This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: Midjourney Mentalität praktisch erklärt.

In a field experiment with 758 consultants at Boston Consulting Group, the participants with access to GPT-4 completed 12.2 percent more tasks, worked 25.1 percent faster, and produced work of significantly higher quality.

The same study carries the counter-story. One task was deliberately placed beyond current AI capability, past what the authors call the jagged technological frontier. There, consultants using AI were 19 percentage points less likely to produce a correct solution than those working without it. Access to the machine made the work worse.

That pattern now shows up everywhere. Managers report faster output with thinner argument. Universities see formally convincing work whose authors can no longer explain their own reasoning. Recruiters receive applications that sound more professional than ever and are almost impossible to tell apart.

The problem was never the use of AI. It is the moment at which it enters.

What the Midjourney mentality looks like at a desk

Earlier pieces on this site describe the Midjourney mentality as a cultural pattern: the tendency to delegate creative and cognitive work to generative systems early. In professional and academic practice the definition can be made precise.

The Midjourney mentality is the use of AI at a point in the thinking process where understanding, structuring and judging have not yet happened.

The shift feels efficient, and it reaches into the development of basic capabilities, measurably so.

Michael Gerlich at SBS Swiss Business School studied the relationship between AI use and critical thinking in 666 participants of varying age and educational background. Frequent AI users scored significantly worse on the Halpern Critical Thinking Assessment, and younger participants were hit hardest: they used AI most intensively and showed the lowest scores.

The mediating factor was cognitive offloading, the outsourcing of mental work to external tools. Cognitive psychology has described the phenomenon for years, and AI sharpens it in a new way. Calculators took mental arithmetic. Search engines took the memorizing of facts. Generative AI takes something qualitatively different: the formulating, weighing and structuring of thought.

Thinking happens in phases, and so should the tool

Knowledge work is not linear, and the useful distinction comes out of learning research.

Phase one is your own understanding. Problem definition, first hypotheses, priorities. The work feels slow and often frustrating, which is exactly what makes it valuable. Robert and Elizabeth Bjork call this principle desirable difficulties: conditions that make learning harder in the short term and improve long-term retention and transfer. Bypass that resistance regularly and judgment stops being trained.

Phase two is confrontation. Testing your own assumptions and examining counter-arguments. Here AI delivers real value as a sparring partner, exposing weaknesses in an argument, supplying alternative perspectives, making blind spots visible.

Phase three is formulation. Linguistic precision, formatting, compression. This is what generative systems were built for.

The Midjourney mentality deploys AI in phase one, and high professional performance comes from deploying it from phase two onward. That distinction sounds abstract and is not.

Three ways it goes wrong

Reaching for AI before your own analysis. Start a thinking process with a generated concept and you inherit its structure and its weighting. Your own prioritization is not consciously skipped; it simply never happens, and you accept a scaffold you never built.

The BCG study makes that concrete. Its authors identified two user types: centaurs, who divided tasks deliberately between themselves and the machine, and cyborgs, who integrated AI into every step. Inside the frontier AI helped. Beyond it, where independent judgment was required, the consultants with AI access did worse as a group.

Using AI in place of forming a hypothesis. Without your own assumptions there is no yardstick against which an AI suggestion could be assessed. Results stop being examined and start being accepted or cosmetically adjusted. In organizations that appears as decision quality falling while productivity metrics rise.

Using AI for confirmation. Most people use these systems to improve text they already have. Very few use them to attack their own reasoning, even though decades of research on decision quality show that working systematically through counter-positions, in the manner of a devil's advocate, raises the quality of complex judgments considerably. AI is mostly deployed as a polishing machine and hardly ever as a test bench.

How the strong users do it

A recurring pattern emerges from conversations with experienced knowledge workers in consulting, product development and research, and it matches the study's findings.

Think first, externalize second. Sketches, notes and half-formed thoughts come before any prompt. The drafts do not have to be good; they have to create the standard against which every AI output can be measured.

Use AI as an instrument of confrontation. The question stops being "write me a summary" and becomes "what are the three strongest objections to this position" or "which assumption in my concept is the most weakly supported".

