PublicAI
PublicAI is a Web3 AI data infrastructure that pays a global contributor network to produce human training data — multimodal annotation, dataset validation, and Trajector, which licenses real agentic coding sessions to AI labs and model teams.
What is PublicAI?
PublicAI describes itself as a Web3 AI Data Infrastructure: a decentralized network that produces human training data and pays the people who supply it. The model is a two-sided marketplace, where clients post requests for specialized data and a global community carries out the tasks in exchange for rewards. The home page claims 3.5M+ workers, 300K+ validators, 200K+ builders, a 100T+ dataset size, and 50+ partners; a documentation page dated six months ago adds more than $14 million in client revenue and 1 million contributors. None of these figures are audited. Four arguments structure the pitch: a global workforce reachable without banking barriers, quality control enforced by on-chain staking, cost efficiency, and multimodal coverage spanning text, audio, video, and mapping. Three activity blocks sit underneath: Data Collection through the Data Hunter Chrome extension, Data Labeling through AI pre-annotation followed by human verification, and Model Evaluation. Since August 2026 the flagship product is Trajector, which "licenses complete, consented Claude Code sessions from real development work — prompts, tool calls, diffs, failures, fixes, and commits — collected from open-source developers who get paid for it." A trajectory is structured JSONL carrying the full event order and model metadata, and the pitch is to capture work that would have happened anyway rather than datasets staged for the bounty. Deliveries pass a five-stage acceptance pipeline: CAPTURE (uploads signed by a device-bound key), SCRUB (secrets stripped locally, then verified server-side), VALIDATE (schema, event ordering, temporal consistency), DEDUPE (exact and near duplicates), and REVIEW (sampled scoring for coherence). On the contributor side, an open-source CLI routes the project's API traffic through a local proxy at 127.0.0.1:41100, records it, masks secrets, and uploads the redacted result. Consent is granted project by project: "Nothing is captured until you opt a project in, and every project you have not opted in is structurally incapable of being captured." The older marketplace remains alongside it: Data Hub for dataset validation by multiple-choice vote with consensus rewards in $PUBLIC, and V1 data types spanning image annotation, 2D semantic segmentation, text collection, document transcription, named entity recognition, video annotation, and 3D LIDAR. Ecosystem roles include task publishers, workers, job reviewers, labeling guilds, approvers, and councils. Since 2026, joining the contributor network requires an application, and DataBabies NFT holders get a priority queue.
What it does
- Capture complete agentic coding trajectories: prompts, tool calls, diffs, test results, failures, fixes, and linked commits
- Annotate and label multimodal data through an AI-assisted pass verified by human validators
- Validate datasets by multiple-choice vote on the Data Hub and earn consensus rewards
- Collect social content and GPT conversations passively through the Data Hunter Chrome extension
- Evaluate models with metrics, weakness identification, and scenario-based testing
- Pay contributors in stablecoin, bypassing the international banking system
- Enforce quality through on-chain staking, entry tests, validator SBTs, and slashing
When to use PublicAI / When not to
A quick filter to help you decide if PublicAI is the right fit.
When to use PublicAI
- AI labs and model teams sourcing real agentic coding trajectories rather than tasks staged for a benchmark
- Companies needing multimodal annotation at scale: image labeling, 2D semantic segmentation, video annotation, document transcription, named entity recognition, and 3D LIDAR
- Open-source developers already working with Claude Code who want to monetize sessions they would produce anyway
- Contributors in countries poorly served by international banking, since payouts settle in stablecoin rather than through bank transfers
- Data buyers who require a documented consent chain and an explicit commercial training license
When not to use PublicAI
- Anyone unwilling to have their data resold to commercial third parties: resale is the business model here, not an option
- Contributors who expect to pull their data back later, since the license granted is perpetual and irrevocable once the content is folded into a training dataset
- Buyers who need a published price, as no client-side figure appears anywhere and access runs through a demo request form
- Users who want nothing to do with crypto: withdrawals settle in USDC or USDT on Ethereum or BSC and require a wallet
- Minors and anyone located in Crimea, Cuba, Iran, North Korea, or Syria, all excluded outright by the terms of service
How to use PublicAI
A typical end-to-end flow, from setup to results.
