work. The same applies to the introduction of legal information systems, which made case law and legislation far more accessible and thereby raised the quality of justice immensely. So too with remote access to case information and electronic case management, which spared litigators from queuing at the courthouse and allowed them to think more about what to write in their submissions, rather than how to serve them.
This essay, however, is not an apology for AI systems in law, nor a catalogue of the benefits of using them. It is a warning that, alongside all the benefits, there are very serious risks, and it is time to talk about them.
The Risks
A great deal has been written about the risks of using AI systems in legal practice. And that is normal: the legal profession is not just some business, but a key activity within the justice system, a pillar of democracy, legality and the rule of law. Any sudden changes should be received with scepticism and doubt. This is a healthy conservatism that we owe to every system that is key to the existence of critical functions of the state, as justice is. That is why it is understandable that the first reaction of the European legal profession was to warn of the risks.
I. The CCBE Guidelines
In March 2022, the Council of Bars and Law Societies of Europe (CCBE), in partnership with the European Lawyers Foundation (ELF), published the “Guidelines on the Use of AI-Based Tools by Lawyers and Law Firms in the EU” (the “Guidelines”)[10]. With this document the authors set themselves an educational aim: to give lawyers, especially small law firms, a practical understanding of the opportunities and risks of the new technologies, so that they do not fall behind in the digital era. The Guidelines were adopted before the explosion of ChatGPT, but by that time AI systems were already entering practice, and the analysis is still relatively up to date. The risks for the legal profession are divided into two groups:
1. Technological risks
These are the risks linked to the nature of AI systems, understood as a broader concept that includes language models but also other applications of machine learning. Since I am not an engineer, I will only briefly list the most significant ones:
Dependence on the provider (vendor lock-in): This is the risk of difficulty in moving data from one cloud provider to another because of a lack of standardisation. Yes, uploaded documents and other files can be downloaded and moved. But the point is that the valuable information that accumulates from the lawyer’s interaction with the system - the history of actions, links between documents, logs, metadata and, in general, all the information on the basis of which the system individualises the user and personalises its results - is lost. This technical feature can limit the lawyer’s freedom and make him or her dependent on the provider.
Lack of explainability: The lack of clarity about how exactly the algorithm arrives at a given result makes it impossible for the lawyer to check the logic behind it. A predictive-justice tool gives a forecast of the outcome of a case, but cannot explain exactly which factors or precedents tipped the scales for that conclusion. It must be said that since 2022 much has been done on the so-called “explainability” of AI systems, and popular chatbots provide some information about the steps they take to arrive at their answers. But it should be clear that this is not the complete account of the path the information takes.
Data privacy risks: In principle, AI systems can be trained on the data that their users, including lawyers, enter in order to use them. This means that, in theory, they can collect information about a lawyer’s clients and, once enough data has been aggregated, recognise them, which can be very dangerous. In principle, there are methods for extracting the information on which a model was trained through special queries (prompts)[11]. According to the Guidelines, this risk exists even when models are fed anonymised documents from which the parties’ data has been erased. With the help of the rest of the content, combined with information on the internet, the data can be de-anonymised and the person identified[12].
The risk of not understanding the fundamental limitations of generative AI: This risk is not included in the CCBE Guidelines, but it was quickly recognised in the years after ChatGPT became popular. Quite simply, language models are not reliable enough to be used for any kind of legal work. Their limitations and errors (e.g. “hallucinations”, false confidence, etc.) must be understood, and the results of their work must be verified[13].
And if these risks can be mitigated by technical means (and probably to some extent have been), the next group of dangers depends on the behaviour of lawyers.
2. Risks to professional duties
These are the risks that directly threaten the lawyer’s ethical standards:
Risks to professional competence when a new technology is introduced: The duty of competence requires a lawyer not to take on work for which he or she is not qualified[14]. This does not change because the lawyer relies on a technological tool to supply the answers, which the lawyer may not fully understand. If a lawyer is not competent in a particular field, the risk of making a mistake at work is greater with the use of AI than without it, because language models create an illusion of competence.
The risk of overload and inability to serve clients: The efficiency of AI tools can create an illusion of unlimited capacity, tempting lawyers to take on far more clients than they can actually serve to a good standard. The legal profession requires a personal understanding of the client’s needs, which cannot be fully automated. Indeed, turning legal services and client service into something robotic is alien to the humane essence of the legal profession.
Risks to professional secrecy: A lawyer has a duty to protect the secrets of his or her clients. The lawyer must take into account the risk that they may reach third parties (e.g. state authorities) when using AI applications. The lawyer must assess not only the technical parameters of the software, but also the strategic risk for the client: who might be interested in the data, and what are the legal powers of the authorities to demand access directly from the provider?
