Candidate fraud in data science and AI hiring: borrowed portfolios, outsourced take-homes, proxy interviews and how to check
On this page
- Why are data and AI roles a target?
- What fraud patterns show up in data and AI hiring?
- How do you check a portfolio, papers and profiles?
- How do you run a take-home that is hard to outsource?
- How do you verify degrees and certifications for these roles?
- How do you confirm the person you hired is the person you interviewed?
- What should data and AI teams do in the first weeks?
- How do you keep these checks fair?
- What should staffing agencies placing data and AI talent do?
- What should you do when something doesn't match?
- Data and AI hiring fraud checklist
- Questions people ask
Candidate fraud in data science, machine learning and AI engineering hiring usually takes one of four forms, as of November 2026: a portfolio that belongs to someone else (copied repositories, notebooks or papers), a take-home assignment done by another person or against the stated rules, a proxy or face-swapped interview, and a false identity used to get a remote job with access to data, models and code. The FBI has warned that deepfakes and stolen personal information are used to apply for remote IT, programming and "database" jobs, some with access to "proprietary information," and that North Korean IT workers have copied employers' code repositories to their own accounts. The checks that work are a live walkthrough of the candidate's own work, an identity check at a fixed stage, issuer verification of credentials, and least-privilege access in the first weeks.
The general checks are on other pages: how to spot a fake GitHub profile, how to verify take-home assignment authorship, deepfake candidates in video interviews and North Korean IT worker red flags. This page is what changes when the role is data or AI.
Why are data and AI roles a target?
Because the job comes with access. Data scientists and machine learning engineers often get read access to customer data, training data and model code in their first weeks, and most of the work can be done remotely. The FBI's 2022 warning on deepfakes in remote job applications said the positions reported "include information technology and computer programming, database, and software related job functions," and that some "include access to customer PII, financial data, corporate IT databases and/or proprietary information" (FBI IC3, I-062822-PSA). The Justice Department's June 30, 2025 announcement of actions against North Korean remote IT worker schemes said the workers "gained access to sensitive employer data and source code," including from "a California-based defense contractor that develops artificial intelligence-powered equipment and technologies," and that the stolen data included information marked as controlled under the ITAR (Justice Department, June 30, 2025). Those are charges and allegations described by prosecutors; they show where the risk sits, not how common it is.
What fraud patterns show up in data and AI hiring?
| Pattern | What it looks like | The check that catches it |
|---|---|---|
| Borrowed portfolio | Repositories, notebooks or competition write-ups copied from someone else; a polished project the candidate cannot explain | Live walkthrough of one project; commit history and authorship; questions about rejected approaches |
| Claimed publications | Papers listed on a resume or a self-maintained profile that the candidate did not write, or an inflated author role | Check the paper's author list and affiliation; ORCID source labels; ask about their specific contribution |
| Outsourced take-home | A strong submission followed by a walkthrough the candidate cannot follow | Rules in the brief; a live extension of the submission; see the take-home authorship page |
| Proxy or face-swapped interview | Lip movement out of step with the audio, a different voice between rounds, a different person on day one | Identity check at a fixed stage; same person confirmed at every round and at onboarding |
| False identity for remote access | Stolen or synthetic identity, laptop shipped to an address that is not the candidate's, remote desktop tools on the work machine | Identity verification, Form I-9 done properly, device and access monitoring |
| Inflated credentials | A PhD that was never completed, a cloud machine learning certification that does not verify | Registrar or degree verification; the issuer's badge or certificate verification |
How do you check a portfolio, papers and profiles?
- Make the walkthrough the test. Pick one project and ask for the data source, how they handled leakage and missing values, the baseline they compared against, the approach they abandoned and why, and what they would do with another month. People who did the work answer with specifics; people who borrowed it answer with the README.
- Read the code history, not the stars. A repository created in one commit, or a fork presented as original work, is a question to ask, not a verdict; the details are in how to spot a fake GitHub profile.
- Treat Google Scholar profiles as self-reported. Google says authors set up and maintain their own profiles and can "Add article manually" or "make changes by hand"; a verified email shows an institutional address, not authorship of each paper (Google Scholar Citations).
- Use ORCID's source labels. Each item on an ORCID record shows its source: the researcher, or a client application such as a university or publisher system that added it (ORCID documentation). An item added by an institution carries more weight than one the researcher typed in.
- Check the paper itself. The author list and affiliation on the published version, and for a key claim, a short call with a co-author or former supervisor found through the institution, not through the candidate.
How do you run a take-home that is hard to outsource?
- State the AI rule in the brief: allowed, allowed with disclosure, or not allowed. Using an assistant within the rules is not fraud.
- Use a small, specific dataset and a question that needs judgment (which metric, which trade-off), not just a model score.
- Ask for a short written note of what they tried, what failed and what tools they used.
- Follow with a live session where the candidate extends the submission: a new feature, a changed constraint, an error analysis on a slice of the data.
- Score the live session, not only the submission, against the same rubric for every candidate.
More detail, including which authorship checks are unreliable, is in how to verify take-home assignment authorship and how to evaluate a take-home assignment.
How do you verify degrees and certifications for these roles?
- Degrees. A PhD or master's that the role or the pay level depends on should be verified with the registrar or its service; see how to verify a candidate's degree and, for unfamiliar schools, how to spot a diploma mill.