- Pick your side first: contribute data and get paid, or buy data as an AI lab or company
- Contributor developers install the Trajector CLI, a one-line curl command on macOS and Linux, or a ZIP to unpack on Windows, which is still in beta
- Run trajector enable on each project you want captured; nothing outside an enabled project is recorded, because the code path does not exist for it
- Keep coding with Claude Code as usual, while secrets and personal data are stripped on your own machine before anything leaves it
- Watch accepted sessions credit your balance, then withdraw from $10 in USDC or USDT on Ethereum or BSC, with a wallet connected at claim time
- Revoke at any point with trajector disable, disable --purge, doctor discard, logout, or uninstall --delete-data
- Non-developer data contributors apply through beta.publicai.io, giving name, contact details, location, background (LinkedIn or CV, education, skills), motivation, and weekly availability
- Expect applications to be reviewed in batches with a reply announced within a few weeks; DataBabies NFT holders move to the front of the queue
- Data Hunter users install the Chrome extension, sign in with a Data Hub account, and earn while browsing or replying to posts on X
- AI labs and companies go through Book a Demo, a Typeform booking that opens onto a scoped pilot batch
Pros & Cons
Pros
- A consent chain documented end to end, with an explicit commercial license for dataset buyers
- The collection client is open source, so the documentation's claims can be checked against the code
- Secrets are redacted on the contributor's own machine, with a second pass on the server side
- Consent is granular per project, revocable, and backed by documented deletion commands
- Pay is transparent and calculable in advance: a published per-token price, an append-only ledger, and a rate that is never applied retroactively
- Stablecoin payouts remove any dependency on the international banking system, which matters for contributors in poorly served countries
- Quality control is spelled out (five-stage pipeline, staking, slashing, human review of doubtful sessions), and a failed payout costs nothing because the balance is returned in full, fees included
Cons
- No client pricing is published: any figure comes only through a demo request
- The license on submitted content is perpetual and irrevocable, and contributors waive every right over the models trained on their work
- No DPA, no subprocessor list, and no declared hosting country or region; the acronym GDPR never appears on the site and no Article 27 representative is designated
- No postal address anywhere, and three different entities appear depending on the page: PublicAI Technology Co., Limited in the terms, PublicAI Foundation, Cayman Islands in the footer, and MetaBlock US Limited on the token subdomain
- A single published contact channel, one email address, served behind Cloudflare obfuscation
- The Trajector CLI is beta: pre-1.0 releases, Windows still in beta, unsigned macOS binaries, no Homebrew or Scoop, and a documented redaction gap that leaves a secret placed in a JSON key position unmasked
- Joining the contributor network requires an application and a wait of several weeks, crypto exposure is assumed (volatility, full wallet responsibility, irreversible transactions), and disputes go to individual arbitration with liability capped at $100 or six months of payments
Pricing & Plans
PublicAI publishes no client-side price. The home page FAQ states that the marketplace includes a free tier providing a complimentary amount of annotation, with pay-as-you-go billing beyond that volume, and that advanced capabilities move to a plan named PublicAI Pro for Enterprise; no amount is given for either. Billing is not per annotator: team members are unlimited and only the work actually completed is charged. Access to Trajector for AI labs begins with a scoped pilot batch arranged through a demo request, again with no published rate. On the contributor side the flow reverses, since PublicAI is the payer: $0.000001 per token of an accepted session ($1.00 per million), withdrawals open at $10, and network fees of $1 to $2 are deducted from the payout.
- free tier with a complimentary amount of annotation
- then pay-as-you-go billing beyond it
- no published amount
- additional features and enterprise capabilities
- no published amount
- scoped pilot batch
- available on request through a demo booking
- no published amount
Data, GDPR & hosting
A consolidated view of how PublicAI handles your data.
GDPR overview
The acronym GDPR appears nowhere on PublicAI's site, and the company never claims compliance. What exists is a privacy policy in force since 8 January 2026 with a section headed ADDITIONAL DISCLOSURES FOR INDIVIDUALS IN EUROPE, covering the EEA, the United Kingdom, and Switzerland. It states legal bases (contract performance, legitimate interest, legal obligation, and consent that can be withdrawn at any time) and lists rights of access, portability, rectification, erasure, objection, and restriction of processing, exercised by writing to support@publicai.io. The right to lodge a complaint with a local data protection authority is expressly recalled. Missing, by contrast: no Article 27 EU representative is designated, no data protection officer is named, no DPA is published or offered, and no subprocessor list exists. A separate section handles the Californian CCPA regime (sale opt-out, access, deletion, non-discrimination, Shine the Light).
Who owns the data?
Under the terms of service you warrant that you own the rights to the User Content you submit, but you then grant PublicAI, its affiliates, successors, assigns, and third-party partners a non-exclusive, worldwide, royalty-free, fully paid-up, transferable, multi-tier sublicensable, perpetual and irrevocable license over it. You also explicitly and irrevocably waive any ownership claim over the models trained on that content, their weights, their outputs, and the commercial products derived from them. Compensation stops at platform rewards: no royalties, no revenue share. On the Trajector side, data already delivered and paid for stays covered by the agreement accepted when the project was enabled, and deletion requests reach only what precedes delivery.
