How to find and hire AI researchers

Search the work, not the job title

Two people carrying the title Research Scientist can be five years apart in capability, and the person who solved the problem you care about may be titled PhD student for another eight months. Titles in this field track institutional stage rather than ability, so a search built on them returns the wrong people confidently.

What a researcher has produced, on the other hand, carries a date and a name on it. That lets a search start from the problem. Describe the subject you are hiring on in your own words, in the terms you would use with a colleague, and match it against the titles and abstracts of recent papers. You get the people whose work sits closest to yours, including the ones no title would have surfaced.

What counts as strong, in numbers

Research quality usually gets proxied by one of three numbers. Two of them mislead in ways that bite hardest when you are hiring.

• Publication count rewards group size. A large lab produces many papers and its members appear on most of them. Counting publications ranks the group.
• Citation count rewards seniority. Two to four years pass before a paper accumulates citations, so the metric systematically favours researchers who published long ago. Since availability moves in the opposite direction, ranking on citations inverts your result set.
• First authorship at a top venue rewards the thing you want. Restrict to NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR and their peers, then count papers where your candidate is first author. The number tells you how many pieces of work this person led through a hard review process, and a doctoral student can score highly on it without being senior.

Where AI researchers are found

The literature is the map, and four other places put names on it earlier or with more context than a paper alone.

• The published record, searched by subject. Every paper names its authors and their institution on the day it was written. An index of the literature lets you search that record by problem rather than by keyword, and lists the co-authors who never became first author. This is where a search starts.
• Conference and workshop programmes. The workshop paper comes a year before the main-track paper, and its authors are a year further from their defence. Read the workshop programmes of NeurIPS, ICML and ICLR in your subject, and walk the poster session if you attend: a poster is a twenty-minute interview the candidate prepared for.
• Lab pages and advisors. A lab page lists its doctoral students with a start year, and the advisor knows who is finishing and who is open to industry. A short, specific note to an advisor gets a name more often than a month of cold outreach, and the advisor's reference comes with it.
• Recent submissions. New arXiv listings and OpenReview submissions in your subject show who is producing work this quarter. A subject alert on either turns the search into a weekly reading habit instead of a project.
• Interns returning to finish. Industry labs take doctoral students for a summer; the ones who go back to finish are listed on the lab's papers with a university affiliation. They have already seen what industry work is and decided to finish first, which is the profile that answers a message.

Why timing decides the outcome

Strong researchers are employed. They move from a doctorate into a postdoc, a lab or a company inside a window a few months wide, and during that window they receive offers from people who knew it was opening. Everyone else writes afterwards and hears that the decision has been made.

The end of a doctorate is public information held in inconvenient places: manuscript submission records, national thesis registries, ORCID education entries carrying an end date, lab pages listing an expected graduation, the acknowledgements of a final paper. Assembled and dated, it becomes the most useful field on a candidate profile.

Six to nine months before a defence is when a first message still gets read and answered. Before that, the person does not know their own date. After it, someone else has already asked. The academic calendar sets the rhythm: most defences fall between May and December, postdoc and faculty offers for the following autumn go out between January and April, and an industry offer that wants to compete has to be on the table in the same months.

Reaching them, and what to write

Researchers are less guarded than the average candidate. Being contacted about their work is a normal part of the job, and a message that engages with the work gets a reply at a rate that surprises people coming from other markets. What gets deleted is anything that could have been sent to a hundred other people: a role description, a salary band, a company paragraph.

Write two sentences about their paper that only apply to their paper. If you cannot, the search has not been done yet.

Finding the address is a real obstacle at volume. Most researchers print one on the first page of a paper, though not on every paper, not in a consistent format, and not always the address they still read.

How to evaluate an AI researcher

The record tells you what they did. Four hours tell you how they think, and whether your team can work with them.

• A research talk in front of your team. Forty minutes on their own work, followed by questions. Listen for what failed before the result worked, and for whether they can explain the idea to an engineer who has not read the paper. Someone who drove the work can; someone who was fourth author on it cannot.
• A paper discussion on your problem. Send one paper close to what you are trying to build, a week ahead. Ask what they would try first, what they expect to break, and how they would know within a month whether it is working. Taste shows here, and so does the habit of measuring before believing.
• A call with the advisor and a co-author. Ask who did what on the key paper, and what the person is like when an experiment fails for three weeks. Advisors are frank about this, because their name goes on the recommendation.
• Skip the coding screen. A researcher who writes slow code is normal and fixable. A researcher who cannot design an experiment is neither, and no algorithm puzzle detects the difference.

