Using ChatGPT for Client Work in Quebec: Law 25 Guide

Executive Summary

Montreal freelancers and small studios that paste client material into ChatGPT, Claude or Copilot face three separate questions, and each has a different official anchor. This report, current as of November 2026, treats them as three gates. Gate A asks whether the material contains personal information. Quebec's private-sector privacy Act, as amended by Law 25, defines personal information as anything relating to a natural person that directly or indirectly allows the person to be identified ^[1]. Before such information is communicated outside Québec, the enterprise must conduct a privacy impact assessment ^[1], the transfer may proceed only if the assessment finds adequate protection ^[1], and a written agreement must follow ^[1]. The Commission d'accès à l'information (CAI) lists cloud storage of personal information outside Québec as an example of a project needing such an assessment ^[2]. The regulator guidance cited here does not say in terms that a chatbot prompt is a "communication" under that section, so the conservative reading is an interpretation, and this report labels it as one.

Gate B asks what to tell the client. No general Canadian AI statute sets a disclosure duty ^[3]. The duty comes from the client's own privacy obligations, from the contract, from Civil Code good faith ^[4], and from regulator principles that ask for openness ^[5] and for labelling of AI-assisted content or decisions ^[6]. Gate C asks who owns the output. Canadian practice ties authorship to a human exercising skill and judgment ^[7], consultation participants opposed protecting AI content without sufficient human contribution ^[7], and an assignment is valid only if written and signed ^[8]. A freelancer should therefore assign only the rights that exist.

The stakes are practical. In the 12 months before the second-quarter 2026 survey, 19.2% of Canadian businesses reported using AI to produce goods or deliver services ^[9], and 26% of self-employed workers had used generative AI at work in the 2024 to 2025 survey period ^[10]. Penalties are real but scaled: individuals face penal fines of $5,000 to $100,000 ^[1], and other enterprises face administrative penalties up to the greater of $10,000,000 or 2% of worldwide turnover ^[11].

The recommended workflow is short: classify the input, choose a tool tier whose documented terms exclude training and offer a data processing addendum, record a proportionate assessment, write the permission into the contract, and assign only protectable rights. This is general information, not legal advice; a Quebec lawyer should adapt the clauses below.

Introduction and Background

Generative AI moved from novelty to routine tool inside two years. Statistics Canada's March 2026 Labour Force Survey supplement found that generative AI tools were by far the most commonly reported AI or automation technology among workers ^[12]. For a freelancer, that routine use now meets a Quebec privacy Act written to cover any person carrying on an enterprise ^[1]. The CAI states that the definition extends to the self-employed individual (travailleur autonome), with a case-by-case analysis ^[13], so a business with one owner is not automatically outside the Act.

Three points frame what follows. First, the Act applies whether the enterprise keeps information itself or through a third person ^[1], which describes a prompt travelling to an AI vendor. Second, the person with the highest authority is by default the person in charge of compliance and remains accountable even after delegating ^[14]; for a sole proprietor, that is the sole proprietor. Third, the regulators say their generative AI principles apply to "individuals acting on behalf of an organization" using a generative AI system ^[15].

Two roles often blur in freelance work. When a client organization hands a freelancer personal information, the freelancer is a service provider and the client remains accountable for the information ^[16]. When the freelancer then sends the material to an AI vendor, a second link is added to the chain, and the question is whether the first contract and the Act permit it. This report reads only documented vendor data terms as of November 2026, records "check vendor's current terms" wherever a page was silent, and does not cover tax, employment-law or professional-order rules.

Key Changes: The Three Gates for Client Work

The 60-Second Triage

Run the three gates in order for each new client file.

• Gate A (personal information): Does the material identify or could it identify a natural person, such as customers, staff, patients or students? If yes, stop and apply the matrix below before any paste.
• Gate B (disclosure): Does the contract or the client's own policy say anything about subcontractors, tools or AI? If not, add wording and tell the client before the first use.
• Gate C (ownership): Will the client expect exclusive, registrable rights in the deliverable? If so, document the human contribution and avoid promising more than the law provides.

