Regulation 10.2.2 of the DIFC Data Protection Regulations requires a notice when an application or website service that uses an AI system to process personal data is used. Paragraphs (a) and (b) say what that notice must contain. Paragraphs (c) to (g) list material the firm must be able to produce on request. Keeping the two apart is the first drafting decision, and the one that decides how long and how candid the notice has to be.
This article reads the published DIFC instruments at a general level for orientation. It is not legal advice. Whether Regulation 10.2.2 applies to a given tool, and what its notice should say, depends on the firm's systems, data and use, and should be decided on its own facts.
When the notice is required
Regulation 10.2.2 opens: “Where an application or website service employing Systems to Process Personal Data is used, the following actions must be undertaken by a Deployer or an Operator in respect of such Processing”. Three terms in that sentence carry the weight.
A System is defined in Regulation 10.1.1(a) as a machine-based system operating in an autonomous or semi-autonomous manner that can process personal data for human-defined purposes, purposes the system defines itself, or both, and generate output on the basis of that processing. The guidance note to that definition says it is not intended to capture purely automated systems “which have no degree of autonomy in their operation and whose operation is deterministically controlled by humans”. A rules engine with fixed logic may fall outside it. A language model, a scoring model or an agent that drafts, ranks or classifies on its own is, on our reading, the kind of tool the definition is written for.
The duty falls on a Deployer or an Operator. A Deployer, under Regulation 10.1.1(b), is the person under whose authority, on whose direction or for whose benefit the System is operated, or who receives the benefit of its operation or output, in each case without regard to whether that person operates, supervises or hosts it. A DIFC firm offering a vendor's AI tool through its own client portal would ordinarily be the Deployer, even though the vendor runs the model. Regulation 10.3.4 deems the Deployer to act as a Controller and an Operator to act as a Processor for the purposes of Regulation 10 and the Law.
The Regulations do not define “application or website service”. A chatbot on a public website and an AI feature inside a client app sit squarely within the words. An internal tool used only by staff is a harder case, and the answer should be written down rather than assumed. The guidance note to paragraph (a) gives the purpose: individuals “who may be impacted by the use of AI in the processing of their personal data” should be made aware of it and given enough detail to assess the risks.
The alert in paragraph (a)
Paragraph (a) requires notice “in clear and explicit terms upon the initial use of, or access to, the System”. It must do two things.
First, it must alert users “to any underlying technology and processes comprising the System that may undertake any Processing of Personal Data by the System that is not human-initiated, controlled or directed”. The examples in the text show what the regulator wants to know: whether the System is restricted to specific human-defined purposes, or whether it can define further purposes on its own or otherwise process personal data for purposes that are not human-defined. A notice that says only “we use AI” does not answer that question.
Second, it must indicate “the impact of the use of the System on the exercise of individual rights as provided under the Law in Article 29(1)(h)(ix)”. That provision covers processing that will restrict or prevent a person from asking for rectification or erasure under Article 33, or from objecting under Article 34. If a model cannot remove a person's data once it has been used, the notice is where that is said. The Commissioner's Regulation 10 FAQ (DIFC-DP-GL-24, Q8) puts the emphasis in the same place, saying the notice should set out appropriate elements “particularly around the potentially negative impact on the ability to exercise their rights to control their data”.
Timing
On our reading, “upon the initial use of, or access to” places the notice at the point of first contact with the System, not in a privacy policy reached through a footer link after the fact. Article 31(3) of the Law lets a Controller meet its Article 29 and 30 duties by clearly directing the person to publicly available policies; Regulation 10.2.2 contains no equivalent. For a chat interface, that suggests the first screen. For an AI feature added to an existing app, it suggests the first time a user reaches the feature, which may be well after the account was opened and the original privacy notice was given.
The five descriptions in paragraph (b)
Paragraph (b) requires the same notice to include “a comprehensive, true and plain description” of five things. Each is quoted, in whole or in part, then applied.
The human-defined purposes
(i) “the human-defined purposes for which Personal Data is Processed by the System”.
