This document sets out the guidelines for the responsible use of Artificial Intelligence (AI) by MACEDO VITORINO.
MACEDO VITORINO is a Portuguese law firm that is implementing an AI multi-model software application, KeyTerms.ai, which integrates the following large language models (LLMs): Perplexity, ChatGPT and Grok 3.
LLMs show extraordinary linguistic capabilities, can access a wealth of information that far outpaces that of any human being. AI can change the way lawyers do their work.
However, not all is perfect with AI. AI poses significant risks to its users, especially to lawyers and law firms. Over-reliance on machine-generated work without proper human verification and validation may lead to errors with a significant impact on the end-product. Moreover, the use of AI generated text as one's own in the final work product, despite lacking sufficient human intervention, is also deeply unethical.
In addition, the use of AI may have pernicious effects on lawyers (specifically on young lawyers), disincentivise proper and careful research, hinder their ability to think about the problems and find grounded solutions, facilitate a non-thoughtful approach to difficult questions and promote overconfidence in AI's outputs against a human and organic approach.
The following guidelines establish a set-by-step process for ensuring that MACEDO VITORINO's lawyers and staff use AI responsibly, respect ethical principles, minimise risks, optimise results and improve the quality of the firm’s work.
BEFORE YOU START
Always use an AI software that does not compromise the safety and secrecy of information that you provide the model. Key Terms ensure confidentiality backed by the API agreements with OpenAI, Perplexity and XAi (Grok 3), which guarantee that no information fed to the model will be collected or used in training. However, the free versions of these models do not give this guarantee and warn users that they may collect and use information provided by the user. You should not use free AI software not approved by MACEDO VITORINO and never provide information belonging to MACEDO VITORINO or its clients. You may only use free applications for searching and reviewing information obtained from the Internet and never with information obtained from the MACEDO VITORINO network.
STEP 1: PLAN
Take some minutes to think about your goals. Make a few notes describing your objectives and the path for achieving them.
Use your notes to make a request (prompt) to Key Terms (Grok 3) asking for a structured plan with key discussion points regarding the subject matter of your work. The best prompts lead to the best results.
For example:
"Develop a plan for a study [opinion / information] on the regime of impossibility of performance in the Portuguese legal system, based on the Civil Code, relevant case law, and doctrine."
Play with the AI system, rephrase and finetune your prompt, ask for clarification or a more in-depth review when needed.
For example:
"I only intend to write a five-page article. Suggest a less ambitious plan."
Reserve 30 minutes for this phase if the matter is of medium complexity (final document expected to have about five pages length).
Insert the AI output in your document, review it carefully and draw up a list of the matters you wish to research.
STEP 2: RESEARCH AND ANALYSE
Use Key Terms (Perplexity) to research each relevant subject matter in your plan, focusing on the key terms and concepts from your list of research points. Reserve 30 minutes for this phase.
Make multiple questions around each subject matter until you reach results that satisfy your requirements. Then insert the results in your document.
At this stage, you should also verify the results from the AI. Check the answers provided by Grok with Perplexity, which is more reliable in finding accurate sources and provides hyperlinks to the web pages on which it based its responses.
STEP 3: VERIFY, STUDY, LEARN AND THINK
Verify the AI-generated results and all sources (weblinks). Reserve at least one hour to verify the AI sources.
Large Language Models (LLMs) can provide (and will provide) incorrect answers in many situations. The incorrect or inaccurate responses from LLMs stem from various causes: (i) a lack of quality information, which is common when there is insufficient information on the subject; and (ii) the model's attention mechanism being drawn to a larger set of data or information sources that appear to be related to the object of the instruction given to the model, even though they are not (for example, providing information originating from Brazil when the intent is to obtain information related to Portugal). The quality of the AI's output depends on the accuracy and quality of the data to which the model has access.
Thus, for example, in the response to the prompt we suggested above, the AI stated the following: "Portuguese doctrine plays a crucial role in the interpretation of the regime. Authors such as Antunes Varela, Pires de Lima, and Roque Nogueira offer valuable contributions on the impossibility of performance." We did not find any published work by Roque Nogueira on the subject. We assume the system was referring to Roque Nogueira, a Judge of the Supreme Court of Justice (STJ).
The reference to Antunes Varela is correct but generic. This matter should be explored in greater depth during the research phase.
This situation exemplifies the inaccuracies of AI. To correct them, you must question all responses. In part, your intuition (grounded in your own knowledge and experience) will be your starting point. The method will always involve a case-by-case verification.
