On this page
Short answer
Can AI replace research assistants? Not as a whole job, based on the public evidence as of November 2026. The products cited here document AI doing the desk half of the work: searching sources, screening papers and extracting data, running statistical analysis in code, and drafting tables, charts and cited reports. In each case the documentation tells you to check the output, and an independent audit shows why. The other half of the job, working with study participants, consent, protocols, quality control and presenting results, has no cited evidence of an agent doing it.
So if you run a lab, a research team or an analysis group, the realistic change is that one research assistant can hand first drafts of the reading, coding and analysis to software and spend more time checking it. Whether that means you need fewer assistants is not something the sources establish.
Some public data for scale. O*NET's closest occupation is social science research assistants (code 19-4061.00), whose reported job titles include research assistant, graduate research assistant and clinical research assistant. The U.S. Bureau of Labor Statistics (BLS) puts the national median annual wage at $61,990 for May 2025; O*NET, citing BLS, gives $29.81 an hour and 38,400 employees. O*NET, citing BLS 2025–2035 projections, rates growth as faster than average (5% to 6%) with 4,900 projected openings. Those are BLS figures for the occupation, not a statement about any single team.
One study looks at the question directly. Scott Wallsten of the Technology Policy Institute, a think tank, had language models read the acknowledgments of all 11,718 NBER economics working papers from 2018 through October 2026. In his working paper, the share thanking a research assistant fell from 40% before ChatGPT to 37% after. He notes that a thank-you reflects norms as well as hiring, that papers appear after the research is done, and that the data may not yet show the effects of the newest tools. It measures credits in one field, not jobs.
Task by task
The table below starts from the 23 task statements O*NET lists for social science research assistants, merged into 12 tasks a research lead would recognize. A task is marked for an agent only where a cited product documents an agent doing it. Where O*NET lists the task but no cited source shows an agent doing it, the task is marked human, and the note says so.
Agent
0of 12
Agent with review
5of 12
- Search the web and library databases for sources
- Screen papers and extract data from them
- Code, clean and transform data for analysis
- Run descriptive and multivariate statistical analyses
- Prepare tables, charts and written summaries of results
Human
7of 12
- Build and maintain large research databases
- Design and run quality control procedures
- Recruit, screen, schedule and track study participants
- Obtain informed consent from participants
- Administer tests, interview participants and supervise interviewers
- Edit protocols, help design surveys and write grant proposals
- Consult on research needs and present findings to groups
12 tasks: 0 agent, 5 agent with review, 7 human
| Task | Who does it | Why |
|---|---|---|
| Search the web and library databases for sources | Agent with review | OpenAI's deep research and Elicit's Research Agent document planning a search, reading many sources and returning cited output, and DeepTRACE found unsupported statements in deep research reports, so a person checks citations. |
| Screen papers and extract data from them | Agent with review | Elicit documents screening criteria you can agree with or edit and extractions supported by quotes, and offers AI as a second reviewer beside humans. |
| Code, clean and transform data for analysis | Agent with review | OpenAI documents ChatGPT running Python transformations on uploaded files and summarizing outliers, and tells users to review the generated code and assumptions. |
| Run descriptive and multivariate statistical analyses | Agent with review | OpenAI documents ChatGPT running Python-based statistical analysis, notes it may pick a method that does not match your intent, and tells users to review it. |
| Prepare tables, charts and written summaries of results | Agent with review | ChatGPT's data analysis creates tables and charts, and deep research and Elicit produce cited reports; a person checks them before they go into a manuscript. |
| Build and maintain large research databases | Human | This is an O*NET task; no cited source documents an agent managing a research database over time, so it stays human here. |
| Design and run quality control procedures | Human | This is an O*NET task; no cited source documents an agent setting up research quality control, so it stays human here. |
| Recruit, screen, schedule and track study participants | Human | These are O*NET tasks; no cited source documents an agent doing them, and they involve direct contact with people. |
| Obtain informed consent from participants | Human | Under 45 CFR 46.116, for research covered by that policy, an investigator obtains legally effective informed consent; no cited source shows an agent doing it. |
| Administer tests, interview participants and supervise interviewers | Human | These are O*NET tasks; no cited source documents an agent running them in a study. |
| Edit protocols, help design surveys and write grant proposals | Human | These are O*NET tasks; no cited source documents an agent drafting IRB protocols, survey instruments or grant proposals, so they stay human here. |
| Consult on research needs and present findings to groups | Human | These are O*NET tasks; no cited source documents an agent doing them. |
On that standard, 0 of 12 tasks are ones an agent completes with only spot checks, 5 of 12 are ones a documented tool does with a person reviewing the output, and 7 of 12 stay with a person. The five are searching for sources, screening and extracting from papers, coding and cleaning data, running statistical analyses, and preparing tables, charts and summaries.
