Artificial intelligence in the United States: research and engineering internships

Research, engineering, data: three ways of doing AI

Under the "AI" label, American postings cover very different realities. A research internship aims to explore an open question: improving a learning method, measuring the robustness of a model, studying the biases of a dataset. The deliverable is a scientific result, sometimes a paper or published code. An engineering internship is more about building, optimising or deploying a software component: a data pipeline, an inference service, an evaluation tool. A data science internship applies statistical methods to a specific business problem.

This distinction decides everything else: the profile expected, the interviews, the calendar and even the type of visa you can consider. Read the posting like a specification. If it cites publications, a laboratory director and a thesis topic, it is research; if it talks about distributed systems, latency and production deployment, it is engineering.

The main kinds of host

Universities host the majority of research projects. Professors recruit students into their teams, often thanks to public funding. Federal science agencies also offer summer programmes. The National Institute of Standards and Technology, for example, describes a Summer Undergraduate Research Fellowship open to undergraduates in science and technology laboratories; its page specifies that applicants must be American citizens or permanent residents, enrolled at an American institution.

Technology companies form the second family. NVIDIA describes on its university pages several programmes for students: internships of at least twelve weeks, introductory pathways for students early in their studies, programmes for doctoral students. Other private players keep comparable pages. Non-profit research institutes, such as the Allen Institute for AI in Seattle, publish research internship offers linked to subjects like AI for science or robotics, with conditions set out in each posting.

Finally, start-ups offer less formalised internships, often for profiles able to deliver quickly. They give great autonomy, but supervision is more variable.

Read the criteria before investing time

Many research programmes funded by the federal government reserve their places for citizens or permanent residents. This is the case for several summer experiences supported by the National Science Foundation under the REU scheme, whose rules specify eligibility. An international student who applies to these programmes counting on academic excellence will run into the rule, with no appeal possible.

Before preparing an application, systematically note four pieces of information:

• the nationality or status required, for the programme and for the host laboratory;
• the level of study expected: undergraduate, master's, doctorate;
• whether the location is mandatory or remote work is possible;
• the pay, any housing and the planned duration.

A good habit is to keep a personal table of your leads, but remember that there is no single catalogue here: each employer publishes its own rules and changes them over the years.

Building a credible AI profile

Recruiters in a laboratory look for solid foundations rather than a list of fashionable tools. Mathematics first: linear algebra, probability, statistics and optimisation. Then programming, with good command of Python and at least one machine learning library. Finally an experimental culture: knowing how to define a baseline, control sources of error and reproduce a published result.

To prove it, favour a few clear achievements:

• a clean, documented code repository, with a file explaining how to re-run the experiments;
• the reproduction of a well-known paper, with an honest account of the differences in results;
• an end-of-semester project that tackles a specific problem with open data;
• a contribution, however modest, to an open-source project.

Five average projects read worse than two in-depth ones. What matters is to explain why you made a given choice and what you would do differently.

Contacting a research team

For a university laboratory, the application often goes through a short email to the principal investigator, or through a form when the programme is structured. Read at least two recent papers from the team before writing. Your message takes only a few lines: who you are, which of the team's work interested you, what skill you can bring, which dates you are aiming for, and a one-page CV attached.

Avoid phrases like "I am passionate about AI" that say nothing. It is better to write: "your approach to evaluating language models on reasoning tasks raises a question I met in my semester project". Professors receive dozens of emails a week and answer those that show careful reading.

Do not expect an immediate reply: follow up once after two weeks, politely, then move on to other leads.

Interviews: coding, reasoning, discussing

In industry, an AI internship interview often includes a programming exercise, a discussion of machine learning concepts and a conversation about a past project. Expect questions like these: why does a model overfit, how do you choose a metric for an imbalanced dataset, how do you detect information leakage between training and testing.

In research, the interview is more like a scientific conversation. You will be asked to explain a paper, criticise an experimental protocol or propose an experiment to test a hypothesis. An answer that acknowledges its limits goes down better than unfounded assurance.

Practise reasoning aloud in English. Describing a mathematical idea out loud is a different exercise from reading it.

Data, ethics and ownership of results

A serious team will explain how it handles data: provenance, licence, anonymisation, access rights. Ask the question, especially if the project touches personal, medical or financial information. Also ask who supervises your work day to day, how results are reviewed and what happens to the code you write.

Intellectual property varies. In a company, the intern's work generally belongs to the employer and may remain confidential. In a university laboratory, publication is a goal, with authorship rules set by the principal investigator. Clarify these points in writing before you begin rather than discovering them at the end.

Some advanced technologies may be subject to export controls; the employer will then specify what a foreign candidate can or cannot handle.

Stay and status for an international candidate

As with any professional activity, a tourist trip does not authorise you to work. The best-known route for an intern who is enrolled or recently graduated abroad is the Intern programme of the J-1 exchange category, administered by the Department of State under the name BridgeUSA. It requires an approved sponsor, a precise training plan and a DS-2019 form. The host organisation must commit to the educational content of the internship.

A student already enrolled at an American university falls under internship rules built into their study visa. A doctoral student invited by a laboratory may have a different exchange status. In every case, raise the question of the administrative framework from the first exchange with the team, because the sponsor, the fees and the delays determine the start date.

Keep in mind that some teams simply cannot host foreign interns: this comes down to their funding or their status, not to your profile.

An example of putting a project together

By way of illustration, with hypothetical durations: Inès, a computer science master's student abroad, is aiming for a four-month research internship on model evaluation. In September she draws up a list of twelve teams and rules out four programmes reserved for American citizens. Eight remain. She reads two papers per team, which takes her three weeks.

In October she sends eight personalised emails. Three teams reply, and one offers an interview. At the same time she contacts two J-1 sponsors to compare fees and delays. In December an offer arrives for June: she still has time to assemble the visa file and look for housing.

This example follows a cautious calendar: an American internship project is prepared over eight to ten months, sometimes more.

Mistakes to avoid

• applying to programmes whose nationality rules you have not read;
• confusing a research experience with an engineering position, and sending the same file to both;
• writing a generic email to twenty laboratories;
• presenting results without specifying the data, the metrics and the limits;
• discovering the sponsor and visa requirements late;
• accepting a project without knowing the supervisor or the publication rules.

For each point, the fix costs time but little money. A few carefully prepared applications are better than a rain of identical messages.