Table 1 below converts Gate A into a decision matrix. Tool tiers are described by their documented terms, never by assumption. The statutory trigger is location outside Québec, not outside Canada: the Act speaks of communications "outside Québec" ^[1], so a server in another province is still outside the province for this purpose.

The matrix is a risk filter, not a legal opinion: the more identifiable the information and the less documented the vendor's handling, the more work shifts to documentation and contract. The tiers differ chiefly in default training behaviour and in whether a data processing addendum (DPA, the contract that governs how a vendor handles customer data) is offered, as Table 3 later shows. Where a row says "allowed", the client contract must still permit it.

(Hypothetical Example) A solo bookkeeper receives a client's payroll export with employee names and salaries and wants an AI tool to summarize variances. Under the matrix, this is identifiable personal information: the bookkeeper would either strip names and employee numbers before pasting, or move to a business-tier tool after the client authorizes it and a short assessment is recorded.

Gate A: Personal Information, Outside-Quebec Transfers and Provider Contracts

What counts as personal information in freelance files

The statutory test is functional. Information is personal if it relates to a natural person and directly or indirectly allows that person to be identified. Typical freelance files that cross the line include customer lists, staff files, support-ticket exports, patient or student records and email threads with named individuals. Ordinary business content, such as a published report, does not. The Act scales security duties to sensitivity: the person must take the security measures necessary to protect the information ^[1], and assessments should be proportionate to sensitivity ^[1].

What Law 25 requires before the transfer

Three statutory pieces work together. First, section 17 requires the enterprise to conduct a privacy impact assessment before communicating personal information outside Québec, and it applies the same rules when the enterprise entrusts a person or body outside Québec with collecting, using or keeping information on its behalf ^[1]. The CAI dates the PIA-before-transfer duty to September 2023 ^[19]. Second, the transfer may proceed only if the assessment establishes adequate protection. The CAI's guide adds the corollary: if protection would not be adequate, the organization must refuse to communicate the information or refrain from entrusting it to a third party outside Québec ^[2]. Third, the communication must be the subject of a written agreement that takes the assessment results into account. A BLG analysis notes that the Québec legislator chose to rely mainly on contractual safeguards for these transfers ^[20].

For the service-provider relationship, section 18.3 allows an enterprise to give information to a provider for a mandate or service contract if the mandate is entrusted in writing ^[1]. The contract must cover four points:

• Use limitation: the information is used only to carry out the mandate ^[1].
• No retention: the provider does not keep the information after the contract ends ^[1].
• Breach notice: the provider notifies the person in charge without delay ^[1].
• Verification: the enterprise may carry out verifications ^[1].

For a freelancer, this is the clause set the client should already have given them, and also the checklist to apply to an AI vendor.

Does a chatbot prompt trigger section 17?

No statute, CAI page or law-firm summary read for this report addresses chatbot prompts specifically, so the reasoning is by analogy. The CAI's guide names cloud storage of personal information outside Québec as an example project requiring an assessment ^[2], and it singles out artificial intelligence, including generative AI, as a technology raising particular issues ^[2]. The CAI's January 2025 brief on AI at work likewise calls opacity around AI systems one of the major problems to correct ^[21]. Because a vendor that stores or processes prompts outside Québec fits the "keeps information through a third person" language, the prudent reading is that a prompt with personal information is a transfer.

What a short assessment should record

The CAI says a privacy impact assessment must be proportionate to the sensitivity of the information ^[22], and Norton Rose Fulbright reports that the CAI repeatedly stresses proportionality ^[23]. McCarthy Tétrault reports that the CAI's template is optional ^[24], and that the duty cannot be delegated to suppliers ^[24], so a vendor's trust page does not substitute for the freelancer's own record. A proportionate freelancer record can fit on one page:

• Purpose and necessity: why AI is needed for this task, and whether de-identified data would do. The regulators ask users to consider whether use is necessary and proportionate ^[15].
• Data inventory: categories of personal information, number of people, sensitivity, and the source client.
• Recipient and location: vendor, plan tier, documented hosting, sub-processors, and whether the vendor is outside Québec.
• Protections: training exclusion, retention period, encryption statements, DPA terms, and access controls on your own account.
• Decision: proceed, proceed with de-identification, or do not use. The CAI guide says that if necessity and proportionality cannot be justified, the project must be substantially modified or stopped ^[2].
• Review date: the assessment should be redone when the vendor's terms or your use materially changes.