The guidance note to paragraph (b) explains the distinction the Regulation draws. Purposes are either human-defined, meaning “externally pre-defined by humans and ‘hard coded’ into the System, which the System cannot change”, or self-defined, meaning purposes the System can generate itself. This item lists the first kind, in terms a user can follow: what the tool is for, and what it does with the person's data.
The principles and limits for self-defined purposes
(ii) “all human-defined principles on the basis of which, and all human-defined limits within which, the System is capable of itself defining further purposes for Processing of Personal Data”.
The guidance note says self-generated purposes “must be contemplated on the basis of an exhaustive set of detailed principles” that are themselves hard coded and that the System cannot change. Regulation 10.3.2(a) makes a similar requirement a condition of commercial use. For an agent that can choose its own next steps, this item is where the firm states the boundaries. If the firm cannot describe them, that is a finding about the tool, not a drafting problem.
The output and how it is used
(iii) “the output which the System produces on the basis of such Processing and the manner in which such output is used”.
The second half matters as much as the first. A model that produces a risk score is one thing. A risk score that decides whether an application proceeds, or that a person reviews before deciding, is what the user needs to know. Where the output feeds automated decision-making, Article 29(1)(h)(vi) of the Law separately requires meaningful information about the logic involved and the significance and possible outcomes for the person.
Design principles and built-in safeguards
(iv) “the principles on the basis of which the System has been developed and designed to operate, including a description of any safeguards built into the System by design to ensure compliance of the Processing of Personal Data by the System with the Law and this Regulation 10”.
Regulation 10.3.1 sets the design concepts a System that may impact a Data Subject must follow: ethical, fairness, transparent, secure and accountability. The transparency limb requires processing to be “explainable to Data Subjects and other stakeholders in non-technical terms, with appropriate supporting evidence”. This item is the user-facing version of that evidence. For a vendor model, the firm will need the vendor to tell it what the safeguards are before it can describe them.
Codes, certifications or principles
(v) “the codes, certifications or principles upon which the System is designed or developed”. The text gives examples that may be used, including those of the Dubai Digital Authority, the OECD, UNESCO, the NIST AI Framework, and the Guidelines for Financial Institutions adopting Enabling Technologies published by the Central Bank of the UAE, the Securities and Commodities Authority, the DFSA and the FSRA.
The FAQ (Q19) notes that Regulation 10.2.2, “which requires that detailed notice is provided, in effect addresses similar concepts pertinent to a conformity assessment (see 10.2.2(b)(v))”. Naming a code is a representation. It should be one the firm, or its vendor in writing, can stand behind.
What stays on file instead
Paragraphs (c) to (g) are not notice content. They are things the Deployer or Operator must be able to provide on request:
- (c) evidence of the System’s compliance with any audit or certification requirements the Commissioner establishes, on request by any affected party.
- (d) evidence of algorithms that make the System seek human intervention where its processing may have an unfair or discriminatory impact, with a risk and impact assessment of unjust bias or High Risk Processing, on request by any affected party.
- (e) and (f) evidence of human intervention algorithms, with risk and impact assessments, for access by competent government authorities, including law enforcement, and for possible non-compliance with Regulation 9, on request by any relevant party.
- (g) a register, on request by any relevant party, covering among other things use cases, necessity and proportionality, how Data Subjects can access information in the System under Articles 32 to 40 of the Law, whether the System is used solely to make automated decisions, third parties and transfers.
Paragraph (h) allows material provided under (c) to (f) to be redacted or summarised, “solely to the minimum extent necessary” to protect intellectual property in the System or its training data or to comply with legal restrictions, on condition that the full and unredacted material is provided to the Commissioner on request. The FAQ (Q9) says that, unless the Commissioner specifies otherwise, this evidence can take the form of policies, governance frameworks, principles frameworks and other documentation. Paragraph (h) also requires any other information the Commissioner requests to demonstrate compliance with the Law, the Regulations or other applicable laws.
