Where Is the AI Scientist for the Social Sciences?

This article examines a specific gap: agents that connect literature, theory, data and writing for a social-science project on your own computer. It compares the announced scope of the tools linked below, then explains Scholardo's approach and limits. It is written by the product's maker, not an independent product review.

For the practical setup, start with a built-in research session. For sensitive material, review what leaves the computer; for manuscript checking, see what the export check can establish.

Say you study how people find work after a factory closes. You have forty interview transcripts, a survey export, three years of reading, and a consent form that told every participant their words would stay with the research team. Then you read about the AI systems being built to do research.

When Google introduced its AI co-scientist, it reported testing the system "in three key biomedical applications: drug repurposing, proposing novel treatment targets, and elucidating the mechanisms underlying antimicrobial resistance." Sakana describes The AI Scientist as "a fully AI-driven system for automated scientific discovery, applied to machine learning research." FutureHouse is "building AI agents to automate research in biology and other complex sciences." Microsoft offers Discovery to industries "including chemistry and materials, silicon design, energy, manufacturing, and pharma." Anthropic's AI for Science program has "a particular focus on biology and life sciences applications."

None of those five announcements mentions psychology, sociology or political science.

The gap is not that researchers don't use AI ​

It would be easy to read that list as a verdict on who uses AI. It isn't one. Measuring AI use across research, Gao and Wang found that "the use and benefits of AI appear widespread throughout the sciences, growing especially rapidly since 2015" (Nature Human Behaviour, 2024). Writing about the social sciences specifically, Bail put the open question plainly: "it is not yet known how such tools might transform social science research" (PNAS, 2024).

So the gap is not access to AI in general. It is the shape the AI takes when it arrives.

The shape on offer today is a row of assistants, one per step. A deep-research service that searches the web and writes a report. A literature tool that finds and summarizes papers. An AI feature inside qualitative analysis software that suggests codes. Each of them is useful. Each of them stops at the edge of its own step, and you carry the work across to the next one: export, import, explain the project again.

The natural sciences are being offered something else: an agent meant to work across a research project rather than inside one step of it. That is the thing I could not find, among the tools I examined, for a study like the one above.

And a social-science version of that agent would meet a constraint before it does anything clever. Much of the data is about people, and those people were told where it would go.

What a study like this asks of a tool ​

Nothing about the interview study in the opening is exotic. Walk through it, though, and five requirements fall out. Each is one you can hold a tool against.

1. The data stays where participants were told it would ​

Some human-subjects offices now say this in writing. Lehigh University's guidance on generative AI and human subjects research states: "It is best practice to de-identify human subjects data before using AI whenever possible." It adds that removing names is not the end of it: "Even when identifiers are removed, AI tools may infer personal information from patterns in the de-identified data." Frostburg State University's AI use policy lists among the added risks "the sharing of information with a third party not approved to be involved in the research study," and says that "generative AI may only be used to assist with data analysis under situations approved by the IRB."

That is two institutions, not a universal rule, and your own board will have its own wording. The direction is the same, though: every service your transcripts pass through is one more party to account for. A tool that needs your data uploaded before it can help starts on the wrong side of that line.

2. The theory is part of the work ​

In much of social science, the deliverable is not a finding alone but an account of it: a construct, how it was measured, the mechanism that links it to an outcome, the conditions under which the link holds. "Trust in local institutions" becomes a set of survey items. A pattern across interviews becomes a proposition someone else can test.

A tool that stores papers and notes holds the evidence but not the argument built on it. So the second requirement is that concepts, constructs and the relations between them (this supports that, this contradicts that, this measures that) are things you can record, revisit and argue with, not sentences buried in a draft.

3. Words and numbers live in one project ​

Plenty of studies mix them: the interviews explain what the survey can only count. When they sit in different tools, the link between a quote and a statistic lives in your head. The requirement is a project that holds both, and analyses that can be run again from their saved inputs. Lehigh's guidance asks for the same thing in its own terms: researchers should "describe how research methods using AI can be replicated."

