Will AI Replace Think Tanks? That Is the Wrong Question.

I analyzed 49 roles across the think tank and policy research sector.

The conclusion is not simply that “AI will replace think tanks.”

That question is too broad, and in many ways too abstract.

The more useful question is much closer to home:

If you work in a think tank, a policy research institute, a public affairs team, a foundation, a consulting firm, an international organization, or a government research unit, is your own role at risk of being replaced?

The answer depends on one thing above all:

Are you producing judgment, or are you processing information?

If your daily work consists mainly of reading reports, collecting materials, writing summaries, preparing slide decks, editing press releases, running basic data analysis, generating charts, drafting social media posts, or preparing early versions of grant proposals, then AI is no longer a future threat. It is already capable of doing a large share of your work.

But if your work involves defining the problem, assessing whether a policy is feasible, understanding stakeholders, building trust with decision-makers, making trade-offs in complex situations, and taking responsibility for the consequences of a recommendation, your risk is much lower.

This means the impact of AI on the think tank sector will not be evenly distributed.

Within the same industry, some roles will be compressed quickly. Others may become more important.

The 49 roles I analyzed fall into nine broad categories: research management, policy research, economic research, social research, international and security studies, data and quantitative analysis, communications and influence, fundraising and operations, and emerging AI-driven policy roles.

The overall estimated replacement exposure across the sector is around 28–38%.

On the surface, that number does not look especially high.

But averages can be misleading.

The estimated replacement exposure for a social media policy communications specialist is 75–85%.
For a policy data scientist, it is 65–75%.
For a grant proposal specialist, it is 60–75%.
For a public opinion polling analyst, it is 60–70%.
For a research publications editor, it is 55–70%.

At the same time, the risk for a think tank director is only 5–10%.
For a senior policy adviser, 15–20%.
For a foreign policy adviser, 10–20%.
For a military and security analyst, 15–25%.
For an AI ethics policy specialist, 15–25%.

So this is not really a story about whether AI will replace think tanks.

It is a story about which tasks are being stripped away, and which people will still be needed after that happens.

The Four Questions That Matter More Than Your Job Title

When assessing the risk to your own role, do not start with your job title.

Titles can be misleading.

A “policy analyst” may be relatively safe, or highly exposed.
A “data scientist” may sound technically advanced, but in some think tank settings that role can be surprisingly vulnerable.
A “communications officer” may be reduced to a smaller content-production function, or may become one of the institution’s most important guardians of political context.

The real test lies in four questions.

The first question is: Who defines the problem?

If someone else has already defined the question, and your role is simply to gather materials, summarize evidence, run a model, or draft the first version of a document, AI can enter that workflow very easily.

But if your role is to decide whether the question itself has been framed correctly, your value is much higher.

The second question is: Is your output standardized text?

Briefings, summaries, press releases, social media posts, grant applications, meeting notes, data explanations, policy scans — all of these are highly exposed to AI.

This does not mean humans will disappear from the process entirely. But it does mean fewer people will be needed to produce the same volume of output.

The third question is: How close are you to real stakeholders?

Do you interact directly with policymakers, foundation program officers, corporate public affairs leaders, community organizations, journalists, expert networks, or institutional partners?

If not, you may be operating mainly as an intermediate processor in the information chain.

AI is particularly good at compressing intermediate layers.

The fourth question is the most important: If the recommendation is wrong, who bears the consequences?

AI can generate recommendations.

But when a recommendation involves budgets, war, sanctions, education reform, health care allocation, immigration policy, or AI regulation, the real value does not lie in producing a sentence that sounds persuasive.

The real value lies in someone being willing to stand behind that sentence.

The closer you are to information, the more exposed you are.

The closer you are to responsibility, the safer you become.

High-Risk Roles: Vulnerable Not Because They Are Unimportant, but Because They Are Process-Driven

The most vulnerable roles are not necessarily low-status roles.

Many of them look professional, specialized, and institutionally important.

Their weakness is that their workflows are highly standardized.

This is one of the highest-risk roles among the 49 positions analyzed.

The reason is straightforward.

After a policy report is published, a communications specialist might once have spent an entire day adapting it into different formats: a LinkedIn post, an X thread, a newsletter summary, a media pitch, an event announcement, a donor briefing, and a plain-language version for the general public.