The practical consequence is a two-layer design. The notice is short enough to read at first use and complete on the points in (a) and (b). Behind it sits a file the firm can hand over when an affected party asks. A notice that tries to carry the register is unlikely to be read, and a register that does not exist leaves the firm unable to answer a request.
The Article 29 notice alongside it
Regulation 10.2.2 does not replace the Law's general notice duties. Where a Controller collects personal data from the person, Article 29(1) requires information “at the time of collecting the Personal Data”, including the purposes and lawful basis, the recipients, any transfers, and the further items in Article 29(1)(h). The AI notice adds to that. In our view the two can be combined in one document, as long as the AI elements are delivered at the point paragraph (a) requires.
Article 29(1)(h)(ix) carries the most weight for AI. Where the Controller intends to process personal data in a way that will restrict or prevent rectification, erasure or objection, it must “include a clear and explicit explanation of the expected impact on such rights” and “satisfy itself that the Data Subject understands and acknowledges the extent of any such restrictions.” An acknowledgement implies a record of it. Article 33(4) then provides that where rectification or erasure is not feasible for technical reasons, the Controller is not in violation for failing to carry it out if it collected the data from the person and the Article 29(1)(h)(ix) information was explicit, clear and prominent with respect to the manner of processing and expressly stated that rectification or erasure would not be feasible. Article 34(3) makes a parallel provision for objections: where the Controller collected the data from the person and can demonstrate that the Article 29(1)(h)(ix) information was explicit, clear and prominent with respect to the manner of processing and expressly stated that an objection could not be implemented, it may continue processing in the same manner, subject to the Law in all other respects.
One gap is worth noticing. Article 30, which applies where personal data was not obtained from the Data Subject, contains no equivalent of Article 29(1)(h)(ix), and both Article 33(4) and Article 34(3) are limited to data the Controller collected from the person. A firm that feeds a model with data from third-party sources should not assume the same notice will cover it.
Schedule 2 of the Law lists a maximum administrative fine of USD 75,000 for failing to comply with Article 29. That is a ceiling, not a tariff.
When a notice misleads
Regulation 6.2.1 allows the Commissioner, where a complaint is made that a Controller or Processor has engaged in unfair or deceptive behaviour, to investigate and take enforcement action as necessary. Regulation 6.2.2 says such practices may include “misleading notices of processing activities or public representations regarding certifications or adherence to principles, codes and compliance standards”, particularly in the context of obligations that include Article 29(1)(h)(ix) and Regulation 10.
That gives paragraph (b)'s word “true” its edge. Three drafting habits create the risk: describing a System as human-controlled when it can define further purposes; describing output as advisory when it is applied without review; and naming a code or certification under item (v) that the System was not in fact designed against.
A working checklist
For each AI tool reached through an application or website service, a firm working through Regulation 10.2.2 would want, in writing:
- A decision on whether the tool is a System under Regulation 10.1.1(a), and whether it is used through an application or website service, with reasons.
- The Deployer and any Operator identified, and the vendor information needed for items (b)(ii), (iv) and (v) requested before launch.
- The notice text, placed at first use or access, covering the paragraph (a) alert and the five paragraph (b) descriptions in plain terms.
- The Article 29(1)(h)(ix) explanation where rights will be restricted, and a record that each person acknowledged it.
- A check that the notice, the Article 29 or 30 privacy notice and any impact assessment describe the processing the same way.
- The on-request file for paragraphs (c) to (g), with any redaction of the (c) to (f) material limited as paragraph (h) requires.
- An owner, and a trigger to revise the notice when the model, its purposes or the use of its output changes.
Where a tool is also used for High Risk Processing Activities, Regulation 10.3.3 adds conditions of its own: compliance with the Commissioner's audit and certification requirements, processing solely for human-defined or human-approved purposes, and an appointed Autonomous Systems Officer. Article 20 then requires an impact assessment before the processing begins; our reading of Article 20(6) covers that document. A written Regulation 10 baseline across a firm's AI tools, including which ones need a 10.2.2 notice, is what the two-week AI Governance Assessment produces, at a fixed AED 5,000, with a costed roadmap for what follows.