4. Every claim can be traced back ​

"AI may yield biased and inaccurate information," the same Lehigh page notes. The answer is not to stop using it but to keep the chain intact: which passage supports which claim, which source stands behind which citation. Citation errors did not start with AI. Reviewing studies of articles in medical journals, Mogull recalculated the rate of incorrectly cited assertions at 14.5% (PLOS ONE, 2017). An assistant that paraphrases freely only adds links to that chain. The requirement is that anything the AI hands you points back to the exact words it came from.

5. The writing survives a reviewer ​

The work ends as a manuscript, a report or a thesis chapter, written to a style guide such as APA and read by someone who may check. The requirement is that the draft cites the works in your own library, in the style you chose, and is checked against those sources before it leaves your hands.

Hold those five against the tools social scientists use today. Several of them meet more than one.

What social scientists have today ​

Credit first. Several of these tools are good, and some reach further than the "row of assistants" picture suggests. Here is what each says it does, in its own words.

Deep research services. Google says Gemini Deep Research can "browse up to hundreds of websites and even your Gmail, Drive and Chat" and "create insightful multi-page reports in minutes." Perplexity says its Deep Research "performs dozens of searches, reads hundreds of sources," and delivers "a comprehensive report." For a fast map of a literature you don't know yet, they are hard to beat. They are built around search, and your study enters them the way anything does: by being uploaded to, or stored in, that service.

Elicit. Elicit makes the broadest claim of the literature tools: "Elicit can support you from beginning to end." Its research agent "[g]athers evidence across scientific databases, academic journals, public data sources, and internally uploaded documents" and "[c]reates fully cited artifacts like reports, tables, figures, bioinformatics analyses, calculations, and slides." Your own documents come in by upload.

NVivo. NVivo now comes in two forms. NVivo Desktop is, in Lumivero's words, its "established qualitative data analysis platform installed on your computer." NVivo AI Cloud is "a separate, browser-based platform that integrates AI-guided assistance throughout the research process, from setting up a project and developing research questions to exploring and analyzing data." Its AURA chat is described as "agentic": it "answers questions grounded in your project data" and "provides click-through citations back to source material." The same post describes AI Cloud this way: "From study design to actionable findings, every part of the research process lives in one project." Of everything here, this is the closest answer to the study in the opening. It lives in the browser.

ATLAS.ti. Its AI Coding is "fueled by OpenAI's GPT model." ATLAS.ti is direct about the data path: "If you do not use AI Coding, none of your data will be uploaded to OpenAI," and before you start, it "will ask for your consent before uploading data to servers by ATLAS.ti, as well as OpenAI servers."

MAXQDA. The strongest answer to the first requirement comes from here. MAXQDA AI Assist can "[r]un AI Assist on a large language model hosted on your own computer or server," and "[u]se all of MAXQDA's AI features without sending any data to the cloud, ideal for sensitive data or strict ethics requirements." Two honest notes from the same page: "Speed and quality depend on your hardware and the chosen model," and local AI is "available with AI Assist Premium." If your study is a qualitative analysis under strict data terms, look at this first.

Zotero. Zotero describes itself as "a free, easy-to-use tool to help you collect, organize, annotate, cite, and share research." It runs on your own computer, and it is the reference library much of this work already rests on.

What is left over ​

That is what the tools say. What follows is how I read them against the five requirements.

Each one covers a real part of the study. MAXQDA keeps qualitative analysis on your machine. NVivo AI Cloud carries an agentic assistant across a qualitative project, in the browser. Elicit runs from question to cited report, on documents you upload. Zotero holds the library and the citations, locally.

What I could not find among them is the combination: an agent, or a team of them, that works across the literature, the data, the theory and the writing, and checks the draft before it leaves, all inside a project on your own computer. MAXQDA's local AI is the closest on privacy, but it is an assistant inside an analysis program rather than an agent that runs the project. That distinction is mine, not MAXQDA's. NVivo AI Cloud is the closest on scope, and it is a hosted service.

