In 2025, most of the major answer engines made an almost identical update. Claude, ChatGPT, Gemini, and many others started to show what they were thinking.
Answer engines would tell the user which websites it was searching, which documents it had opened, and which assumptions it was second-guessing. According to Anthropic, the company that owns Claude, the answer engine does this for three reasons.
- It helps people check and trust answers.
- It reveals mismatches between thinking and answers.
- It’s simply interesting to watch.
But I think there’s a fourth reason that AI providers have purposefully hidden from us.
The Labor Illusion
In 2011, two Harvard professors published a study in Management Science that sought to understand and challenge the idea that faster means better in a customer service context. They asked 266 participants to use a travel search site. Similar to Skyscanner or Kayak, the user could enter a destination, and the site would pull back several flights to potentially book.
However, there was a twist. One group saw just a plain loading wheel on a white background as the search engine filtered through results.
The second saw the plain loading wheel, plus a live scrolling list of which airlines were being searched, with fares visibly stacking up as they were found. Interestingly, the wait was randomly set to 10, 20, 30, 40, 50 or 60 seconds.
Once participants had seen the results, they were asked to rate the website out of seven. Turns out, the results from the loading wheel that show it’s working were perceived as 8.1% higher value than the same results from a blank loading wheel.
Incredibly, people rated the results higher even if the transparent loading wheel took longer. When the search engine showed it was working, people preferred it — even if it took 50 seconds longer to gather the same results.
Later, a new group of 118 participants was asked to pick between using a faster site with no transparent loading screen and a slower site with the transparent loading screen. The results that both sites generated were the same.
Participants were more likely to pick the slower site with the transparent search over the faster site with the blank search.
Labor looks smarter.
As with all good science, one paper from 2011 isn’t enough to draw lofty conclusions.
Fortunately, in 2022, Dimitrios Tsekouras, Ting Li, and Izak Benbasat published research in Information and Management to explore how signalling effort — even through something as simple as a spinning loader — changes how users evaluate a recommendation system.
For the first study, 306 participants were shown a search engine for cars. In a second study, 294 participants were shown an online dating app that used a loading wheel to match potential couples.
Participants were split into four groups across two variables: how much effort they had to put in upfront, and whether the system showed effort back. In the high effort conditions, participants saw a rotating loader with the words “calculating results” for seven seconds. In the low effort conditions, results appeared immediately.
The recommendations themselves were the same. The only difference was seven seconds of a spinner.
Those who saw the loader rated the recommendation agent’s quality significantly higher than those who received instant results — even though the recommendations themselves were identical.
Anthropic and OpenAI might tell us that the transparent loading screen is just there because it’s “simply interesting to watch,” but most marketers know that’s BS.
Our perception is shaped by the effort spent creating something. And most of us will prefer a slower answer engine that shows it’s working to a faster one that doesn’t.