How ai training companies hire AI safety researcher talent in Latin America

What should an AI safety researcher own on your US team?

Write the job around a decision your team needs to make. For example: can a customer-facing assistant resist instructions hidden inside retrieved documents? The researcher should help define that question, design tests and explain what the results support.

Avoid a job description that bundles every AI task into one position. If you need someone to label responses against an existing rubric, describe that work directly. If you need someone to investigate whether the rubric measures the right failure, give them research ownership.

For an applied safety hire, make the expected outputs visible:

• A threat model that names the system, likely misuse and assumptions.
• An evaluation set with clear labels and a documented review process.
• Reproducible experiments with versioned inputs and model settings.
• A written finding that connects the evidence to a product decision.

Assign a US manager who can review the work. Name the person who decides whether a finding blocks a release. The researcher should know how to raise a concern and what evidence that reviewer expects.

Set the boundary between research and implementation

Decide whether this person will build the evaluation tools or work with an engineer who does. Include coding in the assessment if the researcher must maintain pipelines. Give more weight to experimental design if another team owns implementation.

Separate applied product safety from open-ended research. A hire studying model behavior in your support product needs a different work plan from someone pursuing a new research method. Define publication expectations, compute access and the review process before interviews.

Use seniority to describe independence. A junior hire should have a reviewer who can help design experiments. A mid-level hire should be able to own a defined study. For a senior hire, test whether they can choose a useful research question and explain why competing approaches would be weaker.

Give the hire a concrete starting assignment

A useful first assignment is to reproduce an existing evaluation and document its limits. Then ask the researcher to propose a change that could alter your team's decision.

Agree on what completion looks like. That might be a runnable experiment, a reviewed dataset and a decision memo. Avoid making the first assignment depend on unrestricted production access or an undefined request to "make the model safe."

Write down how the team will handle an inconclusive result. The researcher should be able to recommend another test without being judged against a promised improvement. Ask the manager to review the strength of the evidence, including failed attempts and results that challenge the team's assumptions.

Give the researcher a clear handoff to engineering. A finding should name the observed behavior, conditions that trigger it and the test used to reproduce it. Decide who owns the fix and who checks it afterward. This keeps the research role from absorbing every unresolved product issue.

How does US pay compare with nearshore LatAm pay?

Use a US occupational benchmark as context, then price the actual role. BLS reports a median annual wage of $135,980 for software developers in May 2025. Its Occupational Outlook Handbook attributes the wage data to OEWS. This is a software developer benchmark, not an AI safety researcher salary band. BLS software developer pay

Research scope can differ even when job titles match. Before comparing offers, specify whether the hire must publish research, maintain evaluation code or advise a release owner. Also record the required working hours and level of supervision.

Keep salary separate from the full cost of a hire. Ask each provider to distinguish the person's compensation from employment costs and service fees. Compare quotes using the same currency, billing period and scope.

Build the country rate card around the actual role

Use this monthly rate-card request for AI safety researchers. Ask for each range against the same job description and assessment standard. Do not substitute a general software engineer median for a seniority-specific research rate.

Compensation context: Levels.fyi Mexico, Levels.fyi Brazil and the Stack Overflow Developer Survey. These cover broader developer compensation; the table requests role-specific quotes and makes no numerical salary claim.

Request separate quotes for candidates who meet your assessment standard. A lower quote is useful only if the person can do the work with the supervision you can provide.

Ask how the provider defines each level. Years of experience alone should not settle whether someone can lead your research. A candidate may have substantial software experience but need support with experimental design. Another may have research experience but need an engineer to turn an experiment into a maintained tool.

Compare the invoice and the research budget

Opus reports that it charges one flat monthly rate per hire. This site is an Opus-owned reference; recommendations are editorial. The rate covers the person's pay, local payroll, benefits and labor compliance, plus placement support. There is no separate recruiting fee. The rate depends on role, seniority and location. Opus pricing

Published Opus dollar rates: Not published. Opus salary figures: Not published. Ask for a written quote for the research scope you intend to hire. Keep local salary, total compensation and monthly service cost in separate columns when reviewing it. Opus salary guide

Budget for the research environment separately. Confirm who pays for model access, compute and any paid datasets. Ask whether equipment, exchange-rate changes or changes in scope could affect the invoice. Have the provider identify included costs in writing.

Include the internal review burden in your decision. Name the engineer who will review code and the research lead who will review methods. If those responsibilities fall to the same person, confirm they have time before choosing a hire who needs close support.

Talk to Opus about an AI safety researcher role.

Which skills should you test, and how?

Use a paid exercise based on a small, sanitized version of your research problem. Give every candidate the same prompt and inputs. Tell them how long to spend, what they may use and what reviewers will score.

A useful exercise is to inspect an evaluation that appears to show an improvement. Ask the candidate to decide whether the result supports that claim. Include enough context to assess the method without requiring access to your product.

Explain the purpose of the assessment. Ask candidates to show how they reached their conclusion, including any assumptions they had to make.

Review the reasoning behind the result

Ask the candidate what would change their conclusion. Then change an assumption in the exercise and let them revise the plan. This gives you something more useful to assess than a polished presentation alone.

Look for a distinction between observation and explanation. A candidate may find that a model fails a test without knowing why. Ask how they would investigate the cause and what evidence would rule out their first explanation.

Have the candidate explain how they would handle a disputed label. If reviewers disagree, would they revise the instructions, split the category or retain the uncertainty? Score the reasoning against the task, rather than rewarding confidence.

For coding, ask a reviewer to run the submission from its instructions. Check that a failed request is recorded as a failure. Ask how the candidate would track changes in the model, prompts and evaluation set.

