What Counts as AI Misconduct in a Graduate Scheme

In UK early-careers hiring, the rule on AI is no longer one rule. As of October 2026 it splits by stage: most graduate employers now accept some AI help on the written application, and the same employers prohibit it in online assessments, interviews, and assessment centres. That boundary — not the tool you used — is what decides whether your AI use is a normal part of applying or a misconduct finding against your name.

This article reads the Institute of Student Employers' 2026 survey data from the candidate side: what employers reported catching, which stage they reported catching it in, and what they changed about the funnel in response. It covers UK graduate schemes and apprenticeships, where application volume per vacancy is high and the selection process is standardised across hundreds of candidates.

What the 2026 survey reported

The Institute of Student Employers (ISE) publishes an annual student recruitment survey. The ISE Student Recruitment Report 2026 draws on 155 employers, more than 1.7 million applications, and 31,577 hires and work experience placements across the 2025/26 cycle. The full report is members-only, so the figures below come from press coverage published on 14 and 15 October 2026 rather than from the report itself. Treat them as reported figures, not as something independently verified.

The headline number is a trend, not a single year. 92% of employers reported suspected or confirmed candidate misconduct in 2026, against 85% the previous year and 78% the year before (HRreview, 14 October 2026).

The breakdown by stage matters more than the total. PA Media's coverage of the same report puts undeclared AI use in interviews as the most commonly reported form of misconduct, at 66% of respondents reporting at least one case, followed by AI used to complete online assessments at 47%, and AI used to write CVs or cover letters at 10% (PA via AOL, 14 October 2026).

Read that bottom figure alongside the policy figures and it stops looking like candidates have gone quiet on CVs. 64% of employers now accept AI use somewhere in their process, up from 46% the previous year, typically for CVs, cover letters, and written application questions. 82% publish guidance on what is and is not acceptable, up from 55% in 2025 (HRreview, above). The 10% is low in large part because writing a CV with AI is no longer misconduct at most of these employers. It has been reclassified rather than eliminated.

Two important caveats. These are employer reports of suspected or confirmed misconduct, which is a wider category than proven cases, and they are self-reported by the employer, not measured. And one employer reporting "at least one case" says nothing about how many candidates in their funnel were involved.

Where the line sits, stage by stage

Written application, CV, and cover letter

This is where acceptance has moved fastest. With 64% of employers accepting AI somewhere and the acceptance concentrated here, the realistic default for a UK graduate scheme in the 2026/27 cycle is that AI assistance on written application material is allowed unless the employer says otherwise.

"Unless the employer says otherwise" now has a much better chance of being written down somewhere. With 82% publishing guidance, the application portal, the FAQ, or the role page is the first place to look, and the answer you find there overrides any general advice, including this article's.

Online assessments

Prohibited at these employers, and the stage with the most instrumentation behind the rule. Assessment platforms score behavioural signals — paste events, tab focus changes, timing patterns, code similarity — and hand a confidence score to the hiring team rather than a verdict. What that score does to your application is entirely at the employer's discretion, which is covered in detail in what an integrity flag on a coding assessment actually does.

Video and live interviews

The most-reported category in the survey, at 66%. It is also the stage where detection is least systematic and most human: an interviewer forming an impression, rather than a platform computing a score.

Chris Morrow, who runs the Manchester recruitment agency Digitalent, described to PA the pattern he says he looks for — candidates who sound "like they are reading off an Autocue," noticeable pauses before answers, filler phrases such as "that's a good question," and stiff, overly formal phrasing. He said his firm ends interviews when it believes a real-time assistant is in use, calling it "cheating" (PA via AOL, above). That is one recruiter's account of his own practice, not a validated detection method, and the signals he names are also ordinary features of nervous or over-rehearsed candidates. The mechanics of why those signals read the way they do are covered in whether an interviewer can tell you are reading your answer.

Assessment centres

Prohibited, and structurally harder to use anything in. Group exercises, in-tray tasks, and live presentations are observed by multiple assessors scoring against a rubric in real time.

Why the funnel itself changed this cycle

The misconduct numbers are not the only pressure on these employers. The same survey reports an average of 141 applications per graduate vacancy and 96 per apprenticeship, rising to 172 in London and 158 in financial and professional services (HRreview, above). ISE chief executive Stephen Isherwood attributed part of that to candidate-side automation: "Tools that enable candidates to apply for multiple jobs at once are driving up the competition and pressure on students."

77% of employers said they are reviewing or redesigning their recruitment approach. The reported measures split into two groups, and they affect candidates differently.

Measures aimed at the assessment itself:

• time-limited assessments
• randomised question banks
• more personalised interview questions

Measures aimed at application volume:

• shorter application windows and campaigns closed early
• limits on how many roles one candidate may apply for
• automated eligibility screening

The first group changes what preparation is worth doing. A randomised bank and a personalised question set both reduce the value of a pre-written answer, because the question you prepared for is less likely to be the question you get. The second group changes your timing: if a campaign can close as soon as it fills, applying late to a popular scheme is a real risk regardless of how strong the application is.

Worth noting what the survey does not show: a wholesale return to in-person selection. The reported response is redesign within the existing remote formats, not abandonment of them. Employer adoption of AI on their own side also remains limited — 16% currently use AI in recruitment, 52% have no plans to, and 29% intend to within a year (HRreview, above).

What this means for how you prepare

Find the rule before you find a tool. With 82% of employers publishing guidance, the stage-by-stage answer for the specific scheme you are applying to usually exists in writing. A general assumption is worse than a two-minute check of the portal.

Treat the stages as separate decisions. An employer who invites AI help on the application form and disqualifies for it in the interview is not being inconsistent — they are grading different things at each stage, and the written application is the one where the output is the deliverable.

Prepare for variance rather than for a script. Against randomised banks and personalised questions, the useful preparation is being able to talk about your own experience in several directions, not having the most polished version of one answer. Isherwood's advice in the PA coverage pointed the same way: research the role, because that is what interviewers are assessing.

If the employer's rule is specific to a stage, the practical question is which stage you are in. The rules for asynchronous, unproctored work differ substantially from live sessions, which the 2026 rules for take-home coding assignments sets out, and named-employer policies in US tech vary even more widely, as which companies permit AI in engineering interviews documents.

What this evidence does not settle

This is UK early-careers data from 155 employers, weighted toward large graduate-scheme employers; ISE's membership includes organisations such as Marks & Spencer, Rolls-Royce, and Accenture, though respondents were not named. It does not describe experienced-hire technical loops, US employers, or startups.

It also does not tell you the consequence of a finding. The coverage reports what employers suspect and what they prohibit, not what they do about it — whether a case ends in a rejection, a withdrawn offer, or a note on a file. Isherwood's framing was about fit rather than sanction: "If you misrepresent yourself, you could end up doing the wrong job." That is a real risk, but it is not a published sanctions policy, and no one should infer one from these figures.

The practical conclusion

For a UK graduate scheme or apprenticeship in the 2026/27 cycle, assume AI help is acceptable on the written application unless told otherwise, assume it is prohibited from the first online assessment onward, and check the employer's published guidance rather than either assumption. The stage boundary is the rule that matters, and 2026 is the first cycle where most employers wrote it down.

Control is a desktop AI interview assistant for Windows and macOS, built for live interviews, assessments, and screen-share workflows. It does not change whether an employer's rule applies to a stage, and nothing about its design makes a prohibited use permitted. Where you want to practise against varied questions before a real interview, a scored AI mock interview is the more useful starting point, and new accounts include a small free allowance to try the workflow first.