Prolific connects researchers and AI teams with a vetted, demographically filterable pool of paid participants for surveys, behavioural studies, preference data and model evaluation, with quality checks and pay-as-you-go platform fees.
Used this tool? Rate it
Used this tool? Rate it
Prolific is a two-sided marketplace for human data. On one side sit researchers, product teams and AI labs who need answers from real people; on the other sit hundreds of thousands of identity-verified participants who are paid to complete surveys, behavioural tasks, interviews and model evaluations. The company frames its promise plainly on its homepage: real human data, ready in minutes, collected easily from real people rather than synthesised or scraped. It was founded in 2014 with roots at Oxford University and is headquartered in London.
Understanding Prolific as a marketplace rather than a survey tool is the single most important thing for a newcomer. Prolific does not give you a form builder. It gives you a pool, a set of filters for slicing that pool, a payment rail, and a quality apparatus that tries to guarantee the person on the other end is a real, attentive, non-duplicated human. You typically host your questionnaire or task elsewhere — in Qualtrics, Gorilla, a custom web app, or your own annotation interface — and use Prolific to bring qualified people to it and pay them fairly for their time.
The platform's centre of gravity has shifted noticeably. It began as infrastructure for academic psychology and social science, where representative online samples were hard to obtain and Mechanical Turk had a reputation for shallow, gamified participation. It now serves a second and increasingly dominant constituency: AI teams sourcing preference data, safety evaluations, red-team probes and expert annotations. In a Toolso context the useful mental model is that Prolific sits on the supply side of AI — it is where the human judgement that trains and grades models comes from.
A vetted, filterable participant pool: The platform advertises 300k+ active and engaged participants (a vendor-reported figure). What makes the pool usable is not raw size but the targeting layer sitting over it. TechCrunch reported that Prolific built more than 300 filters based on census data and other sources so that customers can better tune the sample they are looking for. These prescreeners cover demographics, employment, health status, language, device ownership and prior submission history, and let you approximate nationally representative samples instead of taking whoever shows up first.
Identity and authenticity enforcement: Quality is handled by a system Prolific calls Protocol. Prolific states that Protocol runs 40+ checks to verify identity, and that of applicants entering the funnel just 55% pass and can start studies and tasks. These are vendor-reported figures rather than independently audited ones, but the direction of travel is clear: the platform competes on rejecting people, not on maximising headcount.
AI-generated response detection: The most contemporary quality problem in paid research is participants pasting a chatbot's answer into a free-text box. Prolific states that its proprietary authenticity checks detect AI-generated free-text responses with 98.7% precision, a vendor-reported figure that has not been independently audited. The company backs this with a commercial promise it markets as a 100% Human Guarantee, offering financial compensation where AI agents slip past pre-study checks.
Preference data and human feedback for model alignment: On the AI side, Prolific generates preference data, human feedback, and judgement signals used to align models with intended outcomes, and measures capability, safety and quality through human evaluations, subject-matter-expert verification and rubric design. This is the RLHF and evaluation supply chain expressed as a self-serve product rather than a bespoke vendor contract.
Expert annotation across languages: For work that a general crowd cannot do, verified domain experts generate SFT data, instruction-response pairs, and specialist annotations across more than 80 languages. This matters for teams building non-English capability or needing clinicians, lawyers, engineers or native speakers rather than generalists.
HUMAINE, a public evaluation framework: HUMAINE is a human-preference leaderboard that evaluates frontier AI models based on real-world usage, scoring them across reasoning, communication style, core task performance, adaptiveness, trust, ethics, and safety. Its pool is structured to represent 22 demographic groups across the US and UK. The framework was featured as a conference paper at ICLR 2026, which makes it one of the more methodologically scrutinised artefacts a data vendor has published.
Programmatic and agentic access: Beyond the web dashboard there is a documented Human Data API and an MCP server, letting evaluation runs be triggered from code or from an agent rather than clicked through by hand.
Academic studies needing defensible samples: A psychology or economics researcher needs 400 UK adults balanced on age and sex, each completing a 12-minute task hosted in Gorilla. They configure prescreeners, set the reward, paste the task link, and let Prolific route matching participants. The defensibility comes from being able to describe recruitment and compensation precisely in a methods section — something a convenience sample from social media cannot offer. The discounted academic fee tier makes this viable on grant budgets.
