■ CRITICAL RISK ■ Science & Research
The tasks that justified hiring research assistants — transcription, literature screening, qualitative coding, data cleaning — are precisely what LLMs now do in hours instead of semesters. The role survives mainly where research needs a human face: recruitment, interviews, and fieldwork.
“AI codes qualitative data in hours. Your highlighting took weeks.”
Our AI replacement risk score — how we score jobs
An RA's workload is the unglamorous substrate of social science: transcribing interviews, screening hundreds of abstracts for a literature review, coding open-ended survey responses into themes, cleaning messy datasets, running participants through protocols, managing IRB paperwork, and formatting citations at 2 a.m. Principal investigators hire RAs because this work is essential, tedious, and — historically — impossible to automate. That last property just expired.
Language models transcribe interviews in minutes, screen abstracts against inclusion criteria, apply codebooks to qualitative data with documented reliability that increasingly rivals human coders, clean and reshape datasets conversationally, and draft literature summaries. Statistical software with AI assistance lets a PI run analyses that once required a quantitatively trained assistant. The honest academic conversation has shifted from 'can AI do qualitative coding' to 'how do we report that it did' — methods sections are already adapting. For grant-funded labs under perpetual budget pressure, an assistant-shaped subscription is irresistible arithmetic.
What resists is embodied and interpersonal research. Recruiting participants from hard-to-reach communities, building rapport in a sensitive interview about addiction or grief, running lab protocols with human subjects, observing behavior in the field, and handling the ethical judgment calls IRBs exist for — these need humans, often specifically trained, culturally fluent ones. There's also a pipeline problem with real stakes: RA positions are how students become researchers, and automating the apprenticeship hollows out training just as it did in law and accounting. Our 78 reflects desk-based RA work collapsing fast, while fieldwork and human-subjects roles hold the line.
Automatability: our editorial assessment of current and near-term AI capability
This is happening mid-grant-cycle: labs are already substituting AI for transcription and first-pass coding, and tightening research budgets accelerate the substitution. Desk-based RA hours contract sharply through the late 2020s. Positions centered on participant interaction, fieldwork, and lab management persist, and the academic labor system's inertia — grants written years ago budgeting for RAs — softens the transition without stopping it.
The desk-based portion is, quickly — transcription, literature screening, qualitative coding, and data cleaning were the core RA workload, and LLMs now do them in a fraction of the time at a fraction of the cost. Roles involving participants directly — interviews, recruitment, lab sessions, fieldwork — remain human. Fewer total positions, reshaped around the interpersonal work.
Increasingly, yes — applied against a well-specified codebook, LLM coding shows agreement with human coders that many teams find acceptable, and methods literature is actively working out validation standards. Where it stumbles: developing codes inductively from unfamiliar contexts, cultural nuance, and knowing when the codebook itself is wrong. Best practice is becoming AI-first-pass with human validation — which needs one RA, not four.
Extract the durable skills: research design, statistics and causal inference, interviewing and fieldwork, and IRB/ethics fluency. Volunteer for anything involving live participants. Learn to use AI tools rigorously — validating machine coding is itself becoming a methods skill worth listing. The students hurt most will be those whose RA experience was purely transcription and spreadsheet janitorial work; make sure yours isn't.
That's the quiet damage. RA positions were the apprenticeship of social science — how students earned mentorship, authorship, and grad-school letters. As AI absorbs the tasks that justified those hires, labs fund fewer assistants, and the training ladder loses rungs. Programs are starting to grapple with it, but if you're aiming at a research career, seek roles with genuine human-subjects contact and methodological responsibility early.