AI can now produce a dozen versions in minutes.

It can imitate an institution’s tone, adjust length, generate headlines, produce multilingual versions, and even suggest timing strategies for publication.

So if the value of a communications role is simply to translate research content into platform-specific language, that role will be heavily compressed.

But communications will not disappear.

The people who remain will look less like content producers and more like political-context editors.

Their real questions will be:

Can this sentence be published today?
Will this headline be taken out of context?
Could this wording be misread by a key stakeholder?
Will publishing this now expand the institution’s influence, or create a crisis?

AI can write content.

But AI does not bear reputational consequences.

Grant Proposal Specialist: 60–75%

Grant writing is another high-risk area.

A grant proposal often follows a predictable structure: problem statement, project objectives, methodology, budget explanation, evaluation framework, organizational capacity, and impact indicators.

These elements are well suited to AI-generated first drafts.

The most vulnerable version of this role is the person who waits for project staff to provide materials, then packages those materials into a polished application.

AI will absorb much of that packaging work.

But this does not mean fundraising expertise becomes less important.

In fact, the real judgment behind fundraising becomes more important.

Which foundation is genuinely aligned with this project?
What are the funder’s unstated priorities this year?
Should the proposal emphasize research impact, community impact, or policy implementation?
After a rejection, how should the narrative be adjusted?
When is it worth applying, and when would an application waste political capital?

AI can help write a proposal.

But AI does not know why a particular program officer may already be tired of a topic.

The people who remain will not simply be “proposal writers.” They will understand funding logic, relationship timing, institutional positioning, and strategic fit.

Policy Data Scientist: 65–75%

This may be the most counterintuitive high-risk role.

Many people assume that knowing Python, R, SQL, visualization tools, and machine learning automatically makes a person safe.

In think tanks, the reality is more complicated.

AI is rapidly reducing the premium on technical execution.

Writing code, cleaning data, running regressions, generating charts, interpreting model outputs, and translating results into policy language are all tasks that AI can already assist with — and will continue to improve at.

If the core value of a policy data scientist is simply “I can get the data to run,” the risk is high.

The safer part of the role lies elsewhere.

How should this policy question be modeled?
Can this variable really represent the underlying reality?
Is there systematic bias in the sample?
Does the causal inference strategy hold up?
Could policymakers misuse the result?

Data roles will not disappear.

But there will be fewer people whose main function is to run data.

The people who become more valuable will be those who can define the analytical question, design the identification strategy, and explain the policy implications responsibly.

Public Opinion Polling Analyst: 60–70%

Polling roles are also exposed.

Not because public opinion is unimportant, but because much of the polling workflow has become increasingly platform-based.

Questionnaire generation, sample management, data cleaning, cross-tabulation, sentiment analysis, and first-draft reporting can all be automated or accelerated.

But the hardest part of polling has never been calculating percentages.

The hard part is understanding why people answered the way they did.

Behind a number there may be fear, identity, culture, media exposure, historical memory — or simply a misleading result caused by poor survey design.

AI can tell you that support has declined.

It may not be able to tell you whether that decline reflects policy failure, communications failure, sample bias, or voters expressing frustration in a deliberately symbolic way.

Routine polling analysis will be compressed.

Survey methodology, cultural interpretation, and political-context judgment will remain valuable.

Research Publications Editor: 55–70%

Editorial roles are also at significant risk.

Grammar, formatting, citations, summaries, headlines, structural suggestions, and style consistency are all tasks AI can handle.

This is especially true in think tanks, where many publications follow standard formats: policy briefs, working papers, research reports, issue notes, and backgrounders.

The more standardized the publication, the easier it is for AI to absorb the editing process.

But strong editors still matter.

Their real work is not merely correcting sentences.

It is asking:

Is the evidence strong enough to support this conclusion?
Is the headline overstated?
Has the author avoided a crucial counterexample?
Who will attack this report after publication?
Is the institution willing to take the reputational risk of making this claim?

Routine editing will shrink.

Quality control, intellectual rigor, and reputational risk management will remain.

Medium-Risk Roles: The Jobs Remain, but the Headcount Shrinks

Many policy research roles will not be directly replaced by AI.

But they will be reorganized.