What an offer has to contain

A researcher chooses between your offer, a postdoc, a faculty position and a large lab. Money decides less than it does elsewhere, and four other things decide more.

• Compute, in numbers. How many accelerators, shared with how many people, with what queue. A vague answer reads as none. A lab with eight GPUs and a clear allocation beats a company with a thousand and a ticketing system.
• The publication policy, in writing. Whether they can publish, on what, with what review. Being unable to publish ends a research career, and researchers know it. Say what is allowed before they ask.
• The problem, and a peer. What they would work on for the first year, and who they would talk to about it every day. A single researcher in a company of engineers leaves within eighteen months; two rarely do.
• A date that beats the calendar. Postdoc and faculty offers arrive between January and April with short deadlines. An offer that needs three more weeks of approvals loses to a worse one that came signed.

On compensation, match the structure the person will compare against: base plus equity, with the equity explained in numbers and a relocation and visa plan written down. Doctoral students have been paid a stipend for five years and will not haggle well, so an honest first number is both fair and effective.

How long it takes

Count backwards from the window. A researcher who defends in September is reachable from January and decided by May, so a search that starts in June finds people who have already chosen. With the record and the dates in hand, the search itself takes two to four weeks; the wait for the right window is what takes months. A standing alert on your subject, read weekly, turns that wait into a list that is ready when the window opens.

Researchers are half of an AI team

A paper names the person who solved the problem. It seldom names the person who made the training run finish, the inference server hold, or the evaluation harness produce a number anyone trusted. Those people leave their record in repositories, and finding them takes a different method, set out in how to find and hire AI engineers. If you are still deciding which profile your problem calls for, start with AI researcher, applied AI engineer or AI product engineer.

Common questions

Where do you find AI researchers?

In the published record first, searched by subject rather than by title: papers name their authors and institutions on the day they are written. Then in conference and workshop programmes, on lab pages and through advisors, in recent arXiv and OpenReview submissions, and among doctoral students who interned in an industry lab and went back to finish.

How do you tell a strong AI researcher from a prolific one?

Count first-authored papers at venues with a hard review process: NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR and their peers. In machine learning the first author led the work, so that number says how many pieces of accepted research the person has driven. Total publication count rewards being in a large group where everyone appears on everything.

Why are citation counts misleading when hiring AI researchers?

Citations take two to four years to accumulate. Someone finishing a doctorate this year has published their strongest work too recently to be cited for it. Ranking candidates on citations therefore sorts the researchers who are about to be available below the researchers who have been unavailable for a decade.

When should you contact a researcher who is finishing a PhD?

Six to nine months before the defence. Earlier, the honest answer from them is that it is too soon to say. Later, the postdoc or the industry offer is already signed and no message changes it. Most defences fall between May and December, and competing academic offers go out between January and April.

How do you evaluate an AI researcher?

With a research talk in front of your team, a discussion of one paper close to your problem sent a week ahead, and a call with the advisor and a co-author about who did what and how the person handles three weeks of failed experiments. Coding screens and algorithm puzzles measure the wrong thing for this role.

What does it take to hire an AI researcher away from a lab?

Compute stated in numbers, a written publication policy, a defined first-year problem and at least one peer to work with, and an offer that arrives before postdoc and faculty deadlines. Compensation follows the same structure as the labs they compare against, base plus equity, with equity explained in numbers.

How do you find a researcher's email address?

Most researchers print a contact address on the first page of their papers or on a personal or lab page. Reading it from a document they published is reliable and lawful. Generating an address from a name and an institution format is a guess, and treating a guess as a fact is how you burn a first contact.

What should a first message to an AI researcher say?

Name the specific paper, say what in it matters to your problem, and ask a question the person is qualified to answer. Researchers are used to being contacted about their work and filter out anything that could have been sent to a hundred people. If you cannot write two specific sentences about their research, you are not ready to write to them.