Gate B: What to Tell Clients and When

No AI-specific disclosure statute

Osler states that there is no law in Canada that sets out a general framework for regulating AI ^[3], and that Canadian privacy laws are consent-based with limited exceptions ^[25]. The Quebec Civil Code gives indirect hooks. A service provider is free to choose the means of performance ^[4], which means the Code does not forbid AI tools, but must act in the client's best interests with prudence and diligence ^[4], and must provide the client, as far as circumstances permit, with any useful information about the task ^[4]. Parties must also act in good faith when an obligation arises and when it is performed. Applying these articles to AI tools is an interpretation: this report cites no Quebec authority that does so.

What the regulators' principles add

The joint principles of the federal, provincial and territorial privacy authorities, launched on 7 December 2023, include being open and transparent about how information is used and the privacy risks involved ^[5], making AI tools explainable to users ^[5], and limiting the sharing of personal, sensitive or confidential information ^[5]. The principles document adds that outputs with significant impact should be meaningfully identified as AI-created ^[15], that prompts should not be retained, used for secondary purposes or disclosed unless required ^[15], and that accountability rests with the organization and not with any automated system ^[15]. These are principles, not a statute written for freelancers.

Disclosure triggers in practice

Disclosure to the client is the contract's job. Three triggers are worth building in:

• Before first use on personal information: tell the client which tool and tier will receive the data, because the client's own written agreement and assessment obligations depend on it.
• When the deliverable is mostly AI-generated: the principles favour labelling where there is significant impact ^[15], and the Canadian Freelance Guild's Story Board advises freelance writers to make clear when and how they use AI ^[26].
• When the client's customers are affected by automated decisions: section 12.1 requires notice to the person when a decision is based exclusively on automated processing ^[1], and the person must be able to submit observations to someone able to review it ^[1]. This duty falls on the enterprise making the decision, so a freelancer building such a workflow should flag it to the client in writing.
• When the client asks: answer accurately; the good-faith rule applies.

One regulated profession shows the high end of expectations. The Barreau du Québec's guide for lawyers says the legal professional must obtain the client's free and informed consent before using generative AI on a file ^[27]. It binds lawyers, not freelancers, but shows how a Quebec professional regulator frames client consent. Quebec's Ministry of Labour guide on AI at work describes itself as non-prescriptive ^[28].

Gate C: Who Owns the Output

What Canadian authorship doctrine says

The Copyright Act makes the author of a work the first owner of the copyright ^[8], and an employer is first owner of works made by an employee in the course of employment, absent contrary agreement ^[8]. A freelancer is not an employee, so the freelancer is first owner of what they author, and an RGD article by a lawyer reminds designers that payment alone does not transfer ownership ^[29]. Educaloi, the Quebec legal-information non-profit, makes the same point for freelancers: without an assignment clause, the freelancer remains owner ^[30].

The difficulty is authorship itself. ISED's consultation paper states that case law suggests authorship goes to a natural person who exercises skill and judgment in creating the work ^[31], and that where output rests only on short user instructions, it is far less probable that the criterion would be met ^[31]. The "What We Heard" report (page modified 11 February 2025) records that existing jurisprudence suggests authorship must be attributed to a human who exercises skill and judgment ^[7], that stakeholders broadly supported keeping human authorship central ^[7], and that there was opposition to protecting AI-generated content without sufficient human contributions ^[7]. It also records support for flexibility for human authors to use AI as a tool ^[7].