Large language models (LLMs) can and will give wrong answers in many instances. Inaccurate and wrong answers result from various causes: (i) lack of quality information, which is more common when there is not sufficient information on the matter; and (ii) the model's attention mechanism being drawn to a larger dataset or information source (for instance Brazil instead of Portugal) that seems related to a query but lacks real relevance. The quality of the AI output depends on the accuracy and quality of the source.
For this reason, you must cross-reference the AI outputs with its sources. Visit all website addresses provided by the AI system to confirm the accuracy and relevance of the information.
Collect and organize the data obtained from those sources. Select and quotes to your document.
Learn from the sources. Keep windows of relevant websites (and PDFs) open for further use as you draft your document. Save windows and information for reuse if the document drafting will take more than one day.
STEP 4: REWRITE THE TEXT SUGGESTED BY AI
Rewrite the original AI-generated content, ensuring it reflects accurate and substantiated information, and add other information that you believe is relevant.
Use Key Terms (Grok 3) to review your text as you write it. Use Key Terms (Perplexity) to double-check the results whether or not they were provided by AI.
Use AI incrementally by focusing on a few paragraphs at a time (two/three at most) rather than requesting AI to write an entire document or a full chapter. This step-by-step approach will ensure that control the end result and that no wrong answers sneak into your document.
Review the AI generated content line by line, as AI results can contain errors in apparently correct answers/assertions.
Always mistrust the AI results.
You must possess or develop in-depth knowledge of the subject matter to confirm even the smallest details.
STEP 5: TAKE OWNERSHIP OF THE DOCUMENT
At this point, your document should have a significant amount of well-structured AI generated text including some inputs of your own.
Now is the moment to take full ownership of the document. Adapt the AI generated text to meet your specific goals and ensure that it aligns with your purpose and style.
When writing a legal article or an opinion, do not be satisfied with a well-written answer, make it your own by rewriting it. This will make you the true author of the text, make the text truly original, escape the robotic style of LLMs and correct imperfections.
STEP 6: FACT-CHECK
Once a pre-final draft is complete, use Key Terms (Perplexity) to fact-check the document thoroughly.
Review and address any flagged inaccuracies or inconsistencies.
Make a second review using Key Terms (Grok 3), which excels in reasoning and writing quality.
If your text is too long or confusing, Key Terms (Grok 3) will simplify the text a bit, create shorter sentences, and usually improve the text without compromising your authorship. In such cases, you will only need to make small adjustments to ensure everything is as you intend.
Carefully review and confirm the AI feedback to refine the document further.
STEP 7: BE TRANSPARENT ABOUT THE USE OF AI
If, after completing all reviews, you see that some parts of the content is mostly AI generated alert the firm’s team involved in the matter. Failure to disclose significant AI contributions in such contexts may undermine the quality of work and our professional credibility. Any unidentified error that ends up in a client work product may harm the firm's reputation and may entail professional liability.
Always highlight the sections with less human input, preferably using a colour scheme (yellow, blue, grey). This will alert reviewers to potential AI errors that may have escaped you.
Reusing past work has been a common practice long before the advent of LLMs in legal work. There is no need to fear the reuse of AI text.
However, the use significant portions of AI generated content in opinion articles or academic work passing as original and personal work is wrong. AI can (and should) be used in academic type of work in law firms, but copy-paste and plagiarism of AI generated content is unethical.
Being true and honest about the AI components of your work enhances trust in your work. If you can produce a full (legal) text using AI in most of it, this means that you have perfected your skills and have strong knowledge of the legal matters involved.
You can rely mostly on AI outputs when preparing contracts and other documents based on standardised text, that you would now cut and paste or repurpose from existing templates and forms. In those cases, do not worry about using the text produced by the LLM if you are satisfied with the result and have verified it against its sources and your own knowledge of the matter. Still, you need to highlight the AI portions of the text to the firm's team working on the matter.
Never misappropriate AI content as your own when writing articles and legal opinions.
© 2025 MACEDO VITORINO
IMPORTANT NOTICE
This document was produced with the assistance of KeyTerms.ai, a multi-model legal AI software that utilises the following large language models (LLMs): Perplexity, Grok 3 and OpenAI. All AI generated output was thoroughly revised and rewritten by the author.
The opinions expressed in this article are of general nature and should not be considered as professional advice. Should you need legal advice on these matters you should contact a lawyer. If you are a client of MACEDO VITORINO, you may contact us by email addressed to mv@macedovitorino.com.