Preference data collection for model alignment: An AI team needs pairwise comparisons between two model outputs across thousands of prompts, from a demographically spread rater pool rather than a homogeneous contractor team. They run the comparisons through Prolific's task tooling and export the results for reward modelling. Because results can be exported as JSONL for direct feed into RLHF, DPO, and reward-model training, the handoff into an existing training pipeline is short.
Safety evaluations and red-team probing before release: Before shipping a model update, a team wants human judgement on refusal behaviour, harmful-content handling and tone under adversarial prompting. Human evaluators with rubric guidance produce scores that automated evals cannot, particularly on subjective dimensions such as condescension, over-refusal or cultural insensitivity.
Product and market research with real customers: A product team validates positioning or pricing perception with a filtered sample — for example, people who have purchased in a category in the last six months. Compared with a traditional panel provider, the appeal is speed and self-service; the trade-off is that Prolific's pool skews toward certain countries and toward people comfortable with online task work.
Benchmarking models against human preference: A team wants to know not just whether a model scores well on static benchmarks but whether people prefer its answers, and whether that preference differs by demographic group. HUMAINE-style evaluation surfaces divergence between groups that a single aggregate score hides.
Agent-in-the-loop human checkpoints: An automated pipeline pauses to ask humans a question — gating a release on an evaluation score, or escalating an ambiguous case — and resumes when results return. This is what the MCP integration is designed for, and it turns human feedback from a scheduled batch exercise into an on-demand call.
Create a researcher account and fund a wallet: Registration is self-serve at the app subdomain. Prolific operates pay-as-you-go: you don't need to sign a contract or pay for a subscription to get started. You add funds, and studies draw down against that balance.
Build your task somewhere else first: Write your survey, experiment or annotation interface in your own tool and get a working link. Make sure it captures the participant ID passed in the URL, and that it redirects back to a completion URL at the end — this is how Prolific knows who finished.
Define your sample with prescreeners: Choose your target population using the filtering layer. Check feasibility before committing: highly specific intersections of filters can shrink your eligible pool to the point where recruitment stalls or takes days rather than minutes.
Set duration and reward honestly: Estimate how long the task really takes and price it against the hourly guidance. Underestimating duration is the most common self-inflicted wound — it lowers your effective hourly rate, attracts complaints and slows recruitment.
Pilot with a small batch: Launch to a handful of participants first. Check that data lands correctly, that completion codes work, and that instructions are unambiguous. Fixing a broken redirect after 400 submissions is expensive; fixing it after 10 is trivial.
Launch, monitor and review submissions: Watch completion times and open text quality as data arrives. Approve promptly — participants are paid on approval, and slow approvals are a frequent complaint. Reject only where you have a defensible reason tied to stated rules.
Export and analyse: Pull your data along with demographic metadata for weighting or subgroup analysis, and record the recruitment parameters you used so the study can be reproduced.
Pay above the floor, not at it: Prolific recommends paying participants at least £9.00 / $12.00 per hour, and the minimum pay allowed on the platform is £6.00 / $8.00 per hour. Studies priced at the floor fill more slowly and attract more rushed responses. Treat the recommended rate as the real baseline and budget accordingly.
Time your task before you price it: Run the task yourself and add a margin for people reading more carefully than you do. Your stated duration determines your effective hourly rate, which in turn determines how fast you fill.
Do not stack filters carelessly: Every additional prescreener multiplies scarcity. If you need three uncommon attributes at once, check feasibility first and be prepared for a longer field period or a larger budget.
Build your own attention checks anyway: Platform-level quality controls reduce bad actors but do not replace instrument design. Include comprehension checks and use response-time distributions to spot inattentive completions.
Design free-text prompts that resist chatbot answers: Ask for personal specifics, concrete recent experiences or reactions to a stimulus shown in the task. Generic essay prompts are the easiest to answer with an LLM and the hardest to police.
Budget the fee, not just the reward: The headline cost of a study is participant pay, but the platform fee sits on top. Model total cost before you commit, and remember the academic and non-profit tier if you qualify.
Approve quickly and communicate clearly: Your reputation with participants affects fill rate. Clear instructions, prompt approvals and fair handling of edge cases make future studies recruit faster.
Pilot every change: Any edit to the task link, redirect or completion code deserves a fresh small pilot. The failure modes are silent and expensive.
Academic researchers: Psychologists, economists, linguists and social scientists needing samples they can describe defensibly in a paper, with ethics-compatible compensation and a discounted fee tier for academic and non-profit work.
