Hiring in the Age of Superintelligence

Hiring in the age of superintelligence means testing less for memorized technique and more for the judgment, communication and process knowledge of someone who will work alongside AI tools. The most direct way to check this is a practical test in which the candidate uses AI on a real task from the role and explains what they accepted, what they corrected and what they rejected. You do not need to predict the future of work to do this. You only need to observe how the person makes decisions today.

Key takeaways

• AI tools already produce text, summaries and quick analyses, so the candidate's edge becomes knowing whether the output is fit for purpose, not just producing it.
• Three skills deserve weight in the interview: judgment, communication and knowledge of the business process.
• A short practical test with AI, using fictional material, shows more than theoretical questions about tools.
• Evaluate the path (questions asked, checks made, corrections), not only the final answer.
• Candidate data needs care: LGPD, transparency and human review apply to the hiring process too.

What changes when you hire for roles where AI already writes, summarizes and analyzes?

What changes is what the interview has to prove. Before, part of the filter was checking whether a person could carry a task from start to finish: write a report, build a spreadsheet, answer a customer. Now part of that execution can be done by a tool in seconds. What remains for the person is deciding what to ask for, checking what comes back and taking responsibility for the result.

It is important not to overstate the situation. What companies use today are language models and agents that make mistakes, invent details and depend on well-supplied context. That is different from superintelligence, and the distinction is worth understanding in Is ChatGPT Superintelligence? What It Is and What It Is Not. Hiring well for this moment does not depend on believing in a specific future scenario. It depends on recognizing that the person you hire will work with tools that help a great deal and make errors that sound convincing.

We also do not cover here which professions will grow or shrink. That debate exists and has its place, and there is a useful conversation about how to discuss it with your team in Will Superintelligence Take Jobs? What to Tell Your Team. The focus of this article is narrower: what to look for in an interview and how to test it.

Which skills to value: judgment, communication and process knowledge

The list below is not a scientific consensus or a forecast. It is a working hypothesis, based on how the relationship between person and tool works: whoever uses AI must decide, explain and know the terrain.

Judgment is the ability to notice that a plausible answer may be wrong or unsuitable for the case. In practice, it shows when the candidate distrusts a number with no source, notices that an answer ignores a company rule, or decides that a particular situation should not be delegated to the tool. People with good judgment also know when to stop and ask a human for help.

Communication has two sides. The first is instructing the tool: describing the goal, the context, the expected format and the limits. Vague instructions get vague results. The second is explaining the result to other people: what was done, on what basis, and what is still uncertain. A colleague or customer does not need to know that an AI took part, but they need to receive something clear and truthful.

Process knowledge means understanding the real work behind the task: who approves, which exceptions occur, where mistakes are costly, which data is sensitive. Without it, the person cannot judge whether the AI output fits. A good support professional, for example, knows that a wrong delivery promise becomes a complaint, and so reviews that part more closely than the rest.

There is one more quality that tends to go with the other three: willingness to verify. People who check a source, test an edge case or ask for a second opinion before sending usually handle tools that err without warning better.

How do you build a practical AI test for interviews?

The practical test puts the candidate in front of a task similar to the one in the role, with access to an AI tool, and observes how they work. The idea is to measure the process, not skill at typing clever prompts. This format is our own proposal and not a method validated by a study, so adjust it and compare it with your team's experience.

A simple outline for a short session:

• Choose a real, recurring task from the role that can be done with fictional material. Avoid customer data and data from other candidates.
• Prepare a small context package: a summarized internal policy, two or three sample messages and a clear objective.
• Include, without announcing it, a realistic trap: a contradictory piece of information, a missing data point or a rule the AI is likely to ignore.
• Tell the candidate they may use any tool from the company's approved list, and that they will be assessed on reasoning, not speed.
• Ask for the final deliverable and then a short verbal explanation: what they asked for, what the tool got wrong, what they changed and what they would still leave for a person to decide.
• Ask one follow-up question: if this work were sent to a customer right now, what would you review first and why?
• Score with a fixed rubric, the same for every candidate, and record notes before the evaluators talk to each other.

The trap is the most revealing part. A candidate who accepts everything the tool returns shows a different pattern from one who asks where each claim came from. Neither behavior is conclusive, but both give you material to discuss.

Illustrative scenario: a support analyst role

This is an illustrative scenario, unrelated to any real company or person. A services company is hiring a support analyst to answer customers on WhatsApp. The test task is to prepare a reply to a customer asking for a refund outside the period set in the policy.

