The HUMAN principle: Using AI safely in social institutions

Artificial intelligence can formulate emails, simplify texts, structure meetings, summarize information and provide ideas for everyday teaching. That sounds like a welcome relief - especially in a field of work where time for relationship work is often limited.

At the same time, legitimate questions arise: Can I have a case report revised by a chatbot? Can I automatically transcribe a conversation? What happens to the data entered? And how do I know whether an answer is technically correct or just convincingly formulated?

This uncertainty is not a sign of a lack of openness. It is an expression of professional responsibility.

AI can support specialists. The responsibility for assessments, relationships and decisions always remains with people.

AI is already here – even without an institutional strategy

AI tools have long been tried out in many teams: via private accounts, free translation tools or chatbots on personal smartphones. A complete ban rarely prevents this use. It tends to shift them into an unregulated area.

Social institutions therefore need understandable guidelines, approved tools, training and space for questions. Experiences from theSwiss social workshow: AI can support text work, information preparation and structuring, but does not replace professional contextual understanding, relationship work and professional assessment.

Legally, we are not in a free space either. As of July 2026, there is still no comprehensive legislation specifically for AI in Switzerland. LoudFederal ChancelleryA consultation template should be available by the end of 2026. However, the existing data protection law already applies directly to AI-supported data processing, such asFederal Data Protection and Information Commissionerholds on.

What AI can do well – and what it can’t

Generative AI is particularly good at manipulating language and existing information. Examples of useful tasks include:

• create a first draft of text from keywords,

• make complicated texts more understandable,

• suggest possible structures for a report,

• summarize published documents,

• Develop variants for an activity or workshop,

• Prepare questions for a team meeting or case reflection.

However, AI does not know whether a statement is true. It calculates linguistically appropriate answers. This allows them to invent information, misrepresent connections or conceal uncertainties. This is particularly relevant in the social sector: a convincingly worded error in a case report can be more serious than an obvious typo.

The HUMAN principle for social institutions

Responsible use of AI can be summarized with six simple principles.

M – People decide

AI is allowed to provide suggestions but not take on professional responsibility. Decisions about services, placements, support measures, threats, sanctions, diagnoses or intensity of care must not be delegated uncritically to a technical system. An AI suggestion can also influence the perception of a specialist, even if a human still formally decides.

Practice rule:The greater the possible impact on a person's life, the smaller the role of AI can be.

E – Limit inputs

Many dossiers contain particularly sensitive personal data: such as information on health and privacy, religious or political views, administrative or criminal proceedings and social assistance measures. Such information does not belong in freely accessible or non-released AI tools.

Even apparent anonymization is often not enough. A person can still be recognizable from the combination of age, place of residence, family constellation, institution, diagnosis and a special event. Where real cases are not required, synthetic or significantly distorted examples should be used.

N – Check instead of adopting

Every AI result is a draft. Before use, at least the following questions must be answered:

• Are names, numbers, dates and quotes correct?

• Did the AI fill in information that wasn’t entered?

• Are the statements comprehensible and technically justified?

• Is the wording respectful and non-stigmatizing?

• Does the text meet the institution's documentation standards?

• Could the wording be explained and represented upon inspection of the files?

Particular caution is necessary when it comes to laws, medical information, crisis situations and cantonal responsibilities. AI can provide initial orientation, but it does not replace a reliable primary source.

S – Ensure protection organizationally

Data protection may not be delegated to individual employees. The institution must determine which applications are approved for which purposes, which categories of data can be entered and how long content is stored. Equally important are processing locations, subcontractors, roles, multi-factor authentication, logging, deletion and the process in the event of a data protection incident.

With cloud services, the institution remains responsible for lawful data processing. TheEDÖBrequires that order processing be contractually secured and possible data disclosures abroad be examined. Swiss hosting can be an important criterion, but it is not enough on its own: ownership structures, support access, subcontractors, applicable legal systems and technical configurations remain relevant.

If there is a high risk to personality or fundamental rights, a data protection impact assessment may be necessary. Public institutions must also examine cantonal data protection law and their legal basis. ThisInformation sheet from the Canton of Zurichprovides helpful guidance for this.

This article does not replace legal advice. Institutions with a public mandate, health data, child and adult protection tasks or special confidentiality obligations should check each specific use case with their responsible data protection and specialist office before introduction. What is crucial is not only whether a tool can be used in principle, but also whether the purpose, scope of data, responsibilities and protective measures for precisely this use are regulated in a sustainable manner.

C – Check equal opportunities

AI systems adopt patterns from their training data - and with them possibly social prejudices. A system can evaluate behavior differently depending on gender, disadvantage people with non-standard language, stereotype cultural backgrounds or describe disability primarily as a deficit.