That gap is narrower than "nobody has built anything for social scientists." It is also the one the opening study falls into. So here are the five requirements again, as questions you can put to any tool, including the one I build.

Five questions for any research tool ​

1. Can the AI work on your study without your data leaving your computer?
2. Can it hold your theory — constructs, mechanisms, the relations between them — as something you can reopen, rather than as sentences in a chat?
3. Can the interviews and the numbers live in one project, with analyses you can run again?
4. Does everything it hands you point back to the exact words or page it came from?
5. Does it cite from your own library, in your style, and check the draft before it leaves?

For each one, ask a second question too: three months from now, when a reviewer asks, where is the thing it produced, and can you open it?

How Scholardo answers them ​

Scholardo is the tool I build, so read this section as the claim it is. I have tried to answer each question with what the product does today, including where the answer is "no." Scholardo's AI is not one agent but a team of three, and each works with you in a different way. The Conversation Agent is on call: you ask, it acts. The Background Agent works around the clock: it proposes, you approve. The Editing Agent works on demand: you select, it edits. All three report to you. Some answers below describe how the agents are instructed to work: Scholardo ships them with skills, plain files you can read and edit, and what follows is what those skills tell them to do.

1. The data stays on your computer. A Scholardo project is an ordinary folder on your own disk, and the research data are plain files in it. The agents can run on models served on your own machine through Ollama, with no API key. With your sources and data already in the project, the work does not need to leave the machine, and Scholardo sends no telemetry. One exception: the license renews online every few days. Fully offline, the app keeps working for at least two weeks; after that it becomes read-only until it connects once.

Three months later: the transcripts are where you put them.

2. The theory is a set of files, not a chat. Each project keeps a Synthesis: concepts, claims and findings, and an academic project can also hold theories, mechanisms, propositions, variables, constructs and boundary conditions. Each is a Markdown file in the project. The relations between them are typed: one object supports, challenges, contradicts, extends, refines, defines or operationalizes another. When Scholardo reads a source and proposes candidates, each one must rest on a quote found word for word in that source; a candidate whose quote isn't there is dropped. What survives waits for you: nothing enters Synthesis until you accept it.

Three months later: the construct, the relation and the sentence it came from are a file you can open.

3. Words and numbers in one place. The transcripts, the survey export and the literature sit in the same project. When you ask the Conversation Agent for an analysis, its analysis skill has it write a Python or R script saved in the project, read the inputs without changing them, and write the results as files beside the script. The same skill tells it to say up front whether the analysis was planned or exploratory, and to report effect sizes with their confidence intervals rather than stop at "significant." What it does not do is code interviews. That is the job of MAXQDA or NVivo, and I come back to it below.

Three months later: the script, the input it read and the table it wrote are side by side, and the script runs again.

4. Everything points back. When the Conversation Agent writes a sentence that rests on a source, its instructions tell it to record an anchor there and then: the sentence as written, the source, and the page it opened. That is a declaration, not a verification, and the tool says so in as many words. But it means a claim can be followed back to the page someone says it came from. The agents can read the highlights and margin notes you made; none of them can make or change them. What you marked stays yours.

Three months later: the anchor is still attached to the sentence when someone asks where a number came from.

5. Cited from your library, checked before it leaves. In an academic project, the Conversation Agent is told to cite works from your own .bib by their citation keys, and the references render in the style you choose: APA 7 unless you pick another of more than 100 CSL styles. Drafts export to PDF, Word or LaTeX. Before an export, Scholardo checks the draft against your sources: a citation key that isn't in your .bib, a page number past the end of the cited PDF, a quotation or a figure with no source. It also notes cited works it has no record of you reading. By default the check reports and lets the export through. Set it to block, and each problem needs a written reason before the file goes out, recorded next to the file with the date. Looking cited works up online, to catch a retraction, happens only in a project where you allow it.

Three months later: any reason you gave for sending a draft anyway is on record beside the file, with its date.

That is the design. Here is what it looks like on the study from the opening.

The opening study, start to finish ​

What follows is an example of the workflow, not an account of a real study. The citation keys are placeholders.

