Ukraine's Ministry of Digital Transformation has published informational recommendations, dated September 2026, that tell psychologists which uses of AI in psychosocial support are limited and which are impermissible, and the document binds nobody

Two dates, and neither is an entry into force

The cover gives the document one date: «Вересень 2026», September 2026. That is the only date the document gives itself. We searched the full text layer for a day of the month, an order number, a protocol number or an approval line and found none; the only numbered instrument in the file is the Cabinet resolution the document cites on p.23.

The 18 September date belongs to the ministry, not to the document. The ministry's news article, headed «Від щоденника емоцій до червоних ліній» (from an emotions diary to red lines, our translation), carries a machine-readable datePublished of 2026-09-18T15:31:29+03:00 and a visible publication stamp of 18 September, 15:31. Its last paragraph links the PDF at the storage URL we read. The storage object itself reports Last-Modified 18 September 2026 at 14:08:05 GMT, which is 17:08 Kyiv time, about 97 minutes after the article's timestamp; that tells us the object was written that day and nothing more.

So the honest statement is the one the desk uses here: the ministry published the recommendations on 18 September 2026, and the document is dated September 2026. Neither date is a commencement. There is nothing in the document to commence. It creates no obligation, sets no deadline and names no body that could enforce it.

What the document says it is, and who it says it is for

The disclaimer on p.3, headed «Застереження», opens by saying the recommendations are «призначені виключно для ознайомлення користувачів та психологів», intended exclusively for the information of users and psychologists in the field of psychological well-being and psychosocial support (our translation). The same paragraph says that the companies, products and platforms named in the text are illustrative and are not advertising, an official recommendation or solutions mandatory for use. That second sentence is about the product examples; it is the first sentence that fixes the status of the whole document.

The introduction on p.3 places the document inside a plan. It calls itself one of the components of the roadmap for AI regulation in Ukraine, and says it was formed on the bottom-up approach described in the White Paper: general and sectoral recommendations first, then the adoption of a law analogous to the European Artificial Intelligence Act (our translation). That is a description of intent about a future law. It is not evidence of what that law will say, and we do not treat it as such.

Attribution is thin and we report it as it stands. The p.4 sentence «Зміст є винятковою відповідальністю Міністерства цифрової трансформації України» makes the ministry solely responsible for the content (our translation). Twenty-one authors are listed alphabetically with no affiliations. The UK DIGIT project, implemented by the Eurasia Foundation and funded by UK International Development, is credited with support, and the same page says the views expressed do not necessarily reflect UK Government policy. There is no deputy-minister foreword, no signature block and no line approving or adopting this document anywhere in the text layer; the year 2026 appears exactly once in the file, on the cover.

The addressees are psychologists, users and, in a chapter starting on p.16, students and trainees without a professional psychology qualification. Employers, schools and app developers are not addressed by name anywhere. The ministry's news article describes the audience the same way: psychologists, and ordinary users who talk to chatbots about their emotions. The article also says the ministry developed the recommendations together with public- and private-sector experts, which is the ministry's statement and not something the document itself says.

One difference between the article and the document is worth marking. The press office writes that AI is «категорично заборонено» from making psychological or psychiatric diagnoses, categorically forbidden (our translation). The document's own row on p.21 says AI «не може самостійно встановлювати» such diagnoses, cannot independently establish them. The document's verb carries a qualifier the press release drops, and it is the document's verb we report.

Three lists with three different verbs

The section that matters starts on p.20 under the heading limits and impermissible practices. Its first table, in this non-binding guidance, is introduced with the words «має бути обмеженим», use of AI is to be limited (our translation), and lists eight areas of heightened risk: psychodiagnostics, mental-health assessment, work with children and minors, work with veterans, service members and people who have lived through traumatic events, crisis psychology, psychological rehabilitation, personnel selection and assessment, and expert work. In each row AI is allowed as an auxiliary tool and the professional act stays with the psychologist. The psychodiagnostics row, for example, permits AI for auxiliary processing of results and statistical analysis while reserving interpretation, conclusions and professional decisions exclusively to the psychologist's competence; all of it is advice from an informational document, not a rule anyone enforces.

The personnel row is the one an HR reader will want. Its opening words are «ШІ не повинен використовуватися як єдина підстава для ухвалення рішень», AI is not to be used as the sole basis for decisions (our translation), and the row goes on to name hiring, professional selection, promotion and other personnel decisions, adding that any results of automated analysis are, in the guidance's own wording, subject to obligatory professional review. That sentence sits in a document addressed to psychologists, in a table headed limited rather than impermissible, inside guidance that binds nobody. It is not a Ukrainian ban on algorithmic hiring and we do not describe it as one.

