In August 2023, the Canadian Association of Law Libraries (CALL) Executive Board approved a proposal for the creation of a new AI Standards Working Group (AI WG). The AI WG was created as a sub-committee of the Vendor Liaison Committee (VLC).

This Guide is governed by Creative Commons license CC BY-NC-SA 4.0.

The Committee members are:

• Annette Demers, University of Windsor (proponent and co-Chair)
• Sandy Hervieux, McGill University (co-Chair)
• James Bachmann, University of British Columbia
• Katarina Daniels, Davies Ward Phillips & Vineberg S.E.N.C.R.L., s.r.l.
• Erica Friesen, Queen’s University
• Sarah Gibbs, Parlee McLaws LLP
• Bryony Livingston, Legislative Assembly of Ontario
• Anita Susac-Bilyk, Goodmans LLP

This group is well-positioned to provide guidance on the topic, as they have decades of combined experience with applying electronic legal research tools in legal practice.

The WG consulted widely in preparing its recommendations. The consultations included members of CALL and legal professionals, vendor representatives and related professionals.

The goal of this Guide is to shape informed consumers who will set base standards and expectations that all vendors of AI legal research and writing (LRW) solutions will aim to achieve.

As many Law Societies in Canada ¹ have stated, licensees who use artificial intelligence have ethical and professional responsibilities in so doing. Accordingly, this Guide was written to help users. Users include law librarians, judges, lawyers, administrators, paralegals, clerks, faculty, and students. The public and other users, such as legislators or policy makers may also benefit from the information provided in this Guide.

Users need to know what to look for, and what to ask for when:

• procuring
• learning how to use
• teaching and mentoring others on the use of
• searching or requesting information from
• employing the results of, and
• managing the risks of

AI applications designed to assist with LRW.

For those users who are judging a particular system for procurement or other purposes, we have also provided a box at the end of each section that allows users to score a system on those criteria and to provide other notes. A summary table of all scores is provided at the end of the document as well that allows users to provide weighted results. An excel spreadsheet has also been developed as a complementary tool to this Guide.

The content of this guide is provided for both AI users and vendors for information only. It is not intended as legal advice. Everyone uses AI products at their own risk and should undertake due diligence to ascertain any liability in using them. The authors of this guide make no warranty whatsoever with respect to the accuracy and reliability of information contained in this guide or any related content provided.

This guide is intended to complement other guidelines and best practices that may be provided in court practice directions, rules of professional conduct, legislation, standards, benchmarks and policies.

This guide does not endeavour to address all possible ethical issues pertaining to AI systems, such as concerns relating to the environment, human rights, etc. Users are encouraged to make their own inquiries.

Algorithm

“Mathematics and Computing. A procedure or set of rules used in calculation and problem-solving; a precisely defined set of mathematical or logical operations for the performance of a particular task.” ²

Artificial Intelligence

“Information technology that performs tasks that would ordinarily require biological brainpower to accomplish, such as making sense of spoken language, learning behaviours or solving problems.” ³

Artificial Intelligence System (AI System) ⁴

An Artificial Intelligence System is “any computing system using artificial intelligence algorithms.”⁵.

Data

Structured and unstructured information that is collected, stored, and analyzed to support decision-making.

Ethical Walls

Also known as “information barriers” or “firewalls”, are procedural safeguards used within organizations to prevent the exchange of confidential information between departments, teams, or individuals whose interactions could result in a conflict of interest or ethical breach. These barriers are designed to protect sensitive information, maintain client confidentiality, and uphold the integrity of professional practices by ensuring that only authorized personnel have access to certain information.

Foundation Models

"Foundation models are artificial intelligence (AI) models trained on vast, immense datasets and can fulfill a broad range of general tasks. They serve as the base or building blocks for crafting more specialized applications."⁷

Generative AI

Generative AI, also known as generative artificial intelligence, is a subset of AI that uses advanced models to create new content such as text, images, videos, and audio based on user inputs. These models, often deep learning algorithms, learn patterns and structures from large datasets and use this knowledge to generate original content.