AI labs and model teams: Groups sourcing preference data, evaluation scores, safety judgements and expert annotation, particularly those wanting a neutral supplier with demographic breadth rather than an opaque contractor pool.
UX and product researchers: Teams validating concepts, comparing designs or measuring comprehension with people outside their own user base, where speed matters more than longitudinal depth.
Market researchers: Practitioners running concept tests, pricing perception and brand studies who want self-service speed and transparent cost, and who can work within the platform's geographic concentration.
Policy and public-sector researchers: Teams needing population-spread views on policy questions, where transparent recruitment and fair pay are themselves part of the credibility of the output.
Engineering teams building evaluation pipelines: Developers who want human judgement callable from code or from an agent, rather than coordinated as a manual side process.
Participants seeking flexible paid work: People who want short, legitimate online tasks with cash payouts, clear rates and no gift-card schemes — accepting that earnings are supplementary and availability varies.
Web application: The primary interface for both researchers and participants, used for study setup, sample configuration, monitoring, approvals and payouts. Participants access studies through a browser and are redirected out to externally hosted tasks.
Human Data API: A documented programmatic interface for creating and managing studies and retrieving results, suitable for teams embedding recruitment into their own systems.
MCP server: The MCP server works with Claude Code, Claude Desktop, Cursor, and any client that implements the Model Context Protocol spec, so an agent can request human judgement mid-pipeline. Prolific states that MCP tools are discovered automatically, with no SDK or wrapper required.
Structured export formats: Results return as structured JSON and can be exported as JSONL for direct feed into RLHF, DPO, and reward-model training.
Integration with external study tools: Because tasks are hosted elsewhere, the platform is designed to interoperate with common survey and experiment builders via link-based redirection rather than replacing them.
No dedicated native mobile app is documented for researchers: Participants generally use the mobile web experience, and some study types explicitly require a desktop browser. (To be verified)
There is no subscription and no free tier in the usual SaaS sense — you fund studies and pay for the human time you consume. Prolific operates two commercial tracks: a self-serve Platform track described as pay-as-you-go with no monthly fees, and a Managed Services track for expert-led data collection with dedicated quality assurance, which is scoped individually and typically governed by a Statement of Work.
The self-serve economics are unusually transparent for this category. The platform fee is usually 42.8% for corporate customers and a discounted 33.3% for academic and non-profit customers, applied on top of what participants are paid. On the participant compensation side, Prolific recommends paying participants at least £9.00 / $12.00 per hour, and the minimum pay allowed on the platform is £6.00 / $8.00 per hour. One detail that materially affects institutional budgets: VAT is charged on platform fees only, and participant rewards are not subject to VAT, which changes the effective cost for VAT-registered institutions.
For participants the arrangement is inverted and equally explicit. Once a participant has earned at least $6/£6 they can cash out directly from their Prolific account, and payouts arrive as cash through PayPal rather than gift cards or points. Third-party analysis estimates that participants receive roughly 70% of what researchers pay in total, which is consistent with the published fee structure. Enterprise and high-volume arrangements are quoted individually and are not published.
Amazon Mechanical Turk: The long-standing general-purpose crowdwork marketplace. Cheaper per response and far larger in raw task volume, but with weaker identity guarantees and a long-documented reputation for attentiveness problems in research contexts.
CloudResearch Connect: Positioned closely to Prolific for behavioural research, offering vetted participants and screening tooling. Often compared directly on cost per complete and on pool composition.
Traditional panel providers such as Cint or Dynata: Much larger and better suited to nationally representative consumer research at scale, particularly outside the UK and US, but generally less self-service and less transparent about participant compensation.
Scale AI, Surge AI and similar data vendors: Aimed squarely at AI labs, offering managed annotation workforces and bespoke pipelines. Stronger for large, complex, long-running annotation programmes; weaker for fast demographic sampling and academic-style studies.
User research recruiters such as User Interviews or Respondent: Better for qualitative work — moderated interviews and usability sessions with professional or hard-to-reach audiences — where sample sizes are small and per-participant fees are much higher.
The platform fee is the most-cited objection: A fee of 42.8% for corporate users means roughly four in ten pounds of your budget buys platform services rather than participant time. For very large samples this compounds quickly, and it is the recurring complaint in researcher reviews.
Geographic concentration: The pool is densest in the UK and United States. Participants outside the core regions report thinner availability; Trustpilot reviews include requests that more activities be opened to participants in Africa. If your research needs coverage of underrepresented regions, verify feasibility before designing around it.