These risks are particularly heavy for smaller law firms and solo practitioners who, on the one hand, are pressed by the need to use AI systems so as not to fall behind. But on the other hand, they cannot invest in advanced applications, professional deployment and training.
II. Practical consequences
Listed like this, these risks sound slightly abstract. I would like to bring the narrative closer to the everyday work of the lawyer. What can happen if these risks materialise, and why is it dangerous for us?
1. The obvious: mistakes can happen.
A few days ago I read a study by the Centre for the Future of the Legal Profession (Australia), which surveyed the known mistakes made in court proceedings in 10 jurisdictions as a result of using AI systems[15]. The mistakes are of every kind: false facts, non-existent legislation and case law (“hallucinations”), drastically inaccurate interpretations of the law, and so on. The main cause is the use of AI without verification, i.e. without checking whether the results of the analysis are correct.
This can mean three things: either the lawyer is not competent and cannot judge whether the answer that the application gives is legally sound, or decides to save the time of reading and verification[16], or both. Each of these options flatly contradicts legal ethics and undermines trust in the legal profession. And trust is an essential component of the value of a lawyer’s work.
2. Devaluation of lawyers’ work
In fact, the benefits of AI systems listed earlier will improve the efficiency and value of lawyers’ work only if they are used for their intended purpose, namely to save the lawyer time that can be invested in activities that are genuinely useful to the client and to justice. If, instead, AI tools are used to feign competence or to slack off, the consequences for the legal profession can be very serious.
In fact, in the study cited in the previous section, a large share of the blunders in court proceedings result from citizens’ wish to represent themselves without a lawyer, with only the help of artificial intelligence. We can only speculate how many cases of self-help outside court proceedings there are, what they are like, and when their negative consequences will emerge. It should be clear what this means: some people simply do not see where the lawyer’s added value lies[17]. This feeling is reinforced if they think the lawyer is just an expensive intermediary between them and, say, ChatGPT.
There is, of course, an argument here that the market will weed out those lawyers and law firms that have no added value and merely relay information from the client to the smart application and back. There is some logic to it, but it is not certain. First, even the biggest can make mistakes, or it may simply show that they used AI without putting in any effort[18]. And second, it is enough for a few big blunders to become public for trust in everyone to suffer. Will clients then still be willing to pay for legal services, or will those services be devalued?
3. Consequences for justice and the legal order
If the role of the legal professional is marginalised and he or she becomes estranged, even partially, from legal dealings, dangerous consequences will follow for justice and for the legal order as a whole. Alongside the undisputed function of the lawyer to help citizens and businesses exercise their rights, the lawyer has the no less important function of “filtering” their wishes in accordance with the law, morality, practice and foreseeable human behaviour. This filtering also happens by virtue of the purely human limits on the provision of legal services, which are limited in time. Generative AI models, for their part, are notoriously inclined to flatter their users, including by giving them wrong and unfounded results[19]. And they can produce millions of “legal products” every day without getting tired.
In practice this means that if people start using chatbots or other applications to write statements of claim and complaints, the court will be swamped. And the court has virtually no ability to use AI systems for applying the law, as these fall into the high-risk category and must go through regulatory approval[20]. The effects on out-of-court dealings may also not be in the interest of the lawful course of legal relations. Who knows how many contracts are already in circulation that are full of nonsense, without the parties to them having any idea. We can assume that there is an enormous number, judging by the fact that OpenAI recently changed its policies regarding the provision of legal services in order to avoid liability[21].
What Is to Be Done?
After this, I hope not overly lengthy, exposition, two questions remain, directed at the Bulgarian legal profession: should we self-regulate in some way, and how should that be done?
I. Should we self-regulate?
I believe that, to answer this, we need to answer two sub-questions: 1) are there important values for the legal profession and for society that must be protected when lawyers use AI systems? and 2) Is it appropriate for this to be done through (self-)regulation?
1. Interests to be protected
From my analysis so far, building on existing guidelines and practices, I believe that several issues have emerged that are critical to the good of the legal profession, its clients and the justice system as a whole. Among them should be:
- Preserving trust in the legal profession
- Ensuring the ability to perform the lawyer’s duties
- Preserving the competitiveness and relevance of lawyers’ work
In fact, each of these values is already at least partly regulated. These are the duties, already mentioned, not to take on work without the necessary competence; the lawyer’s duties are listed in the Bar Act and the Code of Ethics, and Articles 27 and 28 of the Bar Act also impose on us a duty to keep improving our qualifications, including with the help of our governing bodies. Why, then, do we not simply apply the old rules in the new context? Is there any point in additional self-regulation?
2. Justification for regulation
I argue that there are reasons that make self-regulation of the legal profession necessary with regard to the use of AI, rather than alternatives such as education, for example.