Hallucination

A generative AI output “is considered hallucinated if it is either incorrect or misgrounded.” In other words, an AI system has hallucinated if it “makes a false statement or falsely asserts that a source supports a statement.”⁸

Large Language Model (LLM)

“A deep learning algorithm that uses massive amounts of parameters and training data to understand and predict text. This generative artificial intelligence-based model can perform a variety of natural language processing tasks outside of simple text generation, including revising and translating content. LLMs aim to produce the most probable outcome of words for a given prompt.”⁹

Machine Learning (ML)

Machine learning (ML) is a branch of artificial intelligence (AI) focused on enabling computers and machines to imitate the way that humans learn, to perform tasks autonomously, and to improve their performance and accuracy through experience and exposure to more data.¹⁰

Reinforcement Learning

“Reinforcement learning (RL) is a machine learning (ML) technique that trains software to make decisions to achieve the most optimal results. It mimics the trial-and-error learning process that humans use to achieve their goals. Software actions that work towards [the] goal are reinforced, while actions that detract from the goal are ignored.”¹⁴

Model

An AI model is a program that analyzes data to find patterns or make predictions.¹⁵

Natural Language Processing (NLP)

“Natural language processing (NLP) is a subfield of computer science and artificial intelligence (AI) that uses machine learning to enable computers to understand and communicate with human language.”¹⁶

Prompt

(See also “user input”).

Output

Content that is produced by an AI system in response to a prompt/input.

Retrieval-Augmented Generation (RAG)

RAG is an artificial intelligence technique that can improve the quality of generative AI output by allowing large language models (LLMs) to access specialized data sets (e.g. primary or secondary legal databases) outside of its training data sources.

User Input

Instructions or data provided by a user to an AI system which the system uses to produce an output. ¹⁷ (See also “prompt”.)

Vendor

A natural or legal person, public authority, agency or other body that sells or provides access to an AI system.¹⁸

All marketing and product documentation should be provided in plain language.

Marketing

Statements made in marketing should be true, accurate and transparent. Marketing should reflect realistic benefits and risks presented by the product.

When drawing on survey results to market the product, vendors should make available the questions asked, and methodology used.

Documentation for Prospective Users

Vendors should make available to all prospective users, documentation that sets out, in a transparent manner, all aspects of the AI system, as detailed in this Guide.

Documentation Designed to Assist Subscribed Users

Vendors should make available to all subscribed users, documentation that sets out, in a transparent manner, all aspects of the AI system, as detailed in this Guide.

Additionally, subscribed users should have access to tips for how to use the product.

All documentation and tips should be accessible from within the product (e.g. a “Help” page).

The documentation should outline the types of support available to a user, including, for example, access to live support from a human representative.

Users should recognize that most, if not all, AI applications for LRW are built on foundation models trained on the open internet. LRW applications should be built upon a transparent collection of legal information sources. Vendors should clearly provide information about the following:

1. What data sources are used by the AI system? Examples include:

2. Is the scope of each dataset provided? The scope should minimally include, where relevant:

3. What data is NOT included when producing output? For example, is the AI using or not using certain subscribed content, and is unsubscribed content excluded?

4. Do different services offered within the product also specify the scope of data (as described above) upon which each service is trained and from which outputs are constructed?

5. Have data quality standards been established and documented?

Bias

What, if anything is being done to mitigate bias originating from data sources? See also the section on Bias below.

The data on which an AI system has been trained affects the accuracy, validity, and relevance of AI outputs. Vendors should disclose the data from which tools draw, and the quality control mechanisms deployed to assess outputs and prevent errors. If this information is not known for foundation models, the vendor should clearly disclose this. Information on these points should include:

1. On what data is the AI being trained? For example, is the training data limited to content from legal databases (if so, which ones?) or is it drawn from external resources or the open internet? Can the user find out all the sources used in training? What gaps have been identified in the training data that might contribute to hallucinations or anomalous outputs?

2. Has the user been alerted to the possibility of incomplete, irrelevant, or hallucinated outputs resulting from gaps in the specialized training data?

3. Are user inputs used as data to train the AI system? If so, is the data anonymized or otherwise stripped of identifying data? Does the vendor provide the option to opt out?

4. Are the AI system’s own outputs used to train itself?

5. Was data labelling done by humans, by another AI system, or a combination thereof? Describe any processes, standards, or policies relevant to the labelling process.

Bias

What, if anything, is being done to mitigate bias originating from the training data and the training processes? See also the section on Bias below.

Vendors should provide a detailed overview of the AI system, including the components outlined below. If this information is not known for foundation models, the vendor should clearly disclose this.

Algorithms

What methods and algorithms are used to train and operate the AI system (e.g., natural language processing, retrieval augmented generation, type of learning (e.g., semi-supervised learning), transformer-based models, etc.)? Examples of related questions to ask:

1. How do the algorithms work?

2. How are the algorithms trained? For example:

3. What third-party product(s)/AI is the AI system based on (e.g. GPT-5, DeepSeek) and to what extent is the third-party product trained on specialized/subject specific-data?

Output Factors

Vendors should be transparent about how the AI system’s outputs are tested, such as:

1. What constitutes adequate versus inadequate output?

2. Are the testing instrument(s) and test results available? When did the testing take place? How representative of the current system configuration are the test results?

