Fide AI

FID-107 · Open question

What Makes an AI Safety Research Task a Useful First Contribution?

Which task designs help first-time research contributors produce a result another researcher can use within an agreed time and review budget?

Why the question remains open

An open problem can be important yet require unadvertised access, expertise, or reviewer time. Fide already has a way to propose and coordinate tasks. A study can test whether a better starting packet closes a real barrier for contributors and whether maintaining that packet costs less than the rework it prevents.

Working hypothesis

A proposition to test, not a finding.

A scoped task with pinned evidence, a worked example using a different case, and a clear path for recording blockers will yield more independently usable first outputs than a well-written task brief alone. Preparation costs or limited reviewer availability may offset the improvement.

Proposed method

How the question could be tested

  • 01Prepare two small, comparable tasks using the existing research-task template. Start with a two-source synthesis under FID-045 or an offline aggregate check under FID-106. Give both versions a defined output, eligibility criteria, time cap, public starting materials, and equal access to review. Add pinned inputs, a small worked example, and a blocker checklist to one version. Dry-run the tasks for feasibility before any participant study.
  • 02If a separately reviewed opt-in pilot is feasible, compare the versions across matched tasks with comparable prerequisite skills and support. Have an independent reviewer score each artifact against predeclared completion and evidence criteria, blinded to task version when practical.
  • 03Count usable outputs per attempted task, supported negative or inconclusive results, blockers, rework, contributor time, and total preparation and reviewer time. Record what the coordinator would reuse and report the results at cohort level.

Needed controls

What must constrain the study

  • 01Keep access to support and ordinary participation separate from research consent. Use no private contributor records, unsolicited invitations, or public individual rankings. Agree credit, access, and time expectations.
  • 02Treat qualitative feedback and a small voluntary sample as design evidence. Report selection, task type and skill differences. Do not generalize a small pilot to field-wide retention, productivity or neglectedness.
  • 03Keep a supported correction or failed reproduction eligible as a useful output. A passing completion checklist does not validate a scientific claim.

Relationship to existing work

FID-045 supplies source-review tasks and FID-106 supplies reproducibility tasks. This call compares what makes those tasks usable for contributors and reviewers. It works through the existing contribution process.

Expected outputs

Artifacts the work should produce

  • 01Two starter tasks and a short artifact-quality rubric.
  • 02A record of access and review bottlenecks and the cost of addressing them.
  • 03Evidence on which task-packet elements to keep, revise, or retire.

Open questions

Uncertainties the protocol must resolve

  • 01Which barriers come from task wording, access, background knowledge, or reviewer capacity?
  • 02Which parts of a worked example help contributors without giving away the answer?
  • 03Does the packet save enough review time to justify keeping its inputs current?

Open question

Open work

Primary need: research operations, study design, mentoring, and independent artifact review

  • Dry-run one public first task for 30 minutes using the research-task template. Return what you produced, the first blocker if any, and one change an independent reviewer can check.
  • Review a submitted source note or reproduction record against the stated completion criteria.

New to the Contributor Network? Share a short private introduction once, even if you join several calls. You can join the public discussion while introducing yourself.

Join the Contributor Network ↗