How AI Is Reshaping Recruitment — and What Candidates Should Do

Ask a recruiter what has changed in the last three years and you will hear the same list: sourcing is faster, scheduling is automatic, screening is partly scored by software, and the volume of applications has risen sharply because candidates can now generate a tailored application in ninety seconds.

That last point is the one most candidates underestimate. AI has made applying cheap for everyone, which means the market is noisier and employers are leaning harder on filtering. Understanding both sides of that loop is now part of running a competent job search.

Where AI is actually used

Ignore the marketing. In practice, deployment clusters in five places.

What is largely not happening, despite the headlines, is fully autonomous rejection at the final stage. In most regulated markets a human must be meaningfully involved in an adverse decision, and most employers keep a person in the loop for liability reasons alone.

How automated screening works

A modern screening stack does three things with your application.

1. Structures it. Your CV becomes fields: titles, employers, dates, skills, education, seniority estimate.
2. Embeds and compares it. Rather than matching exact strings, the system represents your experience and the requisition as vectors and measures similarity. This is why "led a team of six" can now match "line management" without sharing a word.
3. Applies rules. Hard filters — work authorisation, required licence, minimum experience — are usually explicit rules rather than model outputs, and they are typically driven by your answers in the application form.

AI-assisted and asynchronous interviews

You are increasingly likely to meet one of three formats.

• One-way video. You record answers to fixed questions. A human reviews them, usually with a transcript and sometimes an automated summary.
• AI note-taking in a live interview. The most common use today: a bot joins, transcribes, and produces structured notes against a scorecard. You should be told, and you may ask for it to be off.
• Automated scoring of answers. Content-based scoring of written or spoken responses. Scoring based on facial expression or vocal tone has been withdrawn by most serious vendors after regulatory pressure and poor validation evidence.

Handling a one-way interview well

1. Read all the questions first if the tool allows it, and note which stories you will use.
2. Use the practice attempt. Nearly everyone's first take is stiff.
3. Answer in structure — situation, action, result — because the reviewer is skimming a transcript, not savouring your delivery.
4. Look at the camera lens, not at your own image.
5. Keep answers inside the time limit; being cut off mid-sentence costs you the result of your own story.
6. Neutral background, front lighting, clean audio. The format is unforgiving of poor setup.

Bias, audits and what goes wrong

Models learn from historical hiring data, and historical hiring data encodes historical preferences. The well-documented failures — a CV screener that learned to downgrade women's CVs, tools that correlated scores with background objects in a video — were not exotic. They were ordinary consequences of optimising against biased outcomes.

Responsible employers now do three things: they audit tools for disparate impact across protected groups, they restrict models to job-relevant content, and they keep humans accountable for decisions. New York City requires annual independent bias audits for automated employment decision tools, and the EU AI Act classifies recruitment and worker-management systems as high-risk, with obligations around transparency, data governance and human oversight.

A useful question in any interview process: "Is any part of this assessment automatically scored, and is there a human review before a decision?" A good employer answers plainly. Evasion tells you what you need to know.

Your rights when a machine decides

Rights vary by jurisdiction, but in the EU and UK the core protections under data protection law are reasonably consistent. This is general information rather than legal advice — for a specific dispute, take advice locally.

• Transparency. You should be told when automated processing is used in a way that significantly affects you, and given meaningful information about the logic involved.
• Human involvement. You generally have the right not to be subject to a decision based solely on automated processing where it has legal or similarly significant effects — and to ask for human review.
• Access. You can request the personal data an employer holds about you, including assessment records and scores, in most cases within a month.
• Correction and erasure. Inaccurate data can be corrected; you can ask for deletion once a process has concluded, subject to the employer's retention obligations.
• Objection. You can object to certain processing and ask for the basis on which it is justified.

In practice, a polite written request — "could you confirm whether automated scoring was used in my assessment, and ask for a human review of the outcome?" — is answered more often than people expect.

Using AI in your own search

Used well, these tools remove drudgery. Used badly, they produce applications that sound exactly like the other four hundred.

Good uses:

• Extracting requirements from an advert and comparing them against your CV to find genuine gaps.
• Turning a rambling paragraph about a project into a tight achievement bullet — with your numbers.
• Generating likely interview questions for a specific role, then practising answers out loud.
• Drafting the awkward messages: a follow-up, a networking note, a resignation letter.
• Researching a company's market and preparing questions to ask.
• Summarising a long job description into the five things that actually matter.

Poor uses:

• Generating a cover letter from scratch and sending it unedited. Recruiters recognise the register immediately.
• Inventing experience. It surfaces in the first technical question and ends the process permanently.
• Mass-applying to hundreds of roles. Response rates fall, and some platforms flag the pattern.
• Using a live assistant during an interview. Assume it is visible — eye movement, latency and suspiciously polished phrasing all give it away.

Where AI use damages you

There is a simple test. If the AI produced content — claims, experience, opinions that are not yours — you are misrepresenting yourself. If it helped with form — structure, concision, grammar, a first draft you then rewrote — that is ordinary use of a tool, in the same category as a spell checker or a friend who reads your CV.

The practical risk is homogeneity. When everyone uses the same assistant with the same prompt, applications converge on the same cadence and the same vocabulary, and the reader stops distinguishing between them. Specificity is the antidote: the particular number, the particular constraint, the particular thing that went wrong. No model can invent those for you, which is precisely why they now carry so much weight.

What stays human

After all the automation, hiring decisions still come down to a small number of judgements no system makes reliably: whether someone will be trusted by the team, whether they will still be motivated in eighteen months, whether they can handle a situation that was never in the job description.

Those judgements are made by people, in conversations. Everything AI has changed sits before that point — it decides who gets into the room, and how fast. Your job is to make the machine-readable part of your application unambiguous, and then to be unmistakably yourself in the part that matters.