You create one project and put everything in it: the forty transcripts, the survey export, the papers you have collected over three years. In settings, you point Scholardo's agents at models running in Ollama on your own machine.

You open a paper on job search after plant closures and read it. When you ask the Conversation Agent what the authors mean by "perceived employability," it can read the page you have on screen, so you don't paste anything. You ask Scholardo to extract knowledge from the paper. Three candidates come back, each with the sentence it rests on: a construct, perceived employability; a finding about how it relates to time to re-employment; and a boundary condition about age. You accept the first two. The third quotes a sentence about a different sample, so you reject it.

Your survey has four items meant to measure that construct. You record them as a variable that operationalizes perceived employability. Now the measurement decision is a file with a reason attached, not a choice you will have to reconstruct for a reviewer.

You ask for the analysis. The Conversation Agent writes employability-analysis.R in the project, reads the survey export without touching it, and writes a results table beside the script. It reports an effect size with its confidence interval and says the analysis was planned, because you said so when you asked.

The interviews you code in MAXQDA, with its local AI if you have it. Then you add the exported code summaries to the project, where the agents can read them.

Meanwhile the Background Agent has been watching the journals you follow. It picks out two new papers that bear on your question and proposes them; you keep one and dismiss the other.

You ask for a draft of the discussion section. It cites the papers by key, such as [@closure-panel-placeholder], and, as its instructions say, records an anchor for each number and quotation it uses. Before exporting to Word, you run the check. It reports two things: a page number past the end of one PDF, and a quotation with no source. You fix both, select the paragraph that runs long and have the Editing Agent tighten it, and export.

Nothing in that sequence needed the transcripts to leave your computer.

What it doesn't do ​

A tool that claims a gap should say plainly where it falls short itself.

• It does not code qualitative data. There is no code system and no segment coding. Keep MAXQDA, NVivo or ATLAS.ti for that, and bring the results into the project.
• Local models are slower and less capable than the largest hosted ones. The line MAXQDA prints applies here too: speed and quality depend on your hardware and the model you choose.
• Analyses need Python or R on your computer. The Conversation Agent writes and runs the script; it does not bring its own statistics package.
• Your IRB still decides. Keeping data on your machine makes the case simpler to make. It does not make the case for you.
• An anchor is a declaration, and the check is mechanical. It catches errors of reference, not errors of reasoning. It cannot tell you whether your argument is right.
• The license needs a connection now and then, as described under the first question.
• It runs on macOS and Windows. There is no Linux, web or mobile version.

Back to the five questions ​

The flagship AI scientists were built for problems like a drug target, a new material or a machine-learning experiment. Many social-science problems differ in two ways that matter for tools: the data is about people who were promised something, and the result is an argument as much as a number.

So the five questions are less about features than about where the work lives. Can the AI come to your study, instead of your study going to it? Can it keep your theory, your numbers and your sources connected, in files you can open later? Can it show you where every claim came from before your name goes on it?

Put those questions to any tool you are considering, including this one. If the answer to the first is "upload it," you already know what the rest will cost you.

Sources ​

Every quotation from a source in this article comes from the page linked where it first appears. The vendor descriptions were read on the vendors' own pages in October 2026.

• Bail, C. A. (2024). Can generative AI improve social science? Proceedings of the National Academy of Sciences, 121. doi:10.1073/pnas.2314021121
• Gao, J., & Wang, D. (2024). Quantifying the use and potential benefits of artificial intelligence in scientific research. Nature Human Behaviour, 8, 2281–2292. doi:10.1038/s41562-024-02020-5
• Mogull, S. A. (2017). Accuracy of cited "facts" in medical research articles: A review of study methodology and recalculation of quotation error rate. PLOS ONE, 12, e0184727. doi:10.1371/journal.pone.0184727
• Lehigh University, Office of the Vice Provost for Research. Guidance: Generative AI and Human Subjects Research.
• Frostburg State University. AI Use Policy.

Cover image by GuerrillaBuzz on Unsplash.