A job advertisement is often the first concrete contact between a company and potential applicants. AI can quickly structure a raw text, draft variants or mark hard-to-understand sentences. But it does not know the real workplace or the lived culture and can easily adopt exaggerated requirements, outdated role models or vague promises from its source. Therefore a generated text must never go online unchecked.
Anyone who Review AI job ads möchte, braucht mehr als eine Rechtschreibkontrolle. Tasks must be correct, mandatory criteria factually justified, language accessible and the remuneration information complete. This practical guide shows a transparent editorial workflow for Austrian companies. It concerns the wording of an ad, not the automated assessment or ranking of applicants.
The job advertisement begins with a clear briefing
A model can only work with what it receives. The most common quality error therefore occurs before the prompt: an old ad is copied even though the team, working hours, tools or responsibilities have changed. From an imprecise briefing AI produces a fluent but still inaccurate text. The linguistic polish can even hide the professional shortcoming.
Before generation the specialist department together with recruiting answers five questions: Which three results should the person achieve in the first year? Which tasks actually occur every week? Which qualifications are indispensable on the first day? What can be learned on the job? Which framework conditions have already been decided as binding? Only these facts form the approved source.
The WKO shows, using the example of an apprenticeship advertisement, how AI can rewrite an existing text to be more modern and clearer. Crucially, the employer must adapt the result to the actual position. A fabricated training program, flexible scheduling, or team size is not a creative detail but a false promise.
Visibly separate facts, requirements, and wishes
A fair advertisement distinguishes between what the job objectively requires and what would merely be convenient. A simple three-part separation helps:
- Must: without this prerequisite, a legal, safety-critical, or central technical task cannot be performed;
- Trainable: the skill can be developed during onboarding or further training;
- desirable: experience makes it easier to get started, but it is not the sole deciding factor for an invitation.
This labeling reduces cluttered lists. Ten seemingly mandatory criteria deter people, even though perhaps only three are truly necessary. Do not ask the AI for a "more demanding profile" in general terms. Instead, provide it with the three groups and require that it does not add any new requirements. The reviewer then compares each point with the approved role description.
Every requirement needs a connection to the activity
Phrasings such as "young team," "native speaker level," "physically fit," or an arbitrary number of years of professional experience can unnecessarily exclude people. The better question is: Which observable skill is needed for which specific task? For example, "perfect German" can become "you explain technical work steps in an understandable way in German and document short handovers," provided that is exactly what is required.
Educational qualifications should also be reviewed. Does the role actually require a specific title, or does demonstrable experience count just as much? Is a driver's license indispensable because frequently changing locations without public transport connections must be reached, or was it included out of habit? AI should formulate alternatives here, but the factual justification must come from the company.
The The Equal Treatment Commission explains based on its consulting practice, that requirements for professional experience or excellent German language skills, for example, can have a discriminatory effect depending on the activity. A general list of words therefore does not replace an examination of the specific workplace.
The Fairness Check reviews the text from multiple perspectives
A single round of corrections can miss systematic patterns. Use a consistent perspectives check and document brief responses:
- Would a qualified person of a different age recognize that they could be meant?
- Is the language gender-neutral, and are the image, job title, and body text consistent?
- Are origin, religion, worldview, or sexual orientation not referenced either directly or through unnecessary codes?
- Are requirements for people with disabilities truly job-related, or does the text merely describe a customary way of working?
- Are caregiving responsibilities excluded by unjustified availability requirements?
The AI can flag potential barriers for each perspective. However, it must not claim legal clearance. Notable passages are sent to recruiting, the relevant department, and, if needed, to a qualified legal or anti-discrimination office.
Maintain gender neutrality throughout the entire text
A neutral addition after a masculine job title does not automatically solve all language problems. Review the headline, direct address, examples, pronouns, and image description together. Role terms like "team leader", "specialist", or "customer service staff" can work if they are understandable to the target audience. Alternatively, the job title can be written in an inclusive form that is used consistently across the company.
The The Equal Treatment Authority describes gender-neutral and non-discriminatory language useand points out that positions advertised publicly or internally must not address anyone solely on the basis of gender. For the editorial team, a binding internal rule is therefore advisable: apply the chosen system consistently without sacrificing the readability of the duties.
Inclusive language stays concrete rather than promotional
Many AI-generated texts sound open without providing practical information. Sentences like "Everyone is welcome here" are positive but do not replace clear conditions. State what applicants need to know: Is the workplace step-free accessible? What on-site presence is required? Is there shift work, scheduled core hours, or travel? To what extent is remote work actually possible?
Avoid personality templates such as "Digital Native", "rockstar", "resilient all-rounder" or "always cheerful". Describe behavior in the work context: "You prioritize multiple customer requests and raise bottlenecks early." This allows people to compare their experience without having to identify with a social ideal type.
Pay and working hours are quality data
In Austria, the pay information does not belong in a subsequent standard module. It must match the position, the extent of employment, and the applicable basis. The responsible person enters the specific amount and any willingness to overpay; a language model must not guess either the collective agreement or the classification.
The Recommendation of the Equal Treatment Authority on pay disclosure summarizes the requirements for non-discriminatory advertisements and an amount-based minimum pay information. For each text there should therefore be a four-eyes field: amount, reference period, extent of employment, and overpayment information.
Working time terms also need precision. "Flexible" can mean flexitime, rotating shifts, or short-notice availability. Instead write what is fixed: weekly hours, shift times, lead time for schedules, on-call duties, weekend work, and home office rules. That does not make the ad longer, but more decisive.
