Artificial intelligence researcher (AI Researcher)

An artificial intelligence researcher, or AI Researcher, formulates scientific questions and designs or evaluates new methods and models. Their work gives an organisation insight into what works, under which conditions and with what limitations. They may work in fundamental or applied research, depending on the question being studied and the project’s objectives.

AI research is not limited to language models. Topics may also involve modelling physical systems or processing images and experimental signals. The required specialism therefore needs to be specified when defining the role. Experience with one type of data is not, on its own, enough to establish suitability for a different scientific problem.

The researcher writes code, experiments and analyses their results. Their outputs may include code, a prototype, a technical report or a publication. The expected format depends on the organisation and the scope for sharing the work. A research prototype does not, in itself, establish who will maintain the system used by business teams.

Compare responsibilities before choosing a job title

When comparing roles, clarify the following points for each title:

• AI Researcher: focus the role on the scientific questions to explore, the methods to develop and the evidence to produce.
• Data Scientist: when comparing roles, specify the respective emphasis on data analysis and the design of new methods.
• AI Engineer: make clear how the work is divided between experimentation, model integration and building systems. These activities may overlap with those of the researcher.

Reporting lines and scientific authority need to be clarified with the employer. Conducting research, guiding other researchers’ work and managing a project’s operations are different responsibilities. If they are combined in one role, this must be explicit.

Hiring an AI researcher commits the business to work whose outcome remains uncertain. Research may generate knowledge without achieving the result initially hoped for. For the business leader, the challenge is to specify which decision this knowledge should inform: pursuing a line of enquiry, modifying a method or recognising a limitation of the models being studied.

Connect the scientific question to the business need

Describe the problem before choosing a job title. If you need to build an application using methods that have already been identified, also consider the AI Engineer role. If uncertainty about the methods is holding up the project, articulate that uncertainty and what is needed to investigate it. This distinction helps allocate responsibilities without assuming a rigid boundary between research and engineering.

Avoid making model improvement the sole objective without specifying what will be measured. Link success criteria to the available data and the intended conditions of use. In a language model project, evaluation may focus on factual accuracy, robustness or faithfulness to sources. The choice of criteria must reflect the subject being studied.

Prepare for the decisions that will follow the experiments

Hypothetical example: a company wants to adapt a language model to specialised documentation. Before starting the experiments, it asks the researcher to define how to assess whether answers are faithful to the documents. It also specifies who will review the results and who will decide whether to proceed with integrating the model.

To avoid hiring for a poorly defined role, distinguish the expected scientific deliverables from production responsibilities. Plan for computing resources and arrangements for retaining data and code. Finally, agree on how to discuss results that contradict the initial hypothesis. This will allow you to assess the work carried out without turning the uncertainty inherent in research into a promise of performance.

• Formulate the scientific questions to investigate, based on the problem posed by the organisation.
• Review the current state of knowledge to identify relevant lines of research.
• Design new methods or adapt models to the subject being studied.
• Develop the code and prototypes needed for experiments.
• Define the technical and experimental criteria for evaluating models.
• Analyse results and make the remaining uncertainties explicit.
• Document the approach and scientific choices so that they can be examined.
• Communicate results through reports, publications or presentations, depending on the scope for sharing the work.

A doctorate is a common route into the research roles described in the cited sources. The fields of study mentioned include applied mathematics, statistics, computer science and data science. Some applied research roles also accept substantial research experience. A doctorate is therefore neither a universal requirement nor sufficient evidence of suitability for the role.

Theoretical and practical learning

A dissertation, thesis or research project draws on theoretical knowledge, modelling choices and the interpretation of results. Learning relevant to the role includes turning a scientific question into an experiment, implementing ideas in code or prototypes, and presenting evidence and limitations in reports or publications. Relevance to the company’s field of research helps distinguish knowledge and skills that can be applied directly from areas requiring further study.

Scientific English also matters when the role involves reading, writing or presenting work in that language. Accessible work does not necessarily reflect the full breadth of a person’s experience: sharing it depends, in particular, on confidentiality restrictions.

Distinguish experience from autonomy

Experience is characterised by the responsibilities held: leading a project, defining experimental choices or helping to guide collaborative work. These responsibilities involve different degrees of autonomy. Experience that includes supervising doctoral students or working on collaborative projects may be relevant to a role with these responsibilities. The person’s individual contribution and the decisions they made then clarify the autonomy they have developed.

The foundations of GetPro’s assessment approach

GetPro structures assessment around a grid of prioritised criteria. Each criterion is paired with a way of assessing it, distinguishing elements of the candidate’s experience from capabilities to explore further in an interview. Open questions and concrete examples allow the essential criteria to be examined. The grid guides the discussion towards points that still need clarification.

Reference checks help corroborate the responsibilities held and explore points raised during the interview in greater depth. These discussions place skills in the context of working with others and seek specific examples.

Suggested applications for the AI Researcher role

The questions and case below are suggestions to adapt to the scientific specialism and the responsibilities of the role. They do not constitute a documented GetPro protocol specifically for AI Researchers.

1. Define the criteria and assessors

Create a grid linking each responsibility to an observable capability. For designing methods, focus on theoretical reasoning and modelling choices. For conducting experiments, examine the protocol, code and interpretation of results.

Specify criteria relevant to the subject. In a language model project, you might, for example, examine factual accuracy or faithfulness to sources. Do not turn these examples into mandatory criteria for every specialism.

If your company lacks scientific expertise, involve a specialist in the field in the assessment. Entrust them with examining technical choices and retain responsibility for defining the expectations of the role.

2. Examine a piece of work and the candidate’s personal contribution

Ask the candidate to present work they are able to share. Have them explain the initial question, the methods considered, the decisions made and the conclusions reached.

Distinguish what they personally designed, implemented or analysed from the work of other contributors. Where a publication describes the authors’ contributions, use this information to explore the subject further. Also consider any available protocols, code and reports.

A precise explanation of choices and their limitations is a positive sign. Unclear attribution of contributions calls for further questions. Do not reach a conclusion based solely on the number of publications.

3. Discuss a case relevant to the role

Present a research question, the accessible data and the computing constraints. Ask for a reasoned approach rather than a completed research result.

Hypothetical example: for a role focused on language models, suggest examining a method intended to improve the faithfulness of answers to documentation. Ask which criteria would allow this improvement to be assessed.

Invite the candidate to explain which experiments they would choose and what conclusions those experiments would actually support. Ask them to discuss the limitations of the data and the remaining uncertainties.

A protocol consistent with the question is a positive sign. A promise of improvement without evaluation criteria warrants further exploration. Also observe how the candidate responds to a scientific objection.

4. Assess cooperation and supervision

Ask how the candidate has discussed a choice with specialists from other disciplines. Invite them to explain a technical result to someone unfamiliar with their specialism.

If the role includes supervision, examine a situation in which they supported the work of a doctoral student or researcher. Ask them to specify which decisions were left to that person and what help they provided.

For a project intended to feed into a product, discuss interaction with engineering. Ask what the candidate would hand over to make their results understandable and open to examination.

5. Corroborate responsibilities and reach a conclusion

With the candidate’s consent, prepare a discussion with a referee focused on the responsibilities the candidate has held. Seek to clarify their scientific autonomy, their contribution to collaborative work and the way they discuss the limitations of results.

Then compare the evidence gathered against the criteria defined at the outset. Separate observed capabilities from points that remain uncertain. If an essential skill is still difficult to assess, arrange a focused discussion with the appropriate assessor before reaching a conclusion.