Legal privilege and AI: the Swedish Bar guidance

The duty of confidentiality applies whatever the working method. That is the starting point of the Swedish Bar Association's guidance on generative AI, and this page is about Swedish professional rules. The rest follows from that starting point: which tools may be used at all, what the advokat must have read, how the risk is assessed, and when the client must be asked. Here are the requirements in order, with what the firm actually needs to do for each one.

Updated by Nils-Erik Jansson

The starting point

The guidance on generative AI from June 2025 replaced the general advice issued the year before. It was written by a working group of advokater, the Swedish term for members of the Swedish Bar Association. Its core is in the section on confidentiality: an advokat has a duty of confidentiality for what has been entrusted to them in practice, the guidance calls such information client information, and the duty must be upheld whatever the working method [1]. Everything else in the guidance applies that sentence to a new tool.

The assessment of good professional conduct for advokater (god advokatsed) in an individual case is the firm's own. This page is a reading of what the guidance requires. It is not an opinion on your firm.

Requirement one: the tool must not reuse client information

Generative AI may be used only with tools that ensure client information is not disclosed in breach of the duty of confidentiality. In concrete terms, the tool must not use client information in prompts and other input for training, or to generate answers for anyone other than the advokat or anyone outside the firm's environment [1]. The advokat should also avoid tools that claim ownership of, or a right to use, input or output [1].

The firm reads the vendor's commercial terms and data processing agreement. The marketing material is no basis for the assessment. For Anthropic, the ban on training on customer content and the customer's rights to input and output are in the commercial terms, which apply to Team and Enterprise and do not apply to personal accounts [3]. Two further points to check: whether any right to anonymise data for research comes with the contract, and how content flagged by the safety systems is handled. See the FAQ on the Claude for Legal page.

Requirement two: the advokat has read the terms

As a general rule, every advokat should read, analyse and understand the terms of use, privacy policy and related terms for all generative AI they use, and may need help from IT or security specialists to understand them [1]. This is a personal requirement on each advokat, in addition to the requirement on the firm.

The firm has one person read the terms properly once and write a two page summary with references. The other advokater read the summary and know where the original is. The summary is dated, because the terms change.

Requirement three: the risk assessment is fact based and made per matter

When assessing whether client information is sufficiently protected, the advokat should weigh the risk of disclosure and unauthorised access, the sensitivity of the information, the difficulty of putting safeguards in place and the extent to which the safeguards make the work harder [1]. The assessment is made for each client, matter, task and tool. One assessment for every case is not enough.

The firm makes a general assessment of the tool (data processing agreement, transfer, retention, requests from public authorities, see the checklist) and then writes a short rule for each type of matter on what may be entered. The result is the list of matters kept out of the tool. The individual advokat makes the final assessment in the individual matter, with the list as support.

If no client information is entered, no consent is needed. If client information is entered, and contract terms and technical safeguards mean it is handled as securely as in other external IT services, consent is not needed either. If the advokat considers that risks remain that cannot be dealt with by contract or security measures, the client's consent must be obtained, and a clause in the firm's general terms of engagement is then not enough [1]. The guidance on external IT services says the same and gives AI functions as an example of where the question can arise [2].

The firm makes a consent assessment for each type of matter, linked to the risk assessment, and obtains the client's consent for the types of matter where the answer is yes. Consent is not required when the contract is right and the safeguards are sufficient. In my view, matters where the risk remains are better kept out of the tool than entered with consent.

Requirement five: responsibility for advice is unchanged

The advokat's responsibility for advice applies in full whether the advice is based wholly or partly on output from a tool. The advokat needs to know how the tools work and what risks they carry, and output must be reviewed critically, in particular sources, case law and statutory references [1].

The firm writes the review rule into its guidelines, follows it up with spot checks and makes sure the tool shows its sources where it can, so that review is possible.

Requirement six: guidelines and training

The firm should adopt and implement guidelines on which tools may and may not be used, how they may be used, which legislation may apply, fact checking, transparency and ethics. It should also make sure that staff have training suited to the firm's use. Guidelines and training should be kept up to date [1]. The guidance also points to the AI literacy requirement in the EU AI Act [1].

The guidelines follow the guide on an AI policy for the legal department, which works for law firms with the addition on consent. See also the guide on the EU AI Act.

The contract with the vendor

The guidance on external IT services from April 2026 sets out the requirements for the contract. Its annex contains model clauses for a supplementary agreement: that the vendor is aware of the confidentiality rules in the Swedish Code of Judicial Procedure, that the client information belongs to the firm, that it is stored so that it normally cannot be read in clear text and that access is logged, that the vendor informs the firm and objects when a court or public authority demands disclosure, and that subcontractors are bound by equivalent undertakings [2]. How the clauses compare with a US AI vendor's standard terms is covered in the guide on contracts with AI vendors.

The firm package under Claude for Legal is set up around the six requirements: data protection documentation, a draft supplementary agreement based on the model clauses, a consent assessment, guidelines and training. The data protection package provides the same independently of the tool. The assessment of good professional conduct in the individual matter is always the firm's.

Frequently asked questions

May an advokat use ChatGPT for client work?
Not with client information on a personal account. The guidance requires that the tool does not use client information for training or for answers to anyone other than the firm, and that the advokat has read and understood the terms [1]. A personal account is governed by consumer terms that the firm has not negotiated. The same tool under an enterprise agreement with a data processing agreement and the right settings can be a different matter. The contract and the settings decide the question. The brand name does not.
Must the firm tell the client that AI is used?
The guidance names transparency as one of the questions the firm's guidelines should answer [1], but it sets no general duty to inform. Consent is required only where risks remain that cannot be dealt with by contract or security measures, and in that case a clause in the general terms of engagement is not enough [1] [2]. My recommendation is to say it in the terms of engagement and to answer honestly when asked directly. It costs nothing and it answers the question before it is asked.
Do the requirements also apply to lawyers who are not advokater?
The guidance binds advokater (members of the Swedish Bar Association) and their firms. An in-house legal department, or a law firm without advokater, is not bound by the Bar's rules of professional conduct. It is bound by the GDPR, by the company's own confidentiality undertakings and by its contracts with customers and counterparties. Most of the requirements are sensible to follow anyway, because they describe what it takes to avoid disclosing what you have promised to keep confidential.

Sources

  1. [1]Vägledning om användning av generativ AI i advokatverksamhet, Sveriges advokatsamfund, 12 juni 2025, in Swedish; Swedish Bar Association guidance on generative AI, sections 2 to 5
  2. [2]Uppdaterad vägledning om användningen av externa IT-tjänster i advokatverksamhet, Sveriges advokatsamfund, 15 april 2026, in Swedish; Swedish Bar Association guidance on external IT services, points 9 to 12 and the model clauses in the annex
  3. [3]Commercial Terms of Service, Anthropic

Back to AI for legal

More under AI for legal