You are in a meeting.
The question is messy.
Why are customers leaving?
What should we change?
Why is adoption still low?
For a while, nobody has a good answer.
Then someone opens an AI assistant.
Thirty seconds later, six sensible options appear.
The room changes.
People start talking.
Someone says, “That’s a good starting point.”
And it is. AI has done something genuinely useful. It removed the blank page. It gave the team more ideas, faster. Quite possibly, better ones too.
So what could possibly be wrong? Maybe nothing.
But something else has quietly changed.
Before the AI answered, the team was searching.
After the answer appeared, the team started evaluating.
Which of these six options is strongest?
Which is realistic?
Which can we implement?
The conversation is moving.
But did we explore?
Or did we just get options?
More options are not always more exploration
Generative AI can produce ten ideas in the time it once took us to produce one. But ten ideas do not necessarily represent ten different directions. They may be ten houses on the same street.
Research is beginning to show why. In one experiment, people using AI-generated ideas produced stories judged more creative, better written and more enjoyable. Good news.
But the AI-assisted stories also became more similar to one another.
Other experiments point to a related risk: early AI-generated examples can shape what people explore next, increasing fixation and narrowing variety.
An early suggestion does more than give us an option.
It gives us a reference point before we have explored our own.
And once something plausible is in front of us, the question changes.
Not only:
What else could be true?
But:
Is this good enough?
When do we decide we have searched enough?
Recent research comparing different ways of working with AI makes that question more interesting.
When AI led with questions, people explored a more diverse idea space than when the model led the interaction.
Same broad technology.
Different interaction.
Different human behavior.
The issue may not simply be whether AI joins the search.
It may be what job we give it when it does.
Sometimes stopping early is rational
This does not mean teams should explore forever.
Sometimes the first workable answer is exactly what we need.
If the task is routine, convergence is useful.
Quickly summarize this.
Compare these options.
Draft this email.
Fine.
The problem is different when the task is ambiguous.
Why is this transformation failing?
What caused this quality problem?
Which market should we enter?
What are we missing?
There may be several plausible explanations.
The first plausible answer is not necessarily the stopping point.
Yet AI can make it feel like one.
Not because people are lazy.
Not necessarily because they are biased.
Imagine you need a recommendation by Friday.
Searching further costs you time now. Your manager wants progress. Your target rewards delivery. Challenging the emerging answer may slow everyone down.
And the benefit of finding a much better option?
Perhaps it appears six months from now.
Perhaps another team gets the credit.
Perhaps you will not even be in the role anymore.
What happens when the cost of searching is yours, but the benefit may belong to someone else?
Suddenly stopping looks rather rational.
Research suggests that organizational conditions can make this more likely. Under threat, people and organizations can narrow their search and rely more on familiar responses.
Pressure does not always narrow thinking.
There is a difference between:
I need to get this right.
and
I need to get this done.
AI enters both situations.
But it may affect them differently.
AI does not only change how we think.
It may change when we stop thinking.
AI did not create these pressures.
It changed the economics of responding to them.
A plausible answer that once took hours to produce can now appear in seconds.
Perhaps we are giving AI the wrong first job
Before asking AI for the best answer, there may be a more important decision:
What do we need AI to help us do right now?
For a routine task, let it help you converge.
For an ambiguous problem, don’t ask AI to choose too early.
First use it to widen the search.
Show me three fundamentally different ways to frame this problem.
What assumptions are we making without noticing?
What would someone from a completely different discipline see here?
Then challenge what emerges.
What evidence would prove our favorite explanation wrong?
What possibility have we ruled out too quickly?
Only then move toward choosing.
But its job has changed.
One moves us toward an answer.
The other protects the search a little longer.
That extra friction may feel inefficient.
Sometimes it is exactly what protects the quality of the decision.
Sometimes AI’s first job should be protecting the search.
The question is not whether to stop
Thinking always has to stop somewhere.
The problem is not stopping. It is stopping without noticing that we stopped.