That's why you should consciously look for distortions in relevant content. This test can also be supported by AI, but does not replace human reflection.

H – Inform, participate and document

People should know when AI significantly influences their communication or processing that is relevant to them. Transparency is particularly important when conversations are recorded or transcribed, a chatbot communicates with clients, AI creates recommendations or prioritizations, or a result can influence services and measures.

For children, young people, people with cognitive impairments or people in dependent relationships, formal consent alone is often not enough. Information must be understandable and a real alternative should remain possible.

The AI traffic light for everyday work

“Red” does not necessarily mean that an application is banned in the future. It means: not introduced spontaneously and not by individual employees alone.

Which AI tools are suitable?

No tool is “DSG compliant” just because of its product name. The decisive factors are the specific use case, the contract version, the configuration and the data entered. Instead of just comparing models, institutions should examine appropriate tool categories:

• AI in the existing office environment:obvious for emails, meetings and general documents, as long as permissions and storage are properly maintained.

• Institutional AI Assistant:suitable for general text work, concepts, fictional examples and clearly defined internal assistants.

• Professional translation tool:practical for multilingual writing; In the case of personal data, a legal basis and regulated order processing are still required.

• Swiss or Swiss hosted solution:can improve data sovereignty, but does not replace contract review or role, deletion and security concepts.

• AI directly in the client information system:For case-related applications, it often makes more sense than a separate public chatbot if authorizations are adopted, data access is limited and sources used are clearly identified.

The most important tool question is not: Which model writes best? But rather: What data is allowed in, who can access it, where is it processed and who controls the result?

You can get started in eight weeks

1. Form a small working group.Practice, management, IT, data protection or quality assurance and – depending on the application – the client’s perspective belong at the same table.

2. Collect actual usage.Ask without fear which tools are already used for what and where there are uncertainties.

3. Select three green use cases.For example, simplifying general correspondence, preparing workshops or summarizing published specialist information.

4. Release one or two tools.A few clearly regulated institutional accounts are better than many private solutions.

5. Adopt a one-sided guideline.The rules must be short and specific enough to be used in everyday life.

6. Train employees practically.Data protection, typical model errors, source checking, discrimination-sensitive language and the reporting channel in the event of errors are more important than prompt tricks.

7. Measure the pilot.Time savings including control, correction effort, technical quality, comprehensibility and data protection problems show whether the use really helps.

8. Check rules regularly.Products and contract conditions change quickly. Approvals should be reviewed at least every six months.

Template for a short institutional AI policy

An everyday guideline can fit on one page. The following eight rules form a possible basis and must be adapted to the mission, legal form, canton and data sets of your own institution.

1. AI is a tool.Professional responsibility, assessments and decisions remain with the employees.

2. Only approved tools and institutional accounts are used.Private accounts are excluded for professional use.

3. Public or non-released tools do not include personal data, case information, confidential documents or access data.

4. Data minimization also applies to released tools.Only what is necessary for the specific and documented purpose is used.

5. AI results are checked technically, factually and linguistically before each use.Special attention is paid to fictional or unsubstantiated content.

6. AI does not decide on performance, rankings, diagnoses, threats, sanctions or other significant measures.

7. Conversations are only recorded or transcribed after institutional approval and transparent information.The rights and choices of the data subjects are preserved.

8. Errors and incidents are reported.There is a known internal contact point for unintentional data transmissions, suspicious results and uncertainties.

Three safe prompt examples

Making a general text more understandable

Make the following general information text more understandable for adults without specialist knowledge. Use short sentences, explain technical terms and do not change facts. Mark unclear areas. Text: [Text without personal or confidential data]

Prepare a team meeting

Create a process for a 45-minute team meeting on the topic of “Dealing with challenging handover situations”. The aim is not to assess specific cases, but rather to develop common technical principles. It should include an introductory question, three reflection questions, a small group exercise, securing results and next steps.

Check a text for problematic language

Check the following anonymized text for unsubstantiated assumptions, deficit-oriented or stigmatizing formulations, the mixing of observation and interpretation, and discriminatory statements. Don't create new facts. Show problematic areas first and then suggest more neutral formulations.

Conclusion: Not as much AI as possible, but as good social work as possible

The central question is not how social institutions can use as much AI as quickly as possible. The better question is: Where can AI make administrative or linguistic work easier so that there is more time for relationships, reflection, participation and individual support?

A good use of AI is not invisible, not uncontrolled and not completely automated. It is limited, comprehensible and technically verifiable.

Understood in this way, AI is not in contrast to professional social work. It can become a useful tool – as long as people remain at the center not only in the name of the principle but also in actual practice.

Which applications would actually relieve your everyday working life - and for which tasks must the responsibility remain entirely with the person?

Nikos and Ramon