Describe tasks that applicants will recognize
A good ad shows a realistic workday. Three to six prioritized tasks are more helpful than twenty nouns. Start with active verbs: advise, check, coordinate, maintain, document, or develop. Add the object and the outcome. "You coordinate delivery dates with two production areas and record deviations in the planning system" is more verifiable than "interface management".
Ask the AI to translate abstract terms into examples. Then the specialist department reviews each example. Is a rare crisis situation presented as a daily task? Does the text promise decision-making freedom even though approvals are required? Does it mix junior and senior responsibilities? Correct these differences before the language polishing.
A fixed prompt prevents creative additions
A controlled editorial prompt contains the role, target audience, approved facts, style, and prohibitions. A useful basic pattern is essentially: "Formulate a clear job advertisement from the following confirmed information. Do not invent tasks, benefits, working time models, salary values, or qualifications. Separate mandatory, learnable, and desirable requirements. Mark missing information as an open question."
Add the desired structure and a list of problematic in-house terms. Have the model output a fact table at the end: statement, source in the briefing, uncertainty. This table is working material and is not published. It forces traceability and makes invented details easier to spot.
For a broader but still controlled start, a limited AI pilot in the team helps. A job advertisement is only suitable for this if the approval roles and the permitted input data are fixed. Personal data of real applicants does not belong in this text process.
The cross-check reveals hidden biases
After the first draft, a controlled cross-check follows. In this process, individual neutral characteristics are changed in a thought experiment: Would the same requirement still seem plausible if the person addressed were older, had a non-Austrian-sounding name, mentioned caregiving responsibilities, or had a visible disability? The advertisement itself is not tested with real profiles for this. It is about recognizing generalized expectations in the text.
Let the language model provide hypotheses at most: Which phrasing could be off-putting, which criterion is unclear, which information is missing? Every observation is checked by humans. A model can reproduce stereotypical patterns itself and therefore must not become a fairness arbiter.
A useful method is also the deletion test. Remove each adjective and each requirement one at a time. If the factual description of the job does not change, the phrasing was probably just image language. Words like "young", "dynamic", "down-to-earth" or "culturally fit" may seem harmless but can convey vague social expectations. Replace them with concrete descriptions of collaboration: How are decisions made, how is onboarding conducted, how often does the team coordinate?
The results of the counter-check are recorded in a short editorial log. That way the next reviewer sees not only the final sentence but also the discarded phrasing and the reason for the change. This creates learning material for future postings without collecting applicants' data.
The red-line check before publication
Before approval, wording is no longer reworded; instead it is checked. The final check includes at least:
- Title, location, working hours, contract type and start date match the approved position;
- Every mandatory requirement has a documented factual justification;
- Gender-neutral and non-discriminatory phrasing is used consistently;
- Compensation amount, reference basis and any overpayment have been professionally reviewed;
- Benefits and development paths are up to date and actually available in the organization;
- Contact details, application procedure, deadline and data protection notices lead to the correct process;
- No internal note, prompt instruction or invented statement remains in the text.
A second person reads it without knowledge of the prompt. This reveals whether the advertisement is understandable outside the project context. For recurring roles, the organization can use an approved template, but every new posting requires a fresh fact check.
The quality spot check after publication
Even a carefully reviewed text can have unintended effects. Therefore observe not only clicks but also questions arising from the recruiting process: Do interested parties understand the tasks? Are the same criteria repeatedly misinterpreted? Are there signs that qualified groups do not feel addressed? Document feedback without building hasty profiles of individual people.
Changes to the job posting are versioned. Record which passage was adjusted for what reason and who approved it. If AI is used again to assist, it receives only the current approved factual basis. That prevents an old, already corrected phrasing from reappearing through the chat history.
Text support is not applicant selection
The boundary must remain clearly defined organizationally. Organizing a job posting text is a different use than filtering resumes, assessing people, or prioritizing invitations. The EU AI Act behandelt bestimmte KI-Systeme für Recruiting, Auswahl und Beschäftigtenmanagement als Hochrisiko-Anwendungen. Daraus folgen wesentlich weitergehende Anforderungen als bei einer reinen Schreibassistenz.
If a provider expands its product to include matching, ranking, or automated recommendations, the use case will be reassessed. The previous approval for text drafts does not automatically remain valid. Product description, settings, data flows, human oversight and affected rights must then undergo a separate substantive, legal and technical review.
An eight-step process for fair AI job postings
- The specialist department and Recruiting confirm tasks, objectives, and framework conditions.
- Requirements are divided into Must, Learnable, and Desirable.
- A constrained prompt prohibits new facts and requires open-ended questions.
- Recruiting checks language, comprehensibility and inclusive phrasing.
- The specialist department matches each activity to the actual workplace.
- Compensation, working hours and legally sensitive wording are approved separately.
- A second person performs the redline check on the finished text.
- Versioning, approval and later corrections are documented in a traceable way.
This keeps AI an editorial tool. It speeds up variants and highlights unclear passages, but it neither performs fact-checking nor assumes responsibility.
A fair job posting makes the decision easier for both sides
The best job posting doesn't sound maximally impressive; instead it enables an honest fit assessment. Applicants recognize the task, expectations, conditions and scope for development. The company receives applications from people who know what they are getting into.
Anyone who Review AI job ads as a fixed process, it combines speed with care: confirmed facts in, clear boundaries for the model, multiple professional perspectives and a designated human sign-off. This discipline is at the same time a good foundation for a company-wide AI policy with clearly defined roles. The published text therefore remains a reliable statement from the company – and not a plausible-sounding draft produced by the machine.