So the next time AI gives you a good answer quickly, ask one more question:
Who decided the search was finished?
#AIxBehavioralEconomics #HumanValueInAnAIShapedWorkplace
About the author
Sabina Herwix works at the intersection of organizational change, practical learning design and human behavior. She is currently exploring how people think, decide and adapt in an AI-shaped workplace. Sabina Herwix | LinkedIn
Research note
AI tools supported literature discovery and synthesis. The sources, claims and interpretations were reviewed against the original research, and the final argument and editorial choices are the author’s.
References
1. Generative AI enhances individual creativity but reduces the collective diversity of novel content - Doshi, A. R., & Hauser, O. P. (2024).
https://doi.org/10.1126/sciadv.adn5290
Supports: AI-assisted stories were rated as more creative, better written and more enjoyable, but they also became more similar to one another.
2. The Effects of Generative AI on Design Fixation and Divergent Thinking - Wadinambiarachchi, S., Kelly, R. M., Pareek, S., Zhou, Q., & Velloso, E. (2024).
https://doi.org/10.1145/3613904.3642919
Supports: seeing AI-generated examples can increase fixation and lead people to produce fewer ideas, with less variety and originality.
3. Partnering with Generative AI: Experimental Evaluation of Model-Led and Human-Led Interaction in Human-AI Co-Creation - Maier, S., Schneider, M., & Feuerriegel, S. (2026)
https://doi.org/10.1145/3772318.3791185, open-access manuscript: https://arxiv.org/abs/2510.23324
Supports: When AI guided people mainly through questions, they explored more diverse ideas and felt more ownership than when the AI took the lead in developing the solution.
4. Old Habits Die Hard: A Review and Assessment of the Threat-Rigidity Literature - Mazzei, M. J., DeBode, J. D., Gangloff, K. A., & Song, R. (2025)
https://doi.org/10.1177/01492063241286493
Supports: under threat, organizations can narrow how they process information and fall back on more familiar responses.
5. Changes in the Work Environment for Creativity During Downsizing - Amabile, T. M., & Conti, R. (1999)
https://doi.org/10.5465/256984
Supports: during downsizing, creativity and the conditions that support it can deteriorate.
6. Productivity, counterproductivity and creativity: The ups and downs of job insecurity - Probst, T. M., Stewart, S. M., Gruys, M. L., & Tierney, B. W. (2007).
https://doi.org/10.1348/096317906X159103
Supports: job insecurity can reduce creative problem solving even when visible productivity increases.
7. Horizon problem and firm innovation: The influence of CEO career horizon, exploitation and exploration on breakthrough innovations - Cho, S. Y., & Kim, S. K. (2017).
https://doi.org/10.1016/j.respol.2017.08.007
Supports: shorter CEO career horizons are associated with fewer breakthrough innovations.
8. The Economic Implications of Corporate Financial Reporting - Graham, J. R., Harvey, C. R., & Rajgopal, S. (2005)
https://doi.org/10.1016/j.jacceco.2005.01.002
Supports: near-term performance targets can push executives toward decisions that sacrifice longer-term value. (401 financial executives were surveyed; 78% said they would give up economic value for smooth earnings, and 55% said they would avoid starting a very positive-NPV project if it meant missing the current quarter’s consensus earnings.)
9. How Performance Pressure Influences AI-Assisted Decision Making - Haduong, N., & Smith, N. A. (2024; revised 2025)
https://arxiv.org/abs/2410.16560 (preprint, currently under submission)
Supports: performance pressure does not always make AI-assisted decisions worse; depending on the situation, it can improve or reduce how carefully people use AI advice.
10. When Thinking Pays Off: Incentive Alignment for Human-AI Collaboration - Holstein, J., Hemmer, P., Satzger, G., & Sun, W. (2026)
https://arxiv.org/abs/2511.09612
Supports: incentives can change how people rely on AI. Well-designed incentives can reduce overreliance, while poorly designed ones can make it worse.