The second table, on pp.21 to 22, is where the informational document draws its red lines. It is introduced with «є недопустимим», is impermissible (our translation), and has eight rows: making a psychological or clinical diagnosis; automated decision-making about a person; conducting forensic-psychological examinations without an expert; carrying out crisis interventions without a human; assessing suicide risk as a final decision; manipulative psychological influence; covert psychological profiling; and forming emotional dependency on AI systems. The lead-in says that in these cases AI outputs may not be used as a basis for professional decisions regardless of the technology's level of development (our translation). That is the guidance's own voice; it binds nobody.

Every row of that table carries a qualifier, and the qualifiers are the content of this non-binding guidance. Diagnosis is barred where AI does it independently. Automated decision-making is barred where AI is «єдиного або остаточного інструменту», the sole or final tool, for decisions affecting rights, freedoms, social status, professional activity, education or access to services. Suicide-risk assessment is barred as an independent determination made without the direct participation of a qualified psychologist. The crisis row, which we carry as this article's quotation, bars AI as the sole means of help to a person in acute crisis. Strip the qualifiers and the document reads as a prohibition on AI in mental-health work; keep them and it reads as the ministry's recommended rule about who decides, with no legal force behind it.

The eighth row is drafted at the level of system design. It says, in this informational guidance, that it is impermissible to design or use AI systems aimed at forming a user's psychological, emotional or behavioural dependency, or at imitating interpersonal relationships in order to hold attention or to obtain commercial, political or other benefit (our translation). Developers are not named in that row, but the document does return to them elsewhere. On p.43 it says that if a developer does not provide information about a system's purpose, its sample, its accuracy indicators, its known limitations and its subgroup test results, the system should not be used for psychological assessment (our translation). On p.44, discussing a reported case of a user's attachment to a Replika companion, it says the situation raises ethical questions about developers' responsibility to vulnerable users who rely on AI for mental-health support (our translation). None of that is an obligation on a developer. It is advice to psychologists about what to demand of one, inside guidance that binds nobody.

The third list, on p.22, still inside the same non-binding guidance, is headed «До високоризикових або недопустимих практик належать», high-risk or impermissible practices include (our translation), and gives ten items without saying which are which: uploading transcripts of real sessions without client consent and legal and technical data protection; automatic generation of a psychological opinion; using AI to read emotions from face, voice or behaviour in an educational or work setting; using a chatbot as a replacement for a therapist; recommending that a client turn to AI instead of a psychologist in a crisis; assessing professional fitness with AI without validated procedures; using unvalidated AI tests; automatic risk-ranking of clients without a transparent methodology; using AI with children without parental or legal-representative consent and age adaptation; and covert use of AI in preparing psychological opinions where that matters to the client or commissioning party. The document does not sort those ten into the two categories its heading offers, and neither do we.

Disclosure before the tool touches the session, and a conditional bar on client data

The disclosure duty appears three times in this non-binding guidance and hardens as it goes. Under the transparency principle on p.24 the informational document says clients are to be informed, in a clear and accessible form, of any interaction with an AI system or use of one in preparing or providing their support, before it is used in the specific case, and told about the system's limitations and their own rights (our translation). The same principle asks psychologists not to use an AI system without the client's consent in due form in the cases established by legislation or professional standards, and to document AI use in the cases and manner provided by legislation, professional standards and documentation rules. That documentation line is conditional on other rules. The document does not create a stand-alone log-keeping duty, and it nowhere says that every structured review of an AI output is to be recorded.

On p.32, under transparency of use, the verb changes. The guidance says the psychologist is obliged to proactively inform the client of AI involvement where it is used for transcribing sessions, preliminary analysis of test results, creating psychoeducational materials or forming exercises, and that the information is to be given before data processing begins or before the tool enters the process (our translation). Consent, it continues, is to be voluntary, specific and clear, and should state the purpose of AI use, the data types processed, the level of human control, the main risks, the confidentiality safeguards and the client's right to refuse without losing access to help. Obliged is the strongest word in the document, and it is still the word of an informational text.

Where AI feeds a psychological opinion, as opposed to serving an internal administrative task, pp.34 to 35 of this non-binding guidance give a minimum disclosure list: the fact of AI use; the system's role in the decision, whether auxiliary, advisory or screening; the limits of accuracy and possible errors; that the final decision is the psychologist's; and the possibility of asking questions or getting an explanation in human language (our translation).

The data rule is conditional and the condition is easy to lose. On p.31, under minimisation and anonymisation, this non-binding guidance says personal data of clients, or other information that directly or indirectly identifies a person, may not be entered into AI systems in the absence of a legal basis defined by legislation for that processing, or where the special requirements established by legislation for such processing are not met. In those circumstances, the informational document says, passing names, dates of birth, contact details, place of work or study, home address, identity-document data, identification codes, photos, consultation records, psychodiagnostic results, medical information and unique life circumstances to AI is prohibited, and only de-identified information in the minimum amount for the purpose should be used (our translation). The parallel line on p.28 spells out that the legal basis includes the data subject's unambiguous consent. The guidance applies legal-basis and special-processing requirements to information entered into AI systems. It is not an unconditional bar on client data in AI tools.