Work that once required five people may be done by two people working with AI.

Policy Analyst: 35–45%

The policy analyst role is one of the easiest to misunderstand.

Some people argue that policy analysis requires judgment, so it must be safe.

That is only partly true.

A large share of junior policy analysis consists of literature reviews, policy scans, case summaries, data summaries, and first drafts of briefing notes.

These are exactly the areas where AI is strong.

If your daily work sounds like this —

“Find out what other countries are doing.”
“Summarize the relevant regulations.”
“Write a two-page brief.”
“Summarize this report for senior leadership.”

— then AI has already entered the core of your job.

But there is another kind of policy analyst.

This person identifies policy windows.
They know who will support a proposal and who will resist it.
They understand whether a recommendation is realistic within an election cycle.
They can see why a department publicly supports an idea while privately delaying it.
They can turn a technically correct solution into a politically feasible one.

That kind of policy analyst is much harder to replace.

So two people with the same title — “Policy Analyst” — may face very different levels of risk.

The junior analytical executor is exposed.

The policy feasibility thinker is much safer.

Econometrician: 50–65%

The technical execution side of econometrics is being compressed by AI.

Regressions, robustness checks, instrumental variable tests, model interpretation, code generation, and chart production can all be accelerated.

But the most valuable part of econometrics is not running the model.

It is the identification strategy.

What can function as a natural experiment?
Is the instrumental variable truly exogenous?
Is the parallel trends assumption acceptable?
Could a statistically significant result simply be an artifact of institutional context?

AI can make the technical workflow faster.

But it should not decide for you whether the causal story is credible.

Future econometricians cannot rely only on saying, “I can run the model.”

They will need to explain why the model deserves to be believed.

Regulatory Impact Analyst: 45–55%

This kind of role will be significantly accelerated by AI.

AI can compare regulatory texts, identify compliance obligations, summarize industry effects, and generate cost-estimation frameworks.

But the difficulty of regulatory impact analysis lies in indirect consequences.

Will a rule cause firms to shift costs elsewhere?
Will it create a new opportunity for regulatory arbitrage?
Will it hurt small businesses more than large firms?
Will it make the regulatory goal appear successful on paper while failing in practice?

These judgments require industry knowledge and institutional intuition.

AI can perform the first layer of analysis.

Humans must be responsible for the second-order consequences.

Health Policy Analyst and Education Policy Researcher: 30–40%

The replacement exposure for these roles is not especially high, but their execution layers will still be compressed.

AI can process large bodies of literature, indicators, international comparisons, program evaluations, and implementation data.

But health and education are not purely data problems.

Education policy involves teachers, parents, students, local finance, examination systems, social mobility, and cultural expectations.

Health policy involves doctors, patients, insurers, medicines, ethics, public risk, and resource allocation.

AI can summarize the reform experiences of ten countries.

But AI cannot directly decide which reform is appropriate for a particular country, city, school system, hospital network, or community.

Policy does not operate only on paper.

Policy operates inside human institutions.

Low-Risk Roles: Protected by Trust, Judgment, and Responsibility

Low-risk roles share one feature.

They do not win by simply “knowing more.”

They win because their judgment is trusted.

Think Tank Director: 5–10%

A think tank director’s job is not to write reports.

It is to decide which issues the institution will prioritize, whom it will partner with, whose money it will accept, which positions it will reject, and how it will preserve credibility under political pressure.

AI can generate a daily briefing.

But AI cannot choose an institution’s direction.

It cannot replace a director in meetings with foundations, government officials, universities, media organizations, board members, and strategic partners.

The barriers around this role are reputation, relationships, strategic direction, and responsibility.

AI can assist the director.

It cannot replace the director.

Senior Policy Adviser: 15–20%

The value of a senior policy adviser does not lie in the ability to write a briefing note.

It lies in whether decision-makers are willing to listen at critical moments.

That is a difficult asset to quantify.

It comes from long-term interaction, a record of accurate judgment, sensitivity to political consequences, and the ability to provide clarity in chaotic situations.

AI can prepare materials.

But AI cannot make a minister, legislator, foundation CEO, or corporate public affairs leader trust it.

More precisely, these people may use AI.

But they will not let AI carry political responsibility alone.