Reuse rights
Contributors get no reuse rights: what they submit flows one way. PublicAI states that submitted data serves site and service operation, and the training, development, testing, validation, and improvement of AI models, including commercial training by third parties. Named recipient categories cover LLM developers, voice and multimodal model builders, AI companies, translation platforms, voice agents, autonomous vehicle and robotics teams, and research institutions. Data is de-identified, anonymized, or aggregated before any training use, with personal identifiers removed or masked and identity separated from data, though the site itself cautions that no de-identification method is infallible. Recipients are bound by data protection agreements limiting use to specified purposes, requiring appropriate security, and barring re-identification. Trained models, their weights, and their outputs belong to the third-party developers, not to contributors. Non-training uses are also declared: operations, communication, marketing, support, usage analytics, security, and legal obligations. On Trajector, the resale itself is public by design; buyer identity and commercial terms are not.
Data retention & training
Hosting summary
PublicAI declares no hosting country and no hosting region. The privacy policy in force since 8 January 2026 mentions Other Service Providers including web hosting, but names none of them, and no subprocessor list is published. No certification is claimed either, neither SOC 2 nor ISO 27001. What is documented sits on the contributor's side: with Trajector, recordings wait in a local spool on the contributor's own machine, in files readable only by that user account (0600/0700 permissions) under a 2 GiB quota. Redacted batches then travel over HTTPS, authenticated by a device pairing token. The upload destination can be changed only through config.json, never through an environment variable, and any non-standard destination is announced to the user. The domain's resolved IP address is a Cloudflare anycast node, which says nothing about where the data actually lives. In short: precise, verifiable guarantees about the machine the data leaves from, and no published information at all about the infrastructure it arrives on.
Where PublicAI works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting PublicAI.
- The license granted over submitted content is perpetual and irrevocable, and it survives closing your account
- Data already folded into a training dataset cannot be pulled back, which the site itself calls technically unfeasible, and it is resold to commercial third parties whose identity stays confidential
- De-identification is applied before training use, but the site warns that no method is infallible, and a documented redaction gap leaves a secret placed in a JSON key position unmasked
- Because capture runs quietly behind work you would have done anyway, a single per-project opt-in keeps producing data long after the decision has faded from mind: re-read what you enabled, rather than trusting your memory of it
- Collateral can be slashed at 100% with a permanent ban for activity judged malicious, multi-accounting is forbidden, and no reward amount, rate, or APR is guaranteed since the per-token rate is configurable and can change
- Total liability is capped at the amount paid over six months or $100, disputes go to mandatory individual arbitration under Californian law, and class actions are waived
- Use is barred in Crimea, Cuba, Iran, North Korea, and Syria and to OFAC SDN-listed persons; meanwhile PublicAI publishes a Verify tool precisely because scams impersonate it, and three different legal entities appear across the site with no postal address and no consultable register
Setup & Integrations
Technical difficulty
It depends entirely on which side you are on. Enterprise clients face no technical setup at all, since the entry point is a demo request form, and data contributors simply fill in an application. Data Hunter needs only a Chrome extension and a Data Hub login. Trajector is the technical path: a one-line curl install on macOS and Linux, a ZIP to unpack and unblock on Windows, then trajector enable per project. Documented friction includes adding ~/.local/bin to PATH, unsigned macOS binaries, Windows SmartScreen, and no Homebrew or Scoop. The status, doctor, and doctor bundle commands help diagnose problems.
Deployment
Integrations
Supported languages
Behind PublicAI
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement PublicAI.
Frequently asked questions
Is my data secure with PublicAI?
How does pricing work, and can I pay as I go?
Am I billed per annotator?
Which data types are supported?
What is Trajector?
Is my code safe with Trajector?
What do contributors earn, and how do they claim it?
What counts as an accepted session?
What is the withdrawal threshold?
Who can take part, and is there an age limit?
Should you pick PublicAI?
PublicAI makes a claim that is rare in the AI data business, and makes it concretely: a consent chain documented from the contributor's machine to the buyer's dataset. Opening the collection client's source code turns that from a promise into something checkable, since the five-stage pipeline, the local redaction, the per-project opt-in, and the append-only ledger are readable in code rather than asserted in marketing copy. For an AI lab weary of synthetic data and staged benchmark tasks, that provenance is the product. The counterpart weighs on the other side of the marketplace. The license a contributor grants is perpetual and irrevocable, extends to third-party partners through multiple sublicensing tiers, and comes with an explicit waiver of any claim over the models trained on the work. Compensation stops at platform rewards: $1.00 per million tokens, no royalties, no revenue share. A clear bargain, but a final one. Corporate opacity persists alongside it. No postal address appears anywhere, no client price is published, there is no DPA, no subprocessor list, no declared hosting country, and the acronym GDPR is absent from the entire site. Three distinct legal entities surface depending on which page you read. The flagship product is in beta, and joining the contributor network now runs through an application and a wait of several weeks. The verdict therefore depends on which side of this marketplace you stand. As a buyer, PublicAI offers documented provenance that few competitors match, at a price you will only learn by asking. As a contributor, the terms are transparent and the pay is calculable in advance, but what you give up is permanent: read them before you enable your first project.
- Choosing a selection results in a full page refresh.
- Opens in a new window.