The evidence for those five comes from the products below. Each describes a person reviewing the work.
- OpenAI's deep research in ChatGPT. According to OpenAI's help center, it proposes a research plan you can edit, searches the public web, specific sites or your uploaded files, and returns "a structured report with citations or source links so you can verify the information."
- ChatGPT data analysis. OpenAI documents that ChatGPT can run "Python-based calculations, transformations, and statistical analysis" on uploaded files and create tables and charts. It tells you to review the generated code, outputs and assumptions before relying on the result, and says it may choose a method or chart that does not match your intent on the first try.
- Elicit. Elicit, a research software vendor, says its Research Agent can "search the literature, extract data, analyze figures and tables, model data, and create charts," with every claim linked to a supporting passage. For systematic reviews it says it suggests screening criteria that you can agree with, edit or add to, supports every extraction with quotes, and can act as the second reviewer alongside human reviewers. Elicit also publishes its own accuracy figures, such as 96% data extraction accuracy; those are the vendor's numbers.
How reliable is the output? DeepTRACE, a study in the ICLR 2026 proceedings, audited generative search engines and deep research configurations from several public providers. Its authors report that deep research configurations "still exhibit large fractions of unsupported statements, with citation accuracy ranging from 40–80% across systems." That is the main reason every documented task here keeps a review step: a cited sentence may not be supported by its source.
A worked example
This is an illustration, not a reported case. Assume, for this example, that a principal investigator asks for a summary table of randomized trials on a named intervention published since 2020. The tool searches, proposes screening criteria, and returns 18 included studies with sample sizes and outcomes, each with a supporting quote. The research assistant opens a sample of five papers, finds that one sample size was taken from an abstract that differs from the results table, and corrects it. They also add two trials they knew of that the search missed. The tool did the first pass of search, screening and extraction. The person checked it against the papers and owns the table that goes to the investigator.
Jobs you can hand to an agent today
The reviewable tasks above group into separate jobs, each with its own page covering what to give the agent, a brief you can copy, risks and how to run a trial.
- JobSupervisedAI agent for literature review: what it does and needs
An AI agent can search databases, screen papers against your criteria and extract data with supporting quotes. A reviewer still verifies every citation, settles inclusion calls and owns the conclusions.
- JobSupervisedAI deep research agent: what it can do for reports
An AI research agent can turn a question into a plan, read many sources and draft a cited report. A person still sets the question, checks every claim that matters against its source and owns the conclusions.
- JobAssist onlyAI agent for spreadsheet analysis: what it can do
An AI agent inside Excel can summarize a dataset, add formula columns and PivotTables, flag outliers and trace formula errors. Both vendors cited here tell you to verify the output, so a person checks the numbers before anyone relies on them.
Ready
0of 3
Supervised
2of 3
Not yet
0of 3
- Literature review covers searching, screening and extracting data from papers.
- Deep research reports covers turning a question into a cited report from many sources.
- Spreadsheet analysis covers cleaning, analyzing and charting data from a file.
If you are deciding where to start, choose the literature search or the report on a topic your team already knows well, so a wrong citation is easy to spot. We recommend fixing in advance how many citations you will open per report and how many unsupported claims you will accept, and counting them on the first few jobs.
What stays human
Four groups of work stay with a person in the sources read for this page.