Foreign Policy Adviser and Military/Security Analyst: 10–25%

International and security research is one of the areas where AI is least likely to fully replace human judgment.

Not because AI cannot collect information.

On the contrary, AI is highly useful for open-source intelligence gathering, multilingual news tracking, document summaries, and even assisted satellite image analysis.

But the core of geopolitical judgment is tacit knowledge.

Where is a leader’s real red line?
Is a country’s public statement a negotiating tactic or a policy shift?
Is a military movement a signal of deterrence, a miscalculation, or the beginning of escalation?
Does an ally’s silence mean support, hesitation, or internal division?

These questions do not come with clean datasets.

They do not have standard answers.

The real value of security analysis is the ability to make relatively reliable judgments under conditions of incomplete information.

AI Ethics Policy Specialist: 15–25%

There is an interesting irony here.

People who study AI ethics and AI governance are relatively hard for AI to replace.

That is because their work is not only about what AI can do.

It is about what AI should be allowed to do.

Which systems should AI be used in?
Who should be responsible when an AI system causes harm?
What data should not be used?
How much transparency is necessary?
How should societies balance safety, innovation, privacy, fairness, and accountability?

These questions do not have purely technical answers.

They require ethical reasoning, legal understanding, social negotiation, and political feasibility.

The more widespread AI becomes, the more important these roles become.

The New Roles AI Is Creating May Be Safer Than Many Traditional Ones

AI is not only eliminating or compressing jobs in the think tank and policy research sector.

It is also creating a new class of roles.

AI policy researcher.
Digital governance analyst.
Platform economy regulation specialist.
AI ethics policy specialist.
AI policy simulation engineer.
Policy research AI workflow lead.

These roles have one thing in common:

They are not merely about using AI.

They are about understanding how AI changes society, markets, government, labor, education, health care, security, and law.

Their relative safety comes from a simple reality:

When a technology becomes social infrastructure, the need to govern it increases.

The stronger AI becomes, the more important regulation, auditing, risk assessment, policy design, ethical judgment, and international coordination become.

So the right question is not only:

Will AI replace me?

It is also:

Can I move into the new problems AI is creating?

What This Means for Different People in the Sector

For junior policy analysts, the goal should not be to become faster at writing briefs.

AI will be faster.

The goal should be to ask better questions, identify real stakeholders, and understand why a policy that looks good on paper may fail in practice.

For people in data roles, it is not enough to display models and charts.

You need to show how you define the problem, why a variable can be trusted, and where the analysis might mislead policymakers.

For communications professionals, do not position yourself merely as a content producer.

Content production is becoming cheaper.

Your value lies in becoming a judge of context: what can be said, when it should be said, who it should be said to, and how far the message should go.

For grant writers and operations staff, do not remain only in documents and process.

AI will absorb much of that work.

You need to understand funding logic, organizational capacity, project fit, and relationship timing.

For senior researchers, do not rely only on experience or credentials.

AI will make ordinary research output cheaper.

You need a clear position, an original framework, a credible reputation, and a reason others continue to cite you and invite you into the room.

For institutional leaders, do not take comfort in the idea that the sector’s average replacement exposure is “only” 28–38%.

The average can make organizations complacent.

The real task is to redesign the workflow: let AI handle information processing, and move people toward problem definition, stakeholder coordination, quality control, and influence-building.

The Real Divide

The think tank sector will not disappear because of AI.

But many roles inside it will be repriced.

The most vulnerable people are not necessarily young people.

They are not necessarily non-technical people.

They are the people who have remained too long at the level of information processing.

Reading quickly is no longer enough.
Writing quickly is no longer enough.
Summarizing well is no longer enough.
Making charts is no longer enough.
Running models is no longer enough.

These abilities are being commoditized by AI.

What will become more valuable is something different:

Can you judge whether the question has been framed correctly?
Can you identify who will genuinely oppose a proposal?
Can you turn a technically correct recommendation into a politically feasible one?
Can you make trade-offs under uncertainty?
Can you make others trust your judgment?
Can you take responsibility for the consequences of your advice?

The central career question for this sector is not:

Will AI replace think tanks?

It is:

Is my role processing information, or is it carrying judgment?

If it is the former, AI has already arrived.

If it is the latter, AI may make you stronger.