Short answer: no, and the official data says the opposite in unusually strong terms. The US Bureau of Labor Statistics projects data scientists to grow 34.6 percent between 2025 and 2035, from 275,600 positions to 371,000, an increase of 95,400 jobs, with a 2025 median wage of $120,230. That makes it the third fastest growing occupation in the entire economy, behind only nurse practitioners and solar photovoltaic installers. Against total employment growth of 3.5 percent over the same decade, data science is growing roughly ten times faster than the labour market as a whole.
Why It Feels Like It Is Dying
The projection and the lived experience genuinely disagree, and dismissing the experience is not helpful. Four things are happening at once.
Entry-level hiring got much harder. The bootcamp-to-junior-analyst path that worked in the late 2010s no longer does. There are far more people with a portfolio of notebooks than there are junior openings, and that oversupply is real even while total demand grows.
The easy work got automated. Exploratory analysis, chart production, basic modelling and SQL writing are all dramatically faster now. Anyone whose value came from producing those artefacts has lost most of it, and that loss arrived quickly.
Titles fragmented. Work that would have been posted as “data scientist” in 2019 now appears as analytics engineer, machine learning engineer, decision scientist, research scientist or product analyst. Searching for one title makes the market look emptier than it is.
Disillusionment followed the hype. A large number of organisations hired data scientists without having the data infrastructure, the decision processes or the executive appetite to act on the output. Those roles were unsatisfying and many were cut, which produced a wave of people saying the field was over.
None of those four is evidence of an occupation in decline. Together they describe a field that professionalised: harder to enter, more specialised, and more demanding about what counts as value.
What the Growth Is Actually Made Of
A 34.6 percent projection is not driven by demand for the work that got automated. It is driven by three things that grew because of AI rather than despite it.
Model deployment needs data people. Every organisation putting a model into production needs someone who understands what the training data represents, whether the evaluation is measuring anything real, and why performance degrades over time. That is data science work and it barely existed at scale five years ago.
Evaluation is a data problem. The hardest question in applied AI right now is whether a system is actually working. Answering it requires building datasets, defining metrics, controlling for confounders and reasoning about sampling. That is exactly the skill set.
Decisions moved upstream. As basic analysis became cheap, organisations started asking harder questions: what would happen if we changed this, what is causing that, is this effect real. Those are causal and experimental questions, and they need someone who understands why correlation from an observational dataset is not an answer.
The adjacent occupations tell the same story. Computer and information research scientists are projected to grow 21.8 percent at a median wage of $140,300, and information security analysts 21.0 percent. The technical end of the labour market is not contracting.
The Work That Is Genuinely Disappearing
Being honest about this is more useful than reassurance, because it tells you what to stop doing.
| Task | Status | What replaced it |
|---|---|---|
| Writing SQL for a defined question | Largely automated | Deciding which question to ask |
| Exploratory plots and summary statistics | Automated | Interpreting what the pattern means |
| Baseline model fitting | Automated | Deciding whether a model is the right tool |
| Slide production for stakeholders | Automated | The conversation the slides support |
| Data cleaning for tidy datasets | Largely automated | Judging what counts as an outlier and why |
| Dashboard building | Largely self-service | Defining the metric the dashboard shows |
Total those rows and it is a large share of what a mid-level data scientist did in 2019. That is the honest scale of the change, and it explains the disconnect between a 34.6 percent projection and how the job market feels to someone whose week is made of those tasks.
What replaced it is harder, not easier
The work that remains is the work that was always difficult: framing an ambiguous business question as something answerable, knowing which data can support which claim, designing an experiment that could actually falsify a hypothesis, and telling a senior stakeholder that the answer they wanted is not supported.
That last one is the most underrated skill in the discipline and the least automatable. A model produces an answer regardless of whether the data justifies it. A person who understands the study design is the only safeguard against a confident, well-visualised, completely wrong conclusion reaching a decision-maker.
An evaluation problem that shows where the value moved
A company deploys an assistant that answers customer questions from its documentation. Someone asks whether it is working. That question sounds like a product question and is entirely a data science problem.