The fictional material includes the policy (which allows refunds only within a set period and provides for an exception with supervisor approval), the customer's history and an angry message. The AI tool, without careful instruction, tends to write a polite reply that promises the refund to calm things down.

A candidate with good judgment spots the improper promise and removes it. One with good communication rewrites the text with empathy, explains what is possible and says the case will be passed to the supervisor. One with process knowledge remembers that the exception requires approval and records the reason in the system. A candidate who sends the first version without a careful read hands over a ready-made problem: a promise the company may not be able to keep.

None of these profiles is a verdict. The purpose of the interview is to see how each person reacts when the tool fails in a subtle way, and whether they can explain their own decision calmly.

How to assess the result: criteria and warning signs

A single yardstick for everyone keeps the evaluation from depending on rapport or style. We suggest scoring each criterion qualitatively (insufficient, adequate, strong) and noting one piece of observed evidence.

Criteria to observe:

• Verification: did the candidate check facts, rules and numbers before accepting the output?
• Instruction: did the request to the tool include a goal, context, format and limits?
• Correction: did the candidate identify the trap and fix it, or at least flag it?
• Clarity: are the final deliverable and the verbal explanation understandable to someone who was not there?
• Limits: does the candidate know what they would not delegate to the tool and what they would escalate to a person?
• Data care: did they avoid pasting sensitive information into an unapproved tool?

Warning signs that deserve a longer conversation, not automatic rejection:

• Total trust in the output, with no checking at all.
• Inability to explain why an answer is right.
• Dismissal of the tool, refusing to use it even when it would help.
• Polished answers that ignore the rules of the process.

The last two points include something often forgotten: rejecting AI on principle is also a way of not exercising judgment. The useful question is whether the person knows when the tool helps and when it gets in the way.

To measure performance after hiring, keep in mind that tracking AI use does not have to become excessive control. There is a practical discussion of this in How to Measure AI Productivity Without Surveillance on Your Team.

What LGPD and fairness precautions does the hiring process require?

A test with AI involves the candidate's personal data: answers, recordings, notes and sometimes image and voice. LGPD applies to this processing, and Brazil's data protection authority (ANPD) publishes official guidance that serves as a starting point. Before you begin, define the purpose, tell the candidate what will be recorded and for how long, and keep only what is necessary.

Two practical decisions help. First, use only tools approved by the company, with clear usage rules, such as those described in Company Policy for Employee ChatGPT Use: What to Allow and Ban. Second, do not use real customer material in the test: prepare fictional data.

There is also the question of who decides. If any part of the screening is automated, the candidate may have rights to an explanation and a review, a topic covered in LGPD Automated Decision-Making: The Right to Review (Art. 20). We recommend keeping a person responsible for the final decision and recording the criteria used.

For those who want a risk management structure, the NIST AI Risk Management Framework describes governance, mapping and measurement practices that can be adapted, even in a small company. This is official reference documentation. Adapting it to your hiring process is your decision.

Finally, think about accessibility and equity. A test that requires familiarity with a specific tool may favor people who already use it. Offer the same brief introduction to everyone and assess reasoning, not comfort with the interface.

How does the map, limit, pilot, transfer method apply to hiring?

The same method we use to bring AI into a business process works for designing the hiring test. Map: describe the real task of the role, where it goes wrong and who approves it. Limit: define which tools and data the test may use. Pilot: run the test in a few interviews, with an agreed success criterion, for example whether it helps distinguish profiles in a way the evaluators recognize. Transfer: document the outline and the rubric so that HR and managers can run it on their own, and decide whether to expand, adjust or drop the test.

This is a method in practice, not a promise of results. If the pilot shows that the test does not separate strong from weak candidates, the honest path is to adjust it or stop.

Frequently asked questions

Do I need to hire AI specialists for every role? Not necessarily. For most roles, what matters is knowing how to work with the tool inside the area's process, with judgment and accountability. Technical specialists make sense for roles that build, integrate or maintain AI systems.

Can I ban AI use during the interview? You can, but the result then measures something else. If the role will use AI day to day, a test without it assesses a situation that does not exist in the job. If the ban is for security reasons, offer an approved tool and fictional data.

Doesn't the practical test favor people who already master AI tools? It can, if the test rewards speed and familiarity with the interface. To reduce this effect, give everyone the same introduction and score verification, clarity and limits, which depend more on reasoning than on practice with the tool.

How long should the test last? Long enough for a realistic task and an explanation conversation, without turning the interview into free work. Prefer a small task with fictional material, and tell the candidate in advance about the format and what will be assessed.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• Anthropic: Demystifying evals for AI agents
• OECD: AI Principles