The future of identity operations is here. Meet Persona AI →
How AI data platforms verify who is actually doing the work
AI data platforms increasingly rely on specialized contributors to create and evaluate human data for model training. These contributors may label domain-specific data, design training tasks, create grading rubrics, or otherwise provide the expertise needed to train and evaluate frontier models.
Because this work can pay well, it tends to attract fraudsters. Bad actors may impersonate experts, operate duplicate accounts, buy or take over verified contributor accounts, or use other tactics to access assignments they aren’t qualified or eligible to complete.
To protect the integrity of models trained on their data and keep legitimate experts from losing work to fraudsters, AI data platforms are increasingly turning to identity checks during onboarding, upon job or task acceptance, and throughout the account life cycle.
The guide below takes a closer look at the types of fraud AI data labeling platforms face, how platforms can fight back, and questions you should ask as you begin designing a contributor verification strategy of your own.
Why do fraudsters target AI data platforms?
Fraudsters target AI data platforms because assignments can pay quite well, especially for specialized domains. While pay rates vary by platform and assignment, some platforms advertise average hourly pay rates of $80 to $200 per hour, which translates to a full-time gross annualized pay of $160,000 to $400,000.
These financial incentives attract both individual fraudsters and organized groups, including suspected state-linked actors. For example, Google Threat Intelligence has observed suspected North Korean IT workers maintaining accounts on data annotation platforms that use contributors to train AI models.
The distributed nature of contributor work can make fraud harder for AI data platforms to detect and prevent.
Types of fraud affecting AI data platforms
AI data platforms often see the following types of contributor fraud:
| Type of fraud | What it is | When it occurs |
|---|---|---|
| Impersonation | Fraudsters impersonate experts to qualify for high-paying, specialized assignments | During application and platform onboarding |
| Duplicate accounts | Fraudsters open multiple accounts to double-bill for work | During application |
| Organized fraud rings | Multiple fraudsters work together to open fraudulent accounts on a platform, often using shared or synthetic identities | Throughout the contributor life cycle |
| Account buying or sharing | A legitimate expert opens a verified account and then shares it with or sells it to a fraudster | After verification, throughout the contributor life cycle |
| Account takeover | A legitimate expert opens a verified account, which is then stolen by a fraudster | After verification, throughout the contributor life cycle |
| Country-restriction evasion | A contributor lies about their location to qualify for jobs with country restrictions | During application and when tasks are assigned |
Common techniques and tools fraudsters use to carry out these scams include deepfakes, virtual cameras, manipulated or forged (synthetic) identity evidence, VPNs, local proxies, device reuse, and bot farms.
1. Impersonation
Because AI data platforms often reserve high-paying assignments for experts with specific experience or credentials, fraudsters create profiles impersonating real experts. Alternatively, they misrepresent their qualifications on applications and fake professional profiles like LinkedIn.
How platforms can fight it: Background checks and government ID verification paired with selfie verification and/or database verification to protect against stolen or forged documents
Government ID verification links the claimed identity to an identity document. Selfie verification links the ID to the individual completing the application, while database verification links the ID to an official government record. Background checks can be used to verify employment and educational history.
2. Duplicate accounts
Fraudsters open multiple accounts and complete the same assignments under different profiles to receive multiple payments for the same amount of work. Profiles may share some details but often switch up the username, email address, and contact information.
How platforms can fight it: Device and passive signal collection paired with link analysis
Device and passive signals (like IP address, device fingerprint, and browser fingerprint) give platforms a deeper understanding of each profile. Platforms can then use link analysis to identify suspicious connections between profiles and detect when a single individual may be operating multiple accounts. Layering government ID verification and selfie checks can also make it easier to distinguish between a single person signing up for multiple accounts vs. multiple people using the same network or device, as might be the case when multiple legitimate contributors share a household.
3. Organized fraud rings
Multiple fraudsters work together to open and operate fraudulent accounts. They often share and reuse profile components, including names, identity documents, contact information, and selfies. Fraud rings may involve identity suppliers, account operators, and shared infrastructure like networks and devices.
How platforms can fight it: Link analysis augmented by device and passive signal collection
Link analysis surfaces suspicious connections between accounts that may suggest coordinated activity. Device and passive signals provide additional context, helping platforms identify accounts that share IP addresses, network infrastructure, or other attributes.
4. Account buying or sharing, account takeovers
A legitimate contributor opens an account and completes the identity verification process. They then share or sell their login credentials so fraudsters can access the verified account and complete paid work.
Alternatively, a fraudster may steal (take over) a legitimate account, often using social engineering attacks like phishing to compromise the user’s credentials.
How to stop it: Reverification paired with device and passive signal collection
Platforms can reverify a contributor against the government ID and/or selfie used in the original profile verification any time the user logs in from an unrecognized device, IP address, or suspicious location, or when the user engages in suspicious activity. Device and passive signals collected during login and assignment acceptance can help determine when reverification is necessary.
5. Country-restriction evasion
A contributor misrepresents where they’re working from or where they are located to qualify for assignments that are only open to contributors in certain locations or to skirt international sanctions.
How platforms can fight it: Collecting device and passive signals, especially location signals such as IP address, VPN use, and geolocation. The issuing country for any ID collected during verification can also be a helpful signal, along with sanctions screening.
What to think about as you design your verification strategy
There’s no one-size-fits-all approach to verification. The right strategy depends on a platform’s fraud risks, contributor population, and operational requirements.
As you define what verification should look like for your platform and evaluate potential methods or vendors, ask yourself the following:
What aspects of a contributor’s identity do you need to verify, and why? Examples might include their geolocation (to defend against location-based scams), affiliation with an organization such as an employer or educational institution (to defend against expert impersonation), or selfie (to defend against account sharing and stealing).