What the ISED report does not say

The report is a record of consultation, not a rule. Its conclusion says only that one notable area of agreement was the need for substantial human authorship contributions ^[7]. It does not allocate ownership of AI output between freelancer and client, does not define how much human input suffices, and does not change the Act. Some participants proposed that the Canadian Intellectual Property Office (CIPO) require disclosure of AI-generated elements on registration ^[7], and some supported labelling of mostly or fully AI-generated content ^[7]; neither is law. CIPO's own guidance says registration is voluntary ^[32] and that it does not verify or examine the claims in applications ^[32], so a certificate does not settle protectability. A CIPO page relays the US Copyright Office view that substantively creative transformation of AI material can qualify ^[33]; that is a United States position, not a Canadian ruling.

The ownership risk ladder

Three tiers capture the practical position; the ladder is analytic, not an official classification.

• Fully human work, AI used only for research or proofreading: ordinary authorship rules apply, and a written assignment transfers whatever rights exist.
• AI-assisted with substantial human selection, arrangement and editing: protectable human contribution is plausible, but it must be documented. Fasken recommends that the human creative process be properly documented ^[34] and that human creativity remain at the centre of work leading to final products ^[34].
• Mostly unedited AI output: a freelancer cannot assume the client will own protectable rights. Lexpert cautions that assignment provisions cannot create copyright where none exists ^[35].

Trademark law is different: Fasken notes there is no human-authorship criterion under trademarks law ^[34], so an AI-assisted name or logo may still be protectable as a mark. Moral rights add a second layer. They cannot be assigned but may be waived in whole or in part ^[36], and an assignment of copyright does not by itself waive them ^[36]. RGD advises explicit written waivers for modifications ^[29].

Implementation Considerations and Process Changes

The Clause Menu

Law-firm commentary is consistent on direction. Lexpert recommends updating contractor agreements to capture AI-assisted work product ^[35], and Bennett Jones advises negotiating tailored protections on data use, residency, retention, security and confidentiality in AI vendor terms ^[37]. Table 2 turns those directions into a menu. The sample wording is illustrative drafting by this report's authors, offered for a Quebec lawyer to adapt; it is not quoted from any source. The source column anchors the legal reason for each clause.

The table is a starting set. The three clauses that carry most legal weight are the prohibited-input rule, the tool list and the incident notice, because they track what the Act requires of the client's contract with a provider. The ownership and warranty clauses matter commercially: the more a freelancer promises about originality or exclusivity, the more the contract assumes protectable rights that, on the ISED record, cannot be assumed for unedited AI output. A client may counter-propose a flat ban on AI use, and the freelancer should be ready to price that, since the Civil Code leaves the choice of means to the provider unless the contract says otherwise.

Tool-by-Tool Checklist

Vendor terms change, and the pages read for this report carry different dates, so the checklist records what each vendor's own page documented and leaves the rest as "check vendor's current terms". Table 3 summarizes the documented positions as of November 2026 and leaves a row for any tool not listed. None of this is a ranking.

Read across the rows, three patterns emerge, all documented rather than inferred. Business, enterprise and API tiers read here exclude training by default, while individual tiers differ in whether training is the default or a user choice. Canadian residency options are uneven: OpenAI lists Canada for storage at rest for eligible customers ^[40], Microsoft expects Canadian in-country processing in 2027 ^[52], and the Anthropic and Google pages read here list no Canadian option. And a DPA exists for several business tiers, but it is one input to the section 17 assessment, not a substitute for it.

Incident Handling

Where a confidentiality incident presents a risk of serious injury, the enterprise must promptly notify the CAI and the persons concerned ^[1]. Every enterprise must keep an incident register, which the regulation requires be kept for at least five years after the body became aware of the incident ^[57], and the CAI lists an incident at a third party holding information on the enterprise's behalf among the circumstances to document ^[58]. Two habits follow for a freelancer: log any suspected leak involving client data, including a mistaken paste, and tell the client promptly, since the client carries the duty to the regulator and the contract carries the duty down.

When to See a Lawyer

A Quebec privacy or technology lawyer is worth the fee in five situations:

• Sensitive files: health, financial or children's information is involved.
• Regulated clients: the client is a public body or a regulated professional.
• Automated decisions: the work affects individuals through automated decisions.
• Heavy warranties: the client demands an uncapped indemnity or a flat originality warranty.
• Inadequate protection: an assessment concludes that protection abroad is not adequate.

