The Vettasy interview integrity report: AI use across 14,260 sessions

Across 14,260 interview sessions between September 2025 and September 2026, Verified Session detected AI use in 42.6% of sessions and 47.8% of candidates, after disputes. The number depends entirely on where the line is drawn: count context signals such as a second monitor and it rises to about 66%. The rate is a lower bound, it grew 2.2 times over the period, and 3.0% of sessions were fully cleared on dispute. We publish the method because a single percentage without one means little.

This is the first Vettasy interview integrity report. It covers 14,260 interview sessions run in Vettasy between September 2025 and September 2026, and it reports how often Verified Session detected AI use, by what method, how the rate changed, and how often a signal turned out to be wrong. We publish the method in full, because a detection percentage without a stated method is not comparable to anything.

The headline, and why it needs a line

Across the 14,260 sessions, Verified Session detected AI use in 42.6% of sessions (6,077 of 14,260), after disputes were resolved. Measured by candidate, 47.8% (5,488 of 11,480) had at least one confirmed session. Before disputes, 45.6% of sessions were flagged.

That number is only meaningful next to the line we drew for it. We count a session as showing AI use only when a first-level signal survived review: a running copilot process, a window hidden from screen capture, system audio routed into a third-party app, an active remote-control session, a large paste with no typing, a virtual camera, or the interview running inside a virtual machine. We do not count context-level facts. A second monitor, for example, appeared in 41% of sessions (55% for engineers) and is recorded as a note, not a detection. If we counted context signals, the figure would be about 66% instead of 42.6%.

The gap between 42.6% and 66% is the whole point. A single "AI detected" percentage is not comparable across tools until each says where the line between a context signal and a detection sits. Ours is the conservative figure: a signal that a person could dispute, that survived that dispute.

By method

The 6,077 flagged sessions carried 6,602 surviving signals, about 1.09 per session, because some sessions showed more than one. The share of flagged sessions by method:

Shares add to more than 100% because a session can carry several signals. The two most common methods, a known copilot and a hidden window, together account for most flagged sessions, and both are exactly what a browser tab cannot see.

By role

The rate varies widely by role. Measured by session:

Engineering roles sit around 53%, sales roles around 21%. The pattern fits the tools: copilots are strongest on technical questions with checkable answers and weakest on a conversation about how someone handled a difficult customer.

The trend

The rate rose steadily over the 13 months, from 24.1% of sessions in September 2025 to 52.4% in September 2026, a 2.2 times increase. Monthly, by session: 24.1, 25.8, 27.0, 28.4 (December 2025), 31.2, 33.6, 36.9, 39.5, 42.3, 45.0, 47.2, 49.8, 52.4.

This is the strongest number in the report and also the one to read most carefully. Our detection rules improved over the same period, so part of the rise is more use and part is better detection. The two cannot be fully separated from this data, and we do not claim the whole 2.2 times is growth in use. The sample also leans recent: 64% of the sessions fall in 2026, which is why the whole-period average is lower than where the rate stands now.

Disputes, and our false-positive rate

A signal is a fact, not a verdict, and any line can be disputed within 14 days. Publishing how often disputes succeed is part of the point, so here it is.

Of 7,130 confirming signals, 1,611 were disputed (22.6%), and 528 were withdrawn on review (32.8% of disputed signals, 7.4% of all signals). At the session level, 431 of the 6,508 flagged sessions were fully cleared, so the false-positive rate across the whole sample is 3.0%. The median dispute took 6 days to resolve; the 90th percentile took 14.

Withdrawals are uneven by method, which tells us where detection is weakest:

A running copilot is rarely overturned; a virtual machine, which has many legitimate uses, is overturned almost half the time. This is why a virtual machine is a weaker signal and why no single line should carry a decision on its own.

What this is not

A few limits, stated plainly:

• It is a lower bound. Verified Session detects tool use that leaves an instrumental trace. An attempt with no such trace, roughly one in five by our own estimate, is not caught. The real rate is higher than 42.6%.
• A signal is not cheating. Every method above has innocent uses, and the report prints them. Detection is a fact about the machine, weighed by a person, not a verdict.
• The rules changed. Detection improved across the period, so the trend mixes more use with better detection.
• The sample is skewed recent and to roles that use Vettasy, so it is not a survey of all hiring.

We publish the method and these limits so the figure can be checked and argued with, rather than quoted as a slogan. That is the same principle Verified Session runs on: a fact, its harmless causes, and a person who decides.

Sources

• Vettasy interview integrity report, 2026. Internal analysis of 14,260 interview sessions, September 2025 to September 2026 (September through the 26th).
• Vettasy, Verified Session and Trust center. See also how to read a session report.

Written by the Vettasy team

Vettasy is a desktop application for job interviews. Employers run structured, recorded interviews in it with verified participants. It is free for candidates. Windows and macOS.