Web Research Specialist

What you'll do

You help build an evaluation benchmark for frontier AI browsing agents. The job is to invent research problems that a top model cannot solve even with full web access and several attempts. The ad is firm that this is investigative research, not subject matter expertise and not content writing.

The method runs backwards. You start from a verifiable fact, build a question that makes that fact very hard to find, and then document everything. Each task delivers three things:

• A natural-language question with a short, stable answer that anyone can check
• A set of clues, each independently verifiable, mixing fact types such as dates, people, places, organisations, works, events, records and quantities, with specific constraints
• A validation record showing the obvious searches you ran and what they returned

The ad says the evidence trail is the majority of the work, so expect more time spent documenting than puzzling.

Who fits

• A master's degree or more than 3 years of experience
• Proven open-web research skill: finding primary records, and moving through government and institutional databases, archives, registries and PDFs
• Exact sourcing: citing specific pages, tables and sections, not homepages
• Comfort researching unfamiliar subjects from scratch
• Native or near-native written English
• Patience with structured documentation
• Experience with LLM evaluation, red-teaming or benchmark construction
• Background in at least one of: reference librarianship or archival work, investigative journalism or fact-checking, OSINT, due diligence or KYC, patent or prior-art search, genealogy and records research, or competitive quizzing and puzzle-hunt construction

Familiarity with JSON and structured data delivery is a plus. The ad does not name a location requirement.

What it pays

The ad states estimated earning potential of approximately $30 per approved task, based on the current task mix. Pay is therefore tied to approval, not to hours logged, and the ad does not say how many tasks fit in an hour. Payment is in USD. The ad does not publish payment schedule terms; see Turing for how Turing describes payment generally.

Worth knowing

Good:

• A fixed 8 week project with a clear scope
• Fully remote, paid in USD, with up to 40 hours a week available
• Work that rewards careful, methodical research habits found in librarians, fact-checkers and OSINT analysts
• You start as soon as you pass the assessment

Less good:

• Per-task pay, only for approved tasks; rejected work earns nothing, and the ad gives no hourly equivalent
• Most of the effort goes into documentation, which suits some people far less than the hunting itself
• Questions must resist a strong AI, so difficulty is deliberate and a task can take a long time to get right
• It is a short project, not ongoing work
• Contractor engagement: no benefits, and you handle your own taxes

How to apply

The ad says to submit an application and complete a required assessment; candidates who pass become eligible for onboarding. It does not describe the assessment. Turing's help centre describes skill tests and, for some roles, a live interview (see Turing).

About this listing

Posted by Turing on its job board on 25 August 2026 and read on 18 October 2026. The 8 week length, the hours, the approximate per-task figure and the qualifications are Turing's own text. See Turing.