The CAI itself warns organizations that cannot assess a technology adequately to seek specialist help for artificial intelligence issues ^[2].

Data Analysis and Evidence

Adoption: how many freelancers and firms are affected

Official statistics place the question in the mainstream. Statistics Canada's Canadian Survey on Business Conditions found that 12.2% of Canadian businesses used AI in the second quarter of 2025 ^[59], and 19.2% over the 12 months preceding the second-quarter 2026 survey ^[9]. Among businesses with one to four employees, the closest proxy for solo operators, 19.9% reported using AI ^[9]. The Institut de la statistique du Québec, adapting the same survey, reports that 12.7% of Quebec businesses used AI applications for production in the 2025 reference period ^[60]. One quarter of businesses (25.2%) planned to use AI in the third quarter of 2026 ^[61].

At the worker level, the Canadian Survey on Working Conditions found that more than one in five Canadian workers (22%) had used generative AI ^[62], with use higher among self-employed workers at 26% ^[10], and 21% in Quebec ^[10]. The Labour Force Survey recorded that 93.4% of workers were aware of generative AI tools in March 2026 ^[12]. Self-employed workers were 13.1% of all Canadian workers in March 2025 ^[63].

Université Laval's NETendances 2025 survey (October 2025, 993 Quebec internet users) found that 52% of Quebec internet users had used a generative AI tool ^[64], that 46% of users used it for work ^[64], and that 63% of users were very or fairly concerned about security and confidentiality of personal data ^[64].

Privacy sentiment and compliance awareness

Public concern is high. The Office of the Privacy Commissioner of Canada (OPC) found in a February to March 2025 poll of 1,500 Canadians that 88% had some concern about personal information being used to train AI systems ^[65]. In its January to February 2026 survey of 800 consumer-facing businesses, the share using AI in operations rose from 6% in 2023 to 16% ^[66], while only 2% used customer information to train an AI system ^[67]. Statistics Canada's Q2 2026 data show 13.4% of businesses cited cybersecurity or privacy concerns as a barrier to AI use ^[68]. These surveys differ in population and wording and should not be pooled into one rate.

Regulatory timeline and penalty scale

The Quebec government describes Law 25's entry into force as gradual over three years ^[69]. The milestones that matter here are:

• September 2022: first rules on communication without consent ^[19].
• September 2023: privacy impact assessment before communication outside Québec ^[19].
• 22 September 2024: data portability ^[19].

The penalty scale distinguishes individuals from other persons:

• Administrative penalties: capped at $50,000 for a natural person and $10,000,000 (or 2% of worldwide turnover) in other cases ^[1].
• Penal fines: $5,000 to $100,000 for a natural person and, in all other cases, $15,000 to $25,000,000 or, if greater, 4% of worldwide turnover ^[1].
• Punitive damages: where an infringement is intentional or results from gross fault, a judge awards not less than a statutory minimum ^[1].

None of these figures is specific to AI.

Implications and Future Directions

For freelancers. The cheapest compliance step is classification at intake. A one-page matrix applied to each file, plus a recorded assessment for any file with identifiable personal information, covers the Act's central expectations proportionately. Choosing a business-tier tool with documented no-training defaults and a DPA reduces the work, but does not replace the assessment.

For regulators and legislators. Torys summarizes the CAI's workplace recommendations as non-binding ^[70], including algorithmic impact analysis for high-risk AI activities in addition to the legally required assessment ^[70]. At the federal level, Bill C-36 was introduced on 15 June 2026 ^[71]; its first-reading text would repeal Part 1 of the federal private-sector privacy statute ^[72] and require a general account of an organization's use of automated decision systems ^[72]. Bill status should be checked on LEGISinfo, as the bill had not become law at the time of retrieval and Quebec's private-sector Act continues to govern Quebec enterprises.

For copyright. The consultation record identifies substantial human authorship as the one notable area of agreement ^[7]. Until the law changes, the safe contractual posture is to assign what exists and promise nothing about AI-generated elements.