Quality statistics are vendor-reported: The 55% initial pass rate, the 98.7% AI-text detection precision and the 300k+ pool figure all come from Prolific's own materials and have not, as far as public sources show, been independently audited. Treat them as directional claims rather than verified benchmarks.
Support responsiveness draws criticism from both sides: On Trustpilot the platform holds a TrustScore of 4.5 across 14,854 reviews, with 78% of reviewers awarding five stars and 7% awarding one. Within the negative minority, a recurring theme is that support appeals were occasionally unhelpful when resolving payment disputes.
Access is not open to everyone on the participant side: Joining runs through a waitlist and a verification step; only after being invited does a person complete a roughly 15-minute identity verification and short practice task. Earnings are supplementary and study availability fluctuates.
You still need a separate tool to run the study: Prolific recruits and pays; it does not build your questionnaire or experiment. Budget for a survey or experiment platform alongside it, and for the integration work of passing IDs and completion codes correctly.
Self-selection remains a real constraint: Everyone in the pool chose to join a paid research platform and has practice at completing studies. Filters can balance demographics but cannot fully remove the professionalisation effect inherent to any standing panel.
Legal terms live off the main site: Legal documentation does not sit at the conventional paths; the single entry point is the privacy and legal page, which links out to separate privacy notices, legal terms, and further data protection information. Those documents are hosted on an external Notion workspace, so review them directly before making compliance commitments — specific retention periods and any provisions on secondary use are not restated on the main marketing site. (To be verified)
Not in the sense of a free plan. There is no subscription fee and no contract required to start, but every study costs money because participants are paid for their time and the platform takes a percentage on top. Creating an account and exploring the interface costs nothing; running research does. Budget both the participant reward and the platform fee before you plan a study.
Total cost is participant pay plus the platform fee. The fee is usually 42.8% for corporate customers and 33.3% for academic and non-profit customers, applied to the reward pool. So a study paying participants a total of £700 costs roughly £1,000 for a corporate account and roughly £933 for an academic one. VAT applies to the platform fee only, not to participant rewards.
Prolific recommends researchers pay at least £9.00 / $12.00 per hour, with an absolute floor of £6.00 / $8.00 per hour. Actual earnings depend entirely on how many suitable studies a person is eligible for, which varies by demographics and location. It is best understood as supplementary income rather than a replacement for a job, and third-party analysis suggests participants receive roughly 70% of total researcher spend.
Payouts are made in cash through PayPal. A participant can cash out once they have earned at least $6/£6, and Prolific is explicit that payment is always cash rather than gift cards or points. Researchers approve submissions and approval triggers payment, which is why slow approval is a common source of participant frustration.
Generally no — Prolific is a recruitment and payment layer, not a survey builder. You host the questionnaire, experiment or annotation task in your own tool and share the link, passing participant IDs through the URL and redirecting back on completion. The exception is the platform's own AI task tooling for text annotation and evaluation work.
Through a layered system it calls Protocol: identity verification on entry, ongoing fraud detection across the pool, automatic flagging of implausibly fast submissions, and authenticity checks on free-text answers. Prolific reports 98.7% precision on detecting AI-generated free text and offers compensation under a 100% Human Guarantee when checks are bypassed. These figures are vendor-reported, so pair them with your own attention checks.
Yes, and this is now a major part of the business. The platform supplies preference data, human feedback, safety and capability evaluations, and expert annotation including SFT data and instruction-response pairs across more than 80 languages. It also publishes HUMAINE, a human-preference leaderboard evaluating frontier models, which was featured as a conference paper at ICLR 2026.
Yes. There is a documented Human Data API for programmatic study management, and an MCP server that works with Claude Code, Claude Desktop, Cursor and other MCP-compatible clients, allowing an agent to launch studies and receive structured results without leaving its workflow. Results can be exported as JSONL for direct use in RLHF, DPO and reward-model training.
The pool is concentrated in the United Kingdom and United States, with meaningful coverage across Europe and other regions but noticeably thinner availability elsewhere. Participants in some regions report few eligible studies. If your research depends on a specific country or a rare demographic intersection, check feasibility with the filtering tools before designing the study around that assumption.
The core difference is vetting and compensation policy. Prolific verifies identity, enforces a pay floor, and offers extensive census-informed prescreeners — TechCrunch noted more than 300 filters based on census data and other sources. Mechanical Turk is cheaper and larger but has a long-documented reputation for attentiveness and data-quality problems in research settings. The trade-off is cost against confidence in who answered.