Checking the desk work. Every tool cited above tells you to verify its citations or review its code, and DeepTRACE's 40–80% citation accuracy range shows the check is not a formality. Someone who knows the subject has to open the sources and rerun or read the analysis. In evidence synthesis, Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence issued a joint position statement that, according to Cochrane's summary, "calls for human oversight, transparency, and justification when AI is used" and expects synthesists to report AI use transparently and assume responsibility for their synthesis.
Participants and consent. O*NET lists recruiting, screening, scheduling and tracking participants, administering tests and interviews, and obtaining informed consent. No cited source shows an agent doing these in a study. For research covered by the U.S. federal human-subjects policy in 45 CFR part 46, section 46.116 says that before involving a human subject, "an investigator shall obtain the legally effective informed consent of the subject or the subject's legally authorized representative" (current text on eCFR; 2025 edition on GovInfo). Which rules apply depends on your institution, funder and jurisdiction; ask your IRB or counsel.
Protocols, surveys and grants. Editing and submitting protocols, helping design questionnaires and writing grant proposals are O*NET tasks with no cited agent evidence. We recommend that the investigator who owns the study reviews any draft of these, since they fix the study's design and budget.
Running the project. Building and maintaining research databases, designing quality control, tracking supplies and participant payments, consulting on what research is needed and presenting findings to a room are all O*NET tasks with no cited agent evidence.
A setup the cited documentation supports is AI tools doing first passes of search, extraction, analysis and drafting, and a research assistant checking that work and doing everything that involves participants, protocols and the running of the study. If you want to hire an agent for one of these jobs when the job board opens, sign up below. The guide on how to hire an AI agent covers scope, trials and acceptance.
Sources
- 19-4061.00 - Social Science Research Assistants
O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026
- OEWS national annual median wage, Social Science Research Assistants (series OEUN000000000000019406113)
U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026
- 45 CFR 46.116 General requirements for informed consent
Electronic Code of Federal Regulations (eCFR)PrimaryAccessed Nov 14, 2026
- 45 CFR 46.116 General requirements for informed consent (Code of Federal Regulations, 2025 edition)
U.S. Government Publishing Office (GovInfo)PrimaryAccessed Nov 14, 2026
- Deep research in ChatGPT
OpenAI Help CenterPrimaryAccessed Nov 14, 2026
- Data analysis with ChatGPT
OpenAI Help CenterPrimaryAccessed Nov 14, 2026
- Research Agent
ElicitVendorAccessed Nov 14, 2026
- Systematic Literature Reviews
ElicitVendorAccessed Nov 14, 2026
- DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence (ICLR 2026)
International Conference on Learning Representations (Venkit et al.)PrimaryAccessed Nov 14, 2026
- Is AI Replacing Human Research Assistants? Evidence from NBER Working Papers
Technology Policy Institute (Scott Wallsten)PrimaryAccessed Nov 14, 2026
- Setting the standards for responsible AI use in evidence synthesis
CochranePrimaryAccessed Nov 14, 2026
What we could not verify
- Independent accuracy figures for any single named product on research-assistant work. DeepTRACE reports citation accuracy as a 40–80% range across the systems it tested, and Elicit's accuracy figures are its own; neither is an independent rate for one product.
- How many research assistant hours these tools save, or whether employers hire fewer research assistants because of them. The only direct evidence found is one think-tank working paper on acknowledgments in NBER economics papers, which measures thanks, not hiring.
- Production evidence of agents managing research databases, running quality control, recruiting or tracking participants, administering tests or interviews, editing IRB protocols, writing grant proposals or presenting findings. Those tasks are marked human for that reason, not because an agent provably cannot do them.
- Published per-task prices for an agent covering this role. The tools cited are sold as subscriptions with plan-dependent limits, so no cost comparison is given.
Related
- AI agents for research: jobs they do and how to hire one
- AI agent for literature review: what it does and needs
- AI deep research agent: what it can do for reports
- AI agent for spreadsheet analysis: what it can do
- Can AI replace market research analysts? Task by task
- Can AI replace data analysts? A task-by-task answer
- How to hire an AI agent: scope, brief, trial, supervision