Hallucinations

For generative AI systems:

1. How are hallucinations defined by the vendor? Are there types of hallucinations that are not acknowledged by the vendor? See for example our definition of hallucination above.

2. Does the vendor specify that hallucinations are possible? If so, which type(s)?

3. What percentage of output is hallucinatory? Does (and if so, how does) this percentage vary depending on the type of prompt/task that the AI is asked to perform? Does (and if so, how does) this percentage vary depending on the type of hallucination in question?

4. What controls are put in place to prevent hallucinations?

5. What should a user do when a hallucination is identified? (e.g., does the vendor solicit feedback from users on instances of hallucinations?).

Bias

1. What steps have been taken to create algorithms that reduce and eliminate bias and discrimination?

2. To what extent and how are humans involved in reducing and eliminating bias in algorithms, data, and outputs?

See also: the section on Bias below.

1. Are there instructions embedded in the product to advise users about how to interact with the AI system effectively (e.g. best practices for crafting prompts)?

2. Are training materials and documentation available in a variety of formats (e.g. videos, interactive tutorials, checklists, guides)?

3. Is jurisdiction clearly addressed in the interaction (e.g. through a specific prompt or pre- or post-filtering options)?

4. Does the AI system allow users to create reusable or shareable personas to their accounts?

5. Does the AI system allow users to create prompt libraries or otherwise save or reuse user inputs?

6. Does the AI system provide information about further cases or related questions to explore?

7. Does the AI system learn from user inputs (e.g. From user prompts, from user feedback to outputs, etc.)? If so, how?

Credibility of Outputs

1. How does the AI system respond if it cannot produce an output? Does the AI indicate clearly that it cannot provide an output, does it provide an error message, or will it always produce an output?

2. Is a confidence indicator provided? If so, how is this calculated?

Attributing Generative AI Outputs¹⁹

Does the vendor provide sufficient information to cite AI generated outputs considering the requirements of law societies, court practice directions, etc. to provide attribution of AI generated outputs?

Referencing Sources

1. Are references consistently provided?²⁰ If not, is there a reminder or indicator provided to encourage users to verify the output?

2. Are sources objectively recognizable and verifiable? See also: Algorithm Transparency – Bias

1. Do the training materials, terms of reference, system reminders or warnings include:

2. Does the vendor provide documentation demonstrating compliance with the laws and regulations that apply to this AI system?

1. Does the vendor explain how user data will be used and protected, including:

2. For AI systems that integrate with other applications or software, including but not limited to:

does the vendor explain whether and/or how ethical walls or other conflict or security mechanisms will carry over to ensure that users are only able to see content to which they should have access? How often will the ethical walls or other conflict or security mechanisms synchronise across the systems?

If ethical walls or other conflict or security mechanisms cannot be carried over, does the tool allow for custom security measures? How are these managed?

1. Is there a policy that governs the use of copyrighted data used by the AI system?

2. Is there a policy that governs user uploading of copyrighted content to the AI system?

3. What is the vendor doing to protect users from inadvertently breaching copyright when using content produced by the AI system?

4. Does the vendor confirm that copyrighted materials have been used with permission? If this information is not known for foundation models, the vendor should clearly disclose this.

5. Is citation or attribution made available to users along with the outputs?

6. Does the vendor allow authors and licensed content providers to opt out of having their content used by AI systems?

1. Does the vendor acknowledge the existence of bias in the AI system and the risk of perpetuating or compounding bias?

2. What is the vendor doing to highlight or mitigate bias in the AI system? If this information is not known for foundation models, the vendor should clearly disclose this. For example:

3. Does the vendor ensure that the AI system complies with all legislative and regulatory requirements and conforms to relevant international standards?

4. Does the AI system block some prompts? If so, under what circumstances?

Users should be particularly wary of bias in tools that purport to discern subjective characteristics of a document, for example, truthfulness, tone and sarcasm.

See also:

--Algorithmic Transparency – Bias

--Data Transparency – Bias

1. Does the vendor have a public-facing policy, for both authors and for commentary generated in-house, on the use of AI in generating, updating or editing commentary products?

2. Does the publisher state clearly how AI is used in generating, updating or editing commentary products?

3. In such situations, how is AI usage attributed? See for example: COAL-RJAL Editorial Group, Canadian Open Access Legal Citation Guide, Canadian Legal Information Institute, 2024 CanLIIDocs 830, https://canlii.ca/t/7nc6q section 8.

Criteria	Score	Weight
Marketing and Product Documentation
Data Transparency
Data Used in Training (Training Data)
Algorithm Transparency
User Inputs / Prompt and Prompt Engineering
Outputs
Other Risk Management
Privacy
Copyright
Bias
AI Generated Legal Commentary