- AWS certifications. AWS tells employers to "ask the earner for an AWS Certification digital badge link"; each badge has an option to verify it that shows the issue date and issuer, and AWS does not publish a list of certified people (AWS Certification FAQs).
- Other vendor and online-course certificates. Check each with its issuer's own verification route; course completion certificates show attendance, not a proctored exam. See how to spot a fake professional certificate.
How do you confirm the person you hired is the person you interviewed?
- Check identity at one fixed stage for every finalist, with a live look at a government photo ID, and confirm the same person at each later round and on day one; see the candidate identity verification checklist.
- Watch for the FBI's interview signs. The 2022 PSA describes on-camera "actions and lip movement" that "do not completely coordinate with the audio," and the January 2025 PSA says North Korean IT workers "have been observed using artificial intelligence and face-swapping technology during video job interviews" (I-012325-PSA). Handle a suspicion as in deepfake candidates in video interviews.
- Ship equipment only to the verified address and complete Form I-9 properly; the FBI's October 2026 PSA says facilitators are paid to "appear on camera, participate in job interviews," and keep work devices connected from their own homes (I-101326-PSA).
What should data and AI teams do in the first weeks?
The FBI's January 2025 PSA recommends steps that fit data and model access closely:
- Practice "the Principle of Least Privilege": grant access to the datasets and repositories the first project needs, and widen it later.
- Monitor for data leaving "through easily accessible means such as shared drives, cloud accounts, and private code repositories." The PSA notes these workers "have copied company code repositories, such as GitHub, to their own user profiles and personal cloud accounts."
- Watch for remote desktop software and for "multiple logins into one account in a short period of time from various IP addresses."
- Keep identity verification going "throughout the employment of any remote worker," not only at hire.
If the role touches export-controlled technology, the access and citizenship rules come from the export regulations and the contract, not from a hiring preference; see candidate fraud in government contractor hiring.
How do you keep these checks fair?
- Same checks, same stage, every candidate, not only for people with foreign degrees or thin public profiles. Many strong data scientists have no public code because their work belongs to their employer.
- Never screen on national origin, accent or citizenship status. Ask only the lawful work-authorization questions (see citizenship status discrimination in hiring).
- Use the FCRA when a screening company verifies identity, degrees or employment (see the FCRA background check process).
- Separate "used AI" from "lied." A candidate who followed your AI rule and can explain the work has done nothing wrong.
What should staffing agencies placing data and AI talent do?
- Run the identity check and the portfolio walkthrough yourself before you submit, and tell the client what you checked.
- Confirm the same person at the client interview and on the first day, and agree with the client who ships equipment and to where.
- Check sub-vendor candidates the same way as your own; see candidate fraud in IT staffing.
What should you do when something doesn't match?
- Pause the process or the access, without accusing anyone.
- Verify through an independent channel: the registrar, the certification issuer, the paper's co-author, the identity document checked live.
- Ask the candidate neutrally to explain the mismatch.
- Escalate to HR, legal and security when identity or data access is involved; see what to do if a candidate used a proxy.
- Report suspected identity fraud or North Korean IT worker activity to the FBI at ic3.gov, and keep the record; see how to report candidate fraud.
Data and AI hiring fraud checklist
[ ] AI-use rule written into the take-home brief
[ ] Live walkthrough of one portfolio project, scored on a rubric
[ ] Live extension of the take-home submission
[ ] Key publications checked on the paper itself; ORCID source noted
[ ] Degree the role relies on verified with the registrar
[ ] Certifications verified with the issuer (badge link for AWS)
[ ] Government photo ID checked live at a fixed stage
[ ] Same person confirmed at each round and on day one
[ ] Equipment shipped only to the verified address
[ ] Least-privilege data and repository access for the first project
[ ] Each check documented: date, source, who, result
Not legal advice. FBI, Justice Department and vendor pages checked as of November 2026.
Questions people ask
What kinds of candidate fraud happen in data science and AI hiring?
Four patterns come up: portfolios that belong to someone else (copied repositories, notebooks or papers), take-home assignments done by another person or in a way the rules did not allow, proxy or face-swapped interviews, and false identities used to get remote jobs with access to data and code. The FBI has warned that deepfakes and stolen personal information are used to apply for remote IT, programming and database jobs, and that North Korean IT workers have copied company code repositories to their own accounts.
How do you check whether a data science portfolio is really the candidate's?
Ask the candidate to walk you through one project live: the data, the choices they rejected, what broke and what they would change. Then check what the platforms show: Google Scholar profiles are created and edited by the author, so a listed paper proves nothing on its own, while an ORCID record shows the source of each item, either the researcher or an organization's system. Confirm authorship with a co-author or the publication itself where it matters.
Is using AI on a data science take-home assignment cheating?
Only if your instructions said it was not allowed. Write the rule into the brief (allowed, allowed with disclosure, or not allowed), ask candidates to state what they used, and test understanding in a live walkthrough of the submission. A candidate who used an AI assistant within the rules and can explain every step has not committed fraud.
How can I verify an AWS machine learning certification?
Ask the candidate for the AWS Certification digital badge link. AWS says each badge has an option to verify it that shows the issue date and the issuer, and that AWS does not keep a public list of certified individuals, so there is no lookup by name. Other cloud and vendor certifications have their own verification routes; check each with its issuer.