How much assurance do you need as you verify each attribute? Different verification methods offer different levels of assurance, which can inform the methods you select.
How can you control friction? If verification and reverification introduce too much friction, they can degrade the contributor experience and make it more difficult to attract qualified experts, ultimately resulting in poorer data quality for model training.
How will you determine if additional step-up verification is required? A risk-based approach to verification lets you tailor each contributor’s verification to their risk level, using stronger checks when risk is high and fewer checks when it is low.
Where are the high-risk moments where reverification is most important? It’s not enough to verify a contributor’s identity once during onboarding. Other points in the contributor life cycle, such as when a contributor accepts an assignment or completes it, remain vulnerable to fraud after initial verification. Reverification and other account-based checks help confirm that the person doing the work is still the same verified contributor over the entire life of the account.
Can contributors appeal a failed verification? While uncommon, it’s possible for a legitimate contributor to occasionally fail verification. Options for appeals include allowing contributors to retry verification, reviewing failed applications manually, or investigating edge cases more thoroughly.
How will you secure verification data? Verifying contributor identities may involve collecting sensitive data such as their name, date of birth, government ID, ID number, selfie, and more. It’s important to have a plan for data privacy and security — and to communicate these policies with contributors.
The goal is to protect platform integrity without creating unnecessary friction for legitimate contributors. When well-conceived and thoughtfully implemented, contributor verification helps platforms control access to their contributor networks, prevent fraud, and maintain data quality.
How Persona helps AI data platforms get contributor verification right
AI data platforms face significant challenges when implementing contributor verification. That’s why platforms like Handshake and Labelbox use Persona to establish:
Who’s joining the contributor network
Whether contributors meet identity and location requirements
Whether one person controls multiple accounts
Whether the person who opened an account actually controls it
Whether accounts and profiles are a part of a coordinated operation or fraud network
Persona gives you the flexibility to design a verification strategy specifically informed by your risk appetite and user expectations. Combine government ID, database, and selfie verification to establish and reverify contributor identities. Collect and analyze passive and device signals for deeper context about your contributors. Use Persona’s link analysis tool, Graph, to surface suspicious connections between accounts on your network. Where it makes sense, take advantage of phone and email risk reports and background checks for additional layers of assurance.
Not sure what contributor verification should look like for your platform? Our experts can help you think through these considerations and build the ideal solution for your needs. Reach out today to speak with a member of our team.
The information provided is not intended to constitute legal advice; all information provided is for general informational purposes only and may not constitute the most up-to-date information. Any links to other third-party websites are only for the convenience of the reader.
FAQs
What’s the difference between foundation and frontier AI models?
Toggle description visibility
Foundation models are large AI models trained on broad, general-purpose datasets. Much of that data comes from publicly available sources, including websites. The broad scope of their training data makes these models capable of performing many different types of tasks, which is why many applications are built on top of them — hence the term “foundation.”
Frontier models are advanced models capable of being deployed at the edge of what’s thought to be possible for AI applications. As developers push these models into specialized domains such as healthcare, cybersecurity, finance, and law, they require specialized training data that isn’t readily available from public sources. Often, these companies turn to domain experts to create, label, and evaluate the training data and assess whether model outputs are accurate.
Why do AI data platforms need to verify their contributors?
Toggle description visibility
AI data platforms typically verify contributors for one of three reasons:
- To maintain quality control over training data by ensuring contributors have the relevant experience, expertise, or qualifications
- To protect the user experience by ensuring that assignments are available for qualified contributors and experts instead of going to fraudsters
- To ensure contributors are not located in sanctioned jurisdictions, like North Korea or Iran
How do AI data platforms verify their contributors?
Toggle description visibility
AI data platforms often use multiple verification methods to verify contributors throughout the account life cycle, including government ID verification, database verification, selfie verification, device and passive signal collection, email and phone risk reports, link analysis, and more. The specific methods a platform uses depend on the fraud risks they are controlling for and the dynamics of their specific user population.
How do platforms detect duplicate contributor accounts?
Toggle description visibility
Platforms often use link analysis to surface suspicious connections between accounts, which may indicate that multiple accounts are owned or operated by the same individual. Examples include shared IP addresses, physical addresses, device fingerprints, browser fingerprints, geolocation data, or payment details — though there are exceptions in which multiple accounts may legitimately share these signals.
What triggers reverification on a contributor platform?
Toggle description visibility
Each platform may trigger reverification for different reasons. Common triggers include when a contributor:
- Logs into their account from an unrecognized or suspicious device, IP address, or location
- Logs back in after a prolonged absence
- Attempts to change critical account information, like payment details, login credentials, or contact information
- Accepts or applies for a restricted or high-value project
Reverification can be active or passive. Passive reverification may include collecting and analyzing device signals in the background of a session; active reverification may involve requiring the contributor to resubmit their ID or selfie for comparison against evidence already on record.
How do platforms verify where a contributor is actually working from?
Toggle description visibility
Platforms often analyze multiple risk signals to verify where a contributor is located, both during onboarding and while they complete assignments. Signals can include:
- ID type and issuing country
- IP address
- Geolocation data
- Device fingerprint
- Browser fingerprint
- VPN usage
What happens if a contributor fails identity verification?
Toggle description visibility
When a contributor doesn’t pass an identity verification check, some platforms may allow them to retry a certain number of times before locking them out of the account or platform. Others may offer fallback verification methods, appeals, or manual human review.
How can an AI data platform keep onboarding and reverification smooth for legitimate contributors?
Toggle description visibility
Platforms can protect the contributor experience by implementing targeted friction: dialing verification requirements up when assessed risk is high, and dialing them down when assessed risk is low. Instead of a one-size-fits-all verification process, they tailor verification to each contributor based on the levels and types of risk present at each interaction.