VBT Task Author

What you'll do

You build research questions that current frontier models cannot answer, and prove it. The work starts with finding a fact that is real but hard to surface: something that does not fall out of a simple search. You then write a precise task around it, run it against a leading model, and read the model's search and reasoning trail to see exactly where it went wrong.

The ad is careful about one distinction: a task that fails because the model has a genuine weakness is useful, while a task that fails because the prompt is ambiguous is not. Telling those apart is the core skill. Each task needs a primary-sourced answer key with the source URLs documented. Strong performers also review tasks written by peers and carry working "model-breaking" patterns into new subject areas.

Who fits

• At least 1 year in investigative research, fact-checking, benchmark task authoring, red-teaming or technical QA review
• Hands-on familiarity with AI benchmarks such as GAIA, BrowseComp, SimpleQA, SWE-bench, Terminal-Bench or Humanity's Last Exam
• Intermediate to advanced web search and open source intelligence (OSINT) skills, including academic archives, local records and obscure databases
• The ability to read model reasoning chains and search logs and name the root cause of a failure
• Precise writing: unambiguous prompts with answer keys that hold up

Listed as preferred: experience designing prompt-injection or reasoning-trap tasks, published investigative or academic research, niche record-heavy domains (local government filings, regional sports history, legal archives, historical patents), and past work as a lead reviewer or quality owner on evaluation projects. The ad also states that a bachelor's or higher degree from a relevant engineering domain is a must.

What it pays

The ad does not publish a rate. Turing sets rates per project, so you will usually learn the figure only once you are matched; see Turing for what its listings tend to disclose.

Worth knowing

Good:

• Defined scope: up to 40 hours a week for up to 4 weeks, as a contractor
• Work that rewards research craft rather than coding, which suits investigators and fact-checkers
• Failure diagnosis, not just question writing, so the role teaches how models search and reason

Less good:

• No pay figure, so you cannot compare it with other offers before applying
• Up to 40 hours a week is a heavy load for a short engagement, and the ad asks for 4 hours of overlap with Pacific time
• Contractor status, with no mention of benefits
• A degree requirement in an engineering domain narrows who can apply, even for a research-style task

How hiring works

The ad says candidates submit a Job Interest Form and an assessment, and that those who finish the assessment go through review before proceeding. It gives no further detail; Turing's help centre describes skill tests and, for some roles, a live interview (see Turing).

About this listing

Posted by Turing on its jobs board and read on 18 October 2026. The hours, overlap, engagement type and length, requirements and evaluation steps are Turing's own text. See Turing.