Why the judgment calls stay with people
Will AI replace political scientists? The short answer sits in the hero above, and the reason sits in the shape of the work. Political scientists do not just produce text. They decide which evidence counts, defend that choice in front of officials and reviewers, and carry the blame when a forecast misses. Language models can draft a literature summary in minutes. They cannot sit in a hearing and answer for the method.
Two parts of the job hold the line. The first is advising elected officials, agencies and clients on the likely political effects of a proposal, where the useful answer depends on who holds power this month and what they will tolerate. The second is interpreting conflicting findings about how institutions behave in practice, which means weighing data quality, context and the limits of a sample rather than averaging what has been published. Our Can AI do it? figure for this work reads 45 out of 100, and you can see how that figure is built on the coverage method page.
The erosion is real, though, and it lands on the production side: background research, first drafts, data cleaning, and the long tables of sourced facts that used to fill a junior analyst’s week. That is where fewer hours, and often fewer entry-level posts, show up first. The Bureau of Labor Statistics counts roughly 5,540 political scientists in the United States with a median wage of $142,080, and projects a 2% decline in employment over 2025 to 2035 (BLS, 2025). A small, senior, slow-growing occupation does not need much task erosion to feel it in hiring.
What software drafts, what it assists, what people keep
Work the tools can mostly handle alone: 23% of task time. This is the repeatable end of research production, including pulling and summarizing published studies and documents, and coding or tidying structured survey and voting data before analysis.
Work where a model speeds up a person without owning it: 57% of task time. Analyzing survey and election data is faster with a model that writes and checks the code, and so is drafting reports and articles, as long as a political scientist decides what the numbers mean and signs the claim.
Work that still needs a person: 20% of task time. Advising officials and clients on policy and political strategy sits here, along with developing and defending theories about institutions, parties and behavior, and the interviewing and field contact that produces evidence nobody has published yet.
There is no physical side to speak of. Our robotics tier for this job is “None needed,” which means no hardware has to exist before the software part advances. That is unusual among the jobs we score, and it is one reason the timing range is not pushed far out.
What the evidence does and does not show
Our Is it better than a person? grade here is D, and a D grade means one thing only: nobody has tested AI against qualified political scientists on this job’s real tasks in a way we can score. So we publish no parity number. Guessing one would be worse than leaving it blank.
What would settle it is not hard to describe. A blind comparison where political scientists and a model each produce a policy memo or a forecast on the same question, judged by experienced reviewers who do not know which is which. A tracked record of published political forecasts from models and from people, scored against outcomes over several election cycles. A test of data work, where both are given the same messy survey file and the same analysis question. Until something like that exists, the honest position is uncertainty, which is how our quality parity method treats it. You can see the whole approach on the methodology page.
Good to know: a cheap tool is not the same as a proven one, and the cost figures above compare running software with paying a person, not the quality of the output.
When this could shift
Most likely between 2037 and 2047 (8 in 10 of our scenarios). What that window measures, and how we build it, is set out on the replacement year method page.
Two things could pull it earlier. The cost gap is wide: running the AI side of this work falls between $90 and $9,380 a year in our estimates, against $37,590 to $88,030 for the human equivalent, so budget-squeezed research teams have a strong reason to try. And nothing physical is in the way, so adoption moves at software speed rather than hardware speed.
Two things hold it back. Accountability is the big one. Agencies, committees and clients want a named person who can be questioned about a method, and a model cannot hold that role. The second is evidence: with no measured parity, buyers of political analysis have little basis for handing over the judgment calls, and the most valuable work is exactly the work that has not been tested.
How to stay needed
Lean into the tasks that stay with people. Advising officials and clients directly, where the value is reading the room as much as reading the data. Building and defending original arguments about how institutions and voters behave, rather than restating a consensus. And primary collection, including interviews, fieldwork and survey design that produces data a model has never seen.
Two skills carry most of the weight. First, method auditing: being the person who can find the flaw in a model-generated analysis, from a bad weighting choice to a quietly wrong coding scheme. Second, plain explanation under pressure, in a briefing or a hearing, where a long written answer is useless.
If you are weighing nearby paths, the closest work by task is Survey Researchers, Sociologists and Economists. You can see all of them next to each other in the social scientists job family, and much of this employment sits in the government sector. To test one option against another, put them side by side in the job comparison tool, or see what the AI models say about their own reach into this kind of research.