The AI trust gap needs dates beside its percentages

Slack's 233% rise in daily AI use and its finding that 48% of desk workers felt uncomfortable admitting AI use are both real published statistics. They are not results from one survey. The growth figure appeared in June 2025; the disclosure finding appeared in November 2024. Usage research, disclosure research.

Put them beside each other without dates and a plausible story writes itself: people are using AI more while telling managers less. The sources establish two observations. They do not establish that combined trend.

I would reject that chart before debating its leadership advice. A dashboard can contain only accurate numbers and still make an unsupported claim through the way those numbers are arranged.

Two surveys cannot become one cohort

The 2025 study surveyed 5,156 workers in six countries between April 9 and May 1. The 2024 study surveyed 17,372 workers in 15 countries during August. The second release places the 48% statistic in its detailed global findings. These differences are stated in Salesforce's own methodology sections. 2025 method, 2024 method.

Different samples do not make either survey useless. Repeated surveys can reveal trends when their definitions and sampling support the comparison. The problem here is claiming a relationship between usage and disclosure without a matched measurement of both at the relevant points.

Perhaps disclosure became easier as AI use became ordinary. Perhaps workers became more discreet as adoption expanded. Perhaps the pattern differs by role. The two headline numbers cannot choose among those explanations, so none should become a management diagnosis by implication.

The chart needs a join rule

In a hypothetical adoption review, I would put three items beside every statistic: what was measured, whose answers it represents and when those answers were collected. Then I would require a separate sentence explaining why two statistics belong in the same comparison. That sentence is where weak joins become visible.

Suppose one department reports tool usage from access logs while another reports comfort with disclosure through an anonymous questionnaire. Even when both measurements are from this month, they answer different questions. A login does not reveal whether someone trusts the tool, and discomfort does not prove undisclosed use.

I would use the survey to identify a question worth asking locally: do employees understand when AI assistance is acceptable, and can they discuss mistakes without being penalised for the disclosure itself? I would not use it to identify supposed secret users. That moves from evidence to accusation without an intervening observation.

The practical next step is a small, repeatable measurement under consistent definitions. Ask the same disclosure question alongside the same usage question, record the field dates and retain the wording. Report whether the respondents changed before drawing a line between rounds.

This is an evaluation design, not a survey I have run. It can improve comparability, but it still needs care around response bias and the difference between what people report and what happened. An anonymous answer deserves to remain anonymous; joining it to identifiable usage logs would change the exercise.

The leadership concern survives the correction. Clear expectations and safe disclosure may be worth working on without a dramatic 2026 trust-gap statistic. Better evidence tells us which intervention to try and what result would count as improvement.

Keep the dates and populations attached to the numbers, especially when the story depends on joining them.