2.1. Some new rules
The interests listed earlier cannot be protected by the current professional rules alone. To maintain trust in the legal profession, it must be mandatory to inform clients about the use of AI systems. They must be able to find out how this affects pricing, what additional work is done with the time saved by the technology so that the price remains the same, and what measures are taken to guarantee the confidentiality of the information. Otherwise, trust will decline.
Other rules can also be introduced, such as those providing for special security and traceability measures and other safeguards when handling AI systems in legal practice[22]. The topic is open and constantly evolving.
2.2. Interpreting the existing rules in the context of AI
Yes, most issues are already regulated in general terms, but that does not mean that their specific interpretation in the context of AI is clear. For example, does using an information application compensate for the lawyer’s original lack of competence in the relevant field? Is uploading client documents to general-purpose generative models a breach of professional secrecy? What about transcribing an online meeting with a dedicated bot? Must lawyers maintain their competence on technological matters too, such as the use of AI applications? And should the lawyer’s professional liability insurance also cover liability for damage related to the use of a smart assistant?
For all this, binding rules are needed, or at least interpretations made officially by the governing bodies of the Bar. The UIA Guidelines are a very good document, but uploading them to the website of the Supreme Bar Council binds neither the disciplinary bodies of the bar associations nor, still less, the court to these rules.
2.3. The need for a clear sign that the legal profession stands united in defence of trust in the profession
Such regulation should also send a very clear signal to all lawyers and to society as a whole that the legal profession collectively guarantees trust in the profession. That there are some institutional guarantees of competent and reliable legal services, of effective performance of lawyers’ duties and of adequate, good-faith and intended use of AI systems.
II. How should it be done?
This question is open. It is obvious that the grounds for regulation, the interests and values to be protected, and the specific rules to be adopted are subject to debate. The debate may go on for a long time, and it is good for it to begin very soon, since the technology is entering the profession exponentially.
From there on, there are many different forms of self-regulation. But it seems appropriate for the new rules to be introduced by amending the Code of Ethics, and for the existing rules to be interpreted and specified through guidelines, practical handbooks and the like, officially adopted by the governing bodies of the Bar.
Of course, the education and training of colleagues in working with AI-based tools is not just a possible solution but a mandatory one. It cannot, however, replace the guarantee effect of rules and the lawyer’s responsibility.
In conclusion, artificial intelligence is already here among us, part of our practice, whether we admit it or not. We cannot deny the enormous benefits it will potentially bring to the legal profession. A working life spent more on thinking about the law, about what is just, about philosophy and about life wisdom has been, more or less, the dream of every practitioner today. Behind this optimistic picture, however, lie risks that have the power to hit the profession of the lawyer hard. It depends on our awareness, organisation and proactiveness whether we develop and move forward as a guild or allow our prestige and the trust placed in us to decline.
Notes
- See, for example, the agents of A&O Shearman in cooperation with Harvey - link ↩
- https://www.perplexity.ai/page/legalzoom-perplexity-pair-cons-gcEFTf._TzigM1PYkUofIA ↩
- Here they are in Bulgarian translation - link ↩
- The so-called “low-hanging fruit” ↩
- For example, the COIN project of JP Morgan from 2017 - link ↩
- Such as OneDrive, SharePoint, Google Drive, Dropbox, etc. ↩
- For example, the software service Intapp - link ↩
- See in detail how this is done in: Takov, Kr. Как се решава частноправен казус (How to Solve a Private-Law Case). Sofia: Sibi, 2007 ↩
- https://clp.law.harvard.edu/knowledge-hub/insights/the-impact-of-artificial-intelligence-on-law-law-firms-business-models/ ↩
- CCBE/ELF Guidelines (PDF) - link ↩
- See, e.g., this article - link ↩
- See p. 47 et seq. of the Guidelines ↩
- See the Guidelines of the European Bars Federation from 2024 - link ↩
- Art. 43(1) of the Bar Act and Art. 11(4) and (5) of the Code of Ethics of the Lawyer (the “Code of Ethics”) ↩
- GenAI, Fake Law & Fallout: A review of the misuse of generative artificial intelligence in legal proceedings - link ↩
- Some argue that the need for verification eats up the benefit of AI - link ↩
- Such as this case of an appeal against an act of the National Social Security Institute (NOI), in which the appellant implicitly urges people not to use professional help - link ↩
- Such as this case of Deloitte, which was forced to refund money to the Australian government because of errors in a report prepared with AI - link ↩
- The so-called sycophancy effect - link ↩
- See Art. 6(2) in conjunction with Annex III of Regulation (EU) 2024/1689 (the Artificial Intelligence Act, or AI Act) ↩
- They changed only their policy, not the possibility for any kind of legal services to be generated through ChatGPT. The aim is to allow themselves to claim lack of fault in the event of harm - link ↩
- See, e.g., Art. 12 of the UIA Guidelines cited at the beginning ↩