Two operational lines on p.28 and p.31 are the ones a practice would act on, and both are advice in this non-binding guidance. If confidentiality cannot be secured, help using AI is not started or is stopped early, the client is told why and offered a non-AI alternative. And before using a tool for client materials the psychologist should establish whether prompts are stored, who can access them, whether they train models and whether they can be deleted; the guidance adds that if such a guarantee is absent, using the tool for client materials is risky (our translation).

What the document says about EU law, and what that is worth

The document cites the EU AI Act in three places, always as «Регламент ЄС про ШІ (EU AI Act)» and never by regulation number or article. On p.30 it says the Act treats many psychotherapeutic and medical AI solutions as potentially high-risk where they affect health, safety or fundamental rights, and requires human oversight, transparency, data quality, risk management and the ability of a person to intervene or reject the system's output. On p.32 it says psychological AI solutions often fall into the high-risk category. On pp.50 to 52 it summarises the Act's risk tiers, its penalties and its treatment of emotion recognition in workplaces and schools, and gives its own reading of the implementation timetable under a heading that references the Digital Omnibus.

Those are the document's characterisations of EU law. We report that the document makes them. We have not checked them against the Act's text for this article, and the words potentially and often are the document's own hedges, which we keep. A Ukrainian psychologist reading p.32 should not take from it a legal conclusion that any given tool is a high-risk system under the Act.

On p.31 the guidance describes GDPR compliance conditionally. It says personal-data processing is to comply with the Ukrainian law on personal data protection and, where tools or services connected with EU jurisdiction are involved, or where the client is in the EU or is a citizen of an EU country, with the GDPR (our translation). The GDPR is named again on p.56, in a table of tools, and on p.59, where users are told to check whether an app meets GDPR standards or their local equivalents. No article of the GDPR is cited anywhere in the document.

Ukrainian law is listed, not explained. The p.23 paragraph says the use and development of AI systems in psychosocial support are to be carried out taking into account the requirements of Ukrainian legislation, and names the Fundamentals of Legislation on Health Care, the law on the mental health care system, the personal data protection law and Cabinet of Ministers Resolution No. 1604 of 3 December 2025 on mental-health services delivered by information and communication technologies (our translation). That is the resolution's only appearance. The document does not say what it requires, does not say it contains an AI provision, and does not amend or implement it. We say the same.

What we did not verify

What we opened: the text layer of the ministry's PDF, retrieved from the storage URL and read by line, including the cover and its date, the table of contents, the introduction and disclaimer on pp.3 to 4, the attribution and funding sentences on p.4, the defined terms on p.7, the three lists on pp.20 to 22, the principles on pp.23 to 25, the confidentiality section on pp.28 to 29, the human-in-the-loop section on p.30, the legal section on pp.31 to 33 and the disclosure list on pp.34 to 35. We fetched the ministry's news article of 18 September 2026 and read its JSON-LD, its visible publication stamp and the link it carries to the same PDF, and we confirmed the storage object's size and Last-Modified header and the PDF's metadata with pdfinfo.

What we did not open: a rendered image of any PDF page, so every Ukrainian quotation here is verified against the extracted text layer and not against the printed glyphs. The ministry's White Paper. Cabinet of Ministers Resolution No. 1604 of 3 December 2025, the Ukrainian laws on health care, mental health care and personal data protection, the Council of Europe Framework Convention on AI, the Ukrainian Psychological Association's ethics code and the UNESCO Recommendation, all of which the document names as inputs. The text of the EU AI Act or the GDPR against the document's descriptions of them. The two live Ukraine entries on this desk beyond their slugs.

What we refuse to claim: that the document is dated 18 September 2026, because it is dated September 2026 and nothing else. That it is the first Ukrainian guidance of its kind, because the document does not say so and we did not test it. That it was approved, adopted or is in force, because no approval order, number or commencement exists in the text. That it establishes a standard of care, because we did not verify adoption by a professional body or court, or any standard-of-care effect. That psychologists must keep a written log of every AI review, because the p.24 documentation line is conditional on other rules. That client data may never be entered into AI tools, because the p.31 bar is conditional on the absence of a legal basis. That employers, schools or app developers are regulated by it, because none is addressed by name. That Resolution No. 1604 enacts any of these provisions, because the document only lists it. That any product named in the document does or does not comply with it. Where this article uses the word must or obliged, it is translating the guidance's own text and says so.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Ukraine opens drafting of an EU-aligned AI law
• Colorado HB 1195 restricts AI in psychotherapy
• Italy's Garante fines Luka over the Replika companion chatbot
• Europe AI regulation tracker
• AI Regulation News hub