Understanding How People Change When They Know They Are Watched

The Hawthorne studies from the 1920s and 30s showed something that every manager eventually learns the hard way: when employees know they are being observed, their behavior changes. Not because they are faking anything. Because humans are remarkably sensitive to cues about what is expected of them in a given situation. This is what researchers call demand characteristics, and it shows up everywhere in an organization once you know where to look. I first ran into this problem clearly around 2014 when we rolled out a new sprint retrospective format at a mid-size SaaS company. The premise was simple enough: move from a standard status-update meeting to a blameless post-mortem structure focused on systemic issues rather than individual blame. We spent three weeks training the team leads, sent out pre-read materials, and kicked things off with what we thought was a well-brainwashed group. By the second session, the data started looking suspiciously good. Velocity went up. Complaints dropped. Cross-team dependencies disappeared from the notes almost entirely. Then an external consultant sat in on the third session and asked one question during the retro: did anyone feel pressured to keep things positive because the CEO had personally asked for these to show measurable improvement. A brief silence followed. One senior engineer admitted that several team members had been quietly coaching each other on what the right answers sounded like before the meeting started. The retro was not measuring actual team sentiment. It was measuring performative sentiment optimized for leadership expectations.

What Demand Characteristics Actually Look Like Inside a Company

Demand characteristics occur when participants in any structured interaction infer what behavior is expected and adjust accordingly. In a corporate context this can happen through formal surveys, performance reviews, focus groups, usability testing, or even casual one-on-one conversations with managers who have an agenda. The effect is not necessarily deceptive. People are usually trying to be helpful. They pick up on subtle signals about what outcome is desired and steer toward it unconsciously. The most common forms I see in practice are: Survey fatigue bias: when employees fill out an engagement survey and notice that the last three questions all use positively-framed language with only agree/strongly agree options. They rate everything highly rather than risk seeming ungrateful or politically tone-deaf.

Meeting performance theater: when a team prepares a demo or presentation specifically for leadership visibility rather than as a genuine check of working progress. Features that are half-built get glossed over while minor improvements get inflated into major wins. Manager expectation mirroring: when direct reports sense that their manager prefers a certain type of answer during planning sessions and start pre-filtering their input before speaking. A junior developer might suppress a legitimate technical concern because the last three times they raised it, the response was subtly dismissive.

Get the Full Details

Jak się przygotować do zrobienia nowej strony (dla branży technicznej ...
Jak się przygotować do zrobienia nowej strony (dla branży technicznej ...

How to Detect It Without Making Things Worse

The straightforward approach that works best in practice involves running parallel measurement streams with different incentive structures. Instead of relying on a single survey or meeting, collect data through anonymous dropbox channels, direct observation without agenda, and outcome-based metrics that cannot be gamed by behavior change alone. I developed a practical workaround during that same retrospective experiment around late 2014. We introduced a parallel feedback channel that was completely unconnected to the formal retro process. A Slack bot asked one random question per day to a different team member, something unstructured and low-stakes, like what was one thing that slowed you down this week that you would not normally mention in a meeting. We kept the responses private from management and only aggregated them monthly. Within six weeks the data from this anonymous channel diverged significantly from the formal retro numbers. The actual pain points were roughly three times more frequent and covered different categories entirely than what was being reported in the structured sessions. The key difference was that the anonymous channel removed the demand characteristics entirely. People felt safe reporting real problems because there was no audience, no performance expectation, and no career risk attached to honest input.

Counter-Intuitive Insights Most Beginners Miss

Here is something that surprises people who read the textbook definition: demand characteristics are not always a bad thing. In controlled settings they can actually improve short-term performance. When employees know they are being observed during a new process rollout, they often put in extra effort simply because the attention signals that something matters. The problem is not the behavior change itself. The problem is mistaking temporary compliance for genuine adoption. Another nuance that beginners typically overlook is that removing demand characteristics entirely is impossible. Any structured interaction contains some level of expectation signal. The goal is not elimination but minimization and awareness. You should aim for a 40 to 60 percent reduction in performance distortion rather than a perfect zero, which would require removing all human interaction and defeat the purpose of the exercise. A specific pitfall I encountered involved a client who tried to solve demand characteristics by eliminating all feedback meetings. They replaced monthly retros with an automated dashboard showing raw Jira metrics and removed all team discussions about process. The result was worse, not better. Velocity data looked stable but actual ship rate dropped 18 percent over the next quarter because the team had no venue to surface systemic blockers before they became critical. The dashboard was measuring output but missing the signal about why that output was degrading.

When This Approach Fails Completely

Demand characteristics research has real limitations that practitioners should acknowledge bluntly. The method breaks down in organizations with fewer than 15 people where anonymity is impossible to guarantee. In small teams everyone knows who submitted what feedback, and the social pressure to conform becomes stronger not weaker. I recommend an alternative approach in those cases: structured rotation of facilitators where no single manager controls the agenda, combined with external mediators who have no stake in the outcome. The approach also fails in high-turnover environments where team composition changes faster than measurement cycles. If people leave every 90 days and you measure engagement quarterly, you are capturing ghost data from employees who no longer exist in the organization. In those cases a continuous sentiment tracking tool with rolling 30-day windows provides more reliable signal, though it requires different analysis entirely. Another scenario where demand characteristics work against you involves companies with strong executive visibility culture. When the CEO personally asks for improved numbers during a transformation initiative, the pressure to deliver becomes structural rather than incidental. Employees read this correctly and optimize for the desired outcome. The organization then has accurate data about performance expectations but misleading data about actual working conditions. In these cases the only reliable alternative is external benchmarking against industry standards combined with exit interview analysis from departing employees who have no incentive to protect current leadership.

5 raportów analitycznych Twojej strony, które powinieneś znać | ifirma.pl
5 raportów analitycznych Twojej strony, które powinieneś znać | ifirma.pl

Practical Steps That Actually Cut Down Measurement Time

Here is the streamlined version of what I have found useful across multiple organizations over the past decade. Start by identifying the single highest-stakes measurement event in your current calendar, whether that is an annual engagement survey, a quarterly business review, or a monthly sprint retrospective. Run a parallel anonymous channel alongside it for two full cycles before making any decisions based on the formal data. Compare the divergence between the two streams. If they align within 15 percent you have reasonable confidence. If they diverge beyond 40 percent you need to redesign the formal channel entirely. This usually cuts the process down from about 2 hours of meeting preparation per week to roughly 20 minutes of setup time, depending on your existing tooling and team size. The bulk of the savings comes from not having to manage performative data through multiple layers of review and adjustment. The most practical implementation I have seen involves a simple three-question daily prompt delivered through an existing communication channel that everyone already uses, paired with a monthly aggregate report that is shared with the entire team rather than locked away in HR files. The transparency about what is being measured reduces the demand characteristics naturally because people know the data is public and cannot be used selectively for political purposes.