Working out what “working” means comes first, and it is not obvious. Accuracy against what reference? A sample of real questions, or the ones anyone thought to write down? Weighted how, given that 5 percent of question types account for half the volume? Is a confidently wrong answer worse than a refusal, and by how much, given that one erodes trust and the other merely annoys?
Then the measurement design. You need a labelled set drawn from actual traffic rather than invented examples, with the difficult cases represented in proportion to how often they occur. You need a scoring scheme defined before you look at any results, because a rubric written afterwards will agree with whatever you hoped. You need enough examples that a difference between two versions is distinguishable from noise, which is a power calculation.
Then the ongoing part. Performance drifts as the documentation changes and as the questions people ask change. Detecting that requires monitoring designed in advance, not a dashboard someone checks occasionally.
Every element of that is statistics and study design applied to a new object. It is also work that did not exist five years ago, it is being created faster than people can be hired into it, and it is a large part of why the projection for this occupation looks the way it does.
What to Know Before You Draw Conclusions
Occupational projections are not hiring forecasts. A 34.6 percent decade does not mean junior hiring is easy this quarter. Long-run demand and short-run market conditions are different things, and both can be true.
Exposure and employment are separate measures. BLS published AI exposure categories alongside the projections, sorting occupations into Low, Moderate, High and Very high relative exposure, and states directly that exposure “does not imply job loss, productivity gains, automation probability, or wage effects.” Data science shows substantial task overlap and the fastest technical growth in the economy simultaneously.
Wages have held up. A median of $120,230 is not the profile of a field being commoditised. Occupations in genuine decline show wage compression before they show employment decline.
Domain depth is the differentiator. A data scientist who understands clinical trials, credit risk or supply chain operations is difficult to substitute. A generalist who can fit models is much easier.
The infrastructure grew. More data, more systems, more instrumentation and more regulatory reporting all create work that did not exist. That growth is independent of how good models get.
A Decision Framework for Data Professionals
- Currently producing analyses to order. The most exposed position. Move toward experimentation, causal inference or evaluation work within the year, because those are the areas where demand is growing fastest.
- Mid-career generalist. Choose a domain and go deep. The combination of statistical competence and genuine industry knowledge is what organisations cannot hire easily.
- Interested in the AI side. Model evaluation and data quality for AI systems is currently the most acute shortage in the field, and it builds directly on existing skills rather than requiring a new career.
- Trying to enter. Do not compete on portfolio volume. One project on real, messy data with a documented decision that changed as a result beats ten clean notebooks. Entering through analytics engineering or a domain role and moving across is now the more reliable path than applying directly to junior data scientist openings.
The test that applies to all four: in your last three pieces of work, did anyone change what they were going to do because of what you found? If yes, you are doing the durable part of this job. If not, that gap matters more than any tooling question.
Common mistakes right now
- Concluding the field is dying from a difficult entry-level market rather than from the demand data.
- Building a portfolio of clean-dataset projects, which demonstrates none of the skills that are scarce.
- Treating automated analysis output as a finding rather than as a starting point requiring validation.
- Ignoring experimental design, which is the part of the discipline growing in value fastest.
- Searching for one job title in a market where the work is posted under six.
Building the Fluency the Growing Roles Assume
The data roles growing fastest all sit close to AI systems: evaluating them, supplying and validating their data, and deciding whether their output should be trusted for a given decision. That work needs statistical judgement and a real understanding of how these systems behave.
Concretely, it means knowing why a model can be confidently wrong, what an evaluation set has to contain to be meaningful, how performance degrades as the population shifts, and how to communicate uncertainty to someone who wants a single number. Learning that in a structured sequence is faster than assembling it from papers and product documentation, and it does not become obsolete when the tooling changes. Pairing it with a certificate alongside applied work makes the capability visible when a team is choosing who owns its evaluation. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.
FAQ
Is data science a dying field?
Why is it so hard to get an entry-level job then?
Will AI replace data scientists?
What should I learn to stay relevant?
Your Next Step
Take your last three pieces of work and write down, for each, what decision changed because of it. If you cannot answer for two of the three, that is the actual problem, and it has nothing to do with whether the field is dying. Then pick the single most ambiguous question anyone has asked you recently, the one where the answer depends on how you frame it, and work through what evidence would genuinely settle it. That exercise is the job that is growing 34.6 percent, and it is available to you inside your current role.