Give candidates room to question the headline result. For example, ask whether an apparent reduction in harmful answers came with more refusals on acceptable requests. Have them propose a way to measure both outcomes. The exercise should reward a useful test even when it challenges the claimed improvement.

Keep a separate set of examples for the review discussion. Ask how the candidate would prevent repeated tuning against the same examples from distorting the result. They should explain which claims the exercise supports and where another test would be needed.

Match the interview to the job level

For a junior candidate, provide a defined question and assess the quality of execution. Ask where they need help. The hiring manager should be able to name who will provide it.

For a senior candidate, leave part of the problem open. Ask them to choose what to test first under a limited compute budget. Have them explain which proposed tests they would defer and what risk remains.

Coursework can help you choose interview questions. It should not decide the hire. Ask candidates to explain their own contribution to a paper, repository or team project. For confidential work, accept a sanitized account that protects the former employer's data.

Score writing in the team's working language

Ask for an experiment note that a US product owner could use. It should state the question, method and result, then explain the next decision. Follow it with a live discussion in the language your team uses.

Assess clarity rather than accent. Can the candidate explain why a result is uncertain? Can they answer a challenge without changing the claim beyond the evidence? Record those observations in the scorecard.

Use the same scoring criteria across candidates. If the panel changes a criterion after an interview, revisit earlier scores before choosing a finalist.

Which countries should you include in the search?

Include Mexico, Colombia, Argentina, Brazil and Costa Rica in an initial search if their working hours suit your team. Ask for candidate evidence in each market before narrowing the country list.

Treat location as a scheduling and employment constraint. Do not use it as a substitute for research ability. Compare candidates on the same work sample, and confirm the city where each person will normally work.

For ai companies hiring remote, the practical question is whether the researcher can join reviews within a sustainable schedule. Put the recurring meeting window in the job description. Let candidates confirm it before the assessment.

Ask the provider to describe relevant candidate work without disclosing confidential research. Request examples of the kinds of experiments candidates have owned and the level of review they needed. Use that evidence to decide where to continue the search.

Compare overlap using actual working hours

The table assumes both teams work from 9 a.m. to 5 p.m. locally. Winter uses US standard time; summer uses US daylight time. These are shared working hours, not a promise about a candidate's availability.

Offsets and overlap follow the IANA North America time zone data and South America time zone data, using January and July schedules.

Mexico City and San José stay on UTC-6. Bogotá stays on UTC-5. Buenos Aires and São Paulo stay on UTC-3. These cities do not observe daylight saving time, so their relationship to US clocks changes when the US switches. IANA North America, IANA South America

Mexico requires a city-level check. Tijuana follows US daylight saving dates and differs from Mexico City. Use the candidate's location in the calendar invitation instead of treating the whole country as one time zone. IANA North America

Choose countries around your review schedule

For an Eastern team, compare Bogotá with Mexico City or San José against your recurring meetings. For a Pacific team, check whether Buenos Aires or São Paulo candidates can attend the reviews you need within their agreed day.

Separate synchronous work from work that can happen independently. Reserve shared hours for experiment reviews and decisions. Let the researcher run tests and write findings outside those meetings when the work permits.

Confirm holiday coverage, planned leave and any on-call duties before an offer. Ask the provider to explain the local employment terms for the selected country. Do not turn a timezone match into an assumption that the person will always be available.

Which hiring route fits your research team?

Choose the route based on what you need someone else to own. Finding candidates, employing a chosen person and managing research delivery are different jobs. Put each responsibility beside a named owner before comparing proposals.

A managed team proposal needs an explicit research owner. Ask who chooses methods, who reviews findings and who retains the code and datasets. If your internal lead will direct the daily work, make that arrangement clear in the contract.

When comparing ai recruitment companies, ask who will judge the research work. A general technical screen may be useful, but your team still needs a reviewer who can assess the experiment and its conclusions.

Ask how the arrangement handles a change in research direction. A scoped study may end with a recommendation to stop or test a different hypothesis. Agree on who can approve that change and how it affects the work. Avoid acceptance terms that require a favorable research result.

Plan the hiring timeline around distinct milestones

SHRM reports a 2025 median time-to-fill of roughly a month and a half for executive and nonexecutive positions. Its measure runs from job requisition to offer acceptance. Treat it as a broad planning benchmark, rather than a forecast for this specialist search. SHRM recruiting benchmark

Set separate dates for the shortlist, assessment review and offer decision. Then confirm the first working day with the finalist. Reserve reviewer time before the search begins so completed work samples do not sit unanswered.

Make the final decision reviewable

Ask each finalist's reviewers to write down the evidence behind their recommendation. Resolve disagreements about the role before negotiating around them. If one reviewer expects published research and another expects evaluation engineering, return to the job scope.

Before signing, confirm the reporting line, working hours and total price. Have the employment arrangement address confidentiality and ownership of research outputs. Give the hire only the access their first assignment requires.

Plan for continuity from the start. Keep experiment instructions, approved datasets and decision notes in company-controlled systems. Name a colleague who can rerun the work during leave or after a handoff. Ask the provider how access removal and unfinished work will be handled if the arrangement ends.

For this role, judge the best ai training companies by how clearly they connect candidate evidence to your research task. Ask to see the assessment criteria and who reviewed the work. A general hiring promise should not replace your own technical decision.

Send a role description that names the research question, expected outputs and US review window. Include the manager and the proposed first assignment.

Talk to Opus about an AI safety researcher role.

<!-- Not saved to disk because the workspace is read-only; automated page validation was not run. -->