What would change this analysis. Re-check these developments before relying on this guide:

• CAI statement: whether prompts to external AI services are communications under section 17.
• Civil Code authority: any Quebec authority applying articles 2100 to 2102 to AI tools.
• CIPO position: an official view on registering AI-assisted works.
• Residency: new Canadian data residency options from vendors.

2727 Coworking, the Montreal workspace provider publishing this guide, is not a legal adviser and supplies no AI tools; the compliance steps above belong to the freelancer and their counsel.

Frequently Asked Questions (FAQs)

Can I use ChatGPT with client data in Quebec?

Only after classifying the data. With no personal information and a client contract that permits subcontractor tools, ordinary care applies. With identifiable personal information, the Act requires a privacy impact assessment before communication outside Québec and a written agreement, and regulators advise against sharing sensitive or identifiable personal information with chatbots ^[18]. The OPC also lists that information entered may be used to further train the system ^[18].

Does Law 25 apply to freelancers?

The CAI says the enterprise concept extends to self-employed individuals, case by case ^[13]. A freelancer who is an enterprise must answer for the personal information they hold.

Do I have to tell my clients I use AI?

No source read for this report establishes a general statutory duty for freelancers. Disclosure follows from the contract, the client's own obligations, the openness principle, and the Civil Code duties of good faith and of providing useful information (see Gate B). Writing it into the contract is the safer path.

Who owns AI-generated work for clients in Canada?

A freelancer is first owner of what they author ^[8], but authorship ties to a human exercising skill and judgment. Ownership transfers only by written, signed assignment ^[8], which cannot create rights that do not exist.

Do I need a privacy impact assessment as a freelancer?

Where identifiable personal information leaves Québec, the Act requires one. The CAI says the depth is proportionate to sensitivity ^[22], and its template is optional ^[24], so a short record is the proportionate form for low-risk files.

Is Claude or ChatGPT better for client confidentiality?

Neither page set supports a ranking. Both vendors document no-training defaults for business and API tiers. For consumer tiers, OpenAI says it may use content to train models unless the user opts out ^[17], while Anthropic describes a user choice ^[49]. Retention and residency differ, so use Table 3 and read current terms.

What does Law 25 say about automated decisions?

Section 12.1 requires notice when a decision about a person is based exclusively on automated processing of their personal information. It is narrow: it does not cover AI-assisted drafting.

What should an AI clause in a freelance contract say?

See Table 2: permitted use, prohibited inputs, tool schedule, disclosure, originality, limited assignment, moral-rights waiver and incident notice.

Does sending data to ChatGPT count as a cross-border transfer?

The statute's trigger is communication outside Québec, and the CAI treats cloud storage outside Québec as an assessment project ^[2]. No source read states the point for chatbots, so treat it as a transfer unless a lawyer advises otherwise.

Conclusion

Three gates give a freelancer a defensible routine. At Gate A, identify personal information, assess proportionately, and put the written agreement in place before identifiable data leaves Québec. At Gate B, tell the client which tools receive their data and keep the contract's tool schedule current. At Gate C, assign only the rights that exist, document human contribution, and avoid warranties that assume protectable rights in unedited AI output. The official anchors are the Act's sections 17 and 18.3, the regulators' generative AI principles and the ISED consultation record. None of them is a freelancer-specific rulebook, and the gaps between them are where contract drafting and professional advice do their work.

The practical takeaway is modest and repeatable: classify each file, pick a tool tier whose documented terms fit the file, keep a short written record, and put the permission in the contract. A freelancer who does these four things can explain, on request, what was sent where and why, which is the question a client, an accountant or a regulator is most likely to ask. Because vendor terms and official guidance both continue to change, the record should carry a review date, and the first step in any new engagement should be to check whether anything in this guide has moved.

References and Verification Note

This report reflects the sources as consulted in November 2026, and the Act text as consolidated on the Quebec Official Publisher's site in that month. Open questions: any Quebec regulator statement on chatbot prompts under section 17, Anthropic's sub-processor list, and Canadian residency options for several vendors. Where a vendor page was silent, this report says "check vendor's current terms".