Use AI to compress, not to generate. The machine improves clarity and structure. The substance has to be there beforehand.

That approach fits the OECD's Skills Outlook 2025, which counts adaptive problem-solving among the skills of this century: reaching a goal in a changing situation where no method of solution is immediately available. No tool supplies that capacity to someone who has never practiced it.

Students carry the sharpest version of this

For students the timing question is most acute, because the goal is building competence and not producing output.

Using AI before attempting your own solution accelerates the short term and weakens the learning curve. That is a claim from cognitive science and not a moral position: learning conditions that feel harder produce better long-term results, and AI in phase one removes exactly that productive difficulty.

The labor market is already responding. In the World Economic Forum's Future of Jobs Report 2025, analytical thinking remains the top core skill, and seven in ten companies consider it essential. The analytics firm Lightcast observes that many of the fastest-growing skills in AI jobs go beyond technical expertise and call for work across organizational silos.

The irony is hard to miss: the more capable the tools become, the more valuable the ability to judge independently of them.

This is not only an individual problem

Deploy AI systematically at the wrong point and the effect reaches working cultures, training standards and the quality of decisions across an organization.

Gerlich's study offers an unsettling detail. Education level moderated the negative link between AI use and critical thinking, which suggests that learned habits of reflection, once built, provide some protection. Anyone who never develops those habits, because AI makes them look unnecessary from the start, has no such buffer.

The risk has little to do with AI making us less intelligent. It is that we train a generation of professionals who appear productive without having developed the judgment that controls their productivity.

The whole question is when

Generative AI reaches deeper into cognitive processes than any tool before it. Whether it improves work and learning turns on one question: when do we reach for it?

Preserve the early, formative part of the thinking and use AI deliberately for testing and compression, and you get quality and speed together. Automate that part and you save time and pay in substance.

The lesson worth taking from the BCG experiment is that the tool's value depends on the task. Using it well means knowing precisely when not to.

Sources

Grouped by the section they support.

Opening

  • Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon and Lakhani (2023), Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, Harvard Business School Working Paper 24-013, ssrn.com. Supports the 758 consultants, 12.2 percent more tasks, 25.1 percent faster completion, higher quality, and the 19 percentage points outside the frontier.
  • The observations on managers, universities and recruiters are the author's own.

What the Midjourney mentality looks like at a desk

  • Gerlich (2025), AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies 15(1), 6, doi:10.3390/soc15010006. Supports the 666 participants, the Halpern Critical Thinking Assessment, the results for younger participants and the mediating role of cognitive offloading.
  • Risko and Gilbert (2016), Cognitive Offloading, Trends in Cognitive Sciences 20(9), 676–688, doi:10.1016/j.tics.2016.07.002. Supports cognitive offloading as an established concept.

Thinking happens in phases, and so should the tool

  • Bjork and Bjork (2011), Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning, in Gernsbacher et al. (eds.), Psychology and the Real World, Worth Publishers, 56–64. Supports desirable difficulties.
  • The three phases are the author's own model.

Three ways it goes wrong

  • Dell'Acqua et al. (2023). Supports centaurs and cyborgs and the results inside and beyond the frontier.
  • The remark on devil's advocacy summarizes decision research in general; no single study is cited.

How the strong users do it

  • OECD (2025), OECD Skills Outlook 2025: Building the Skills of the 21st Century for All, 9 December 2025, oecd.org. Supports adaptive problem-solving and its definition.
  • The practices described come from the author's conversations with knowledge workers.

Students carry the sharpest version of this

  • World Economic Forum (2025), The Future of Jobs Report 2025, chapter 3, weforum.org. Supports analytical thinking as the top core skill, considered essential by seven in ten companies.
  • Magrini (2026), Emerging skills in AI jobs, Lightcast, 27 May 2026, lightcast.io. Supports the point that many of the top skills in AI jobs go beyond technical expertise.
  • Bjork and Bjork (2011). Supports the point that harder learning conditions produce better long-term results.

This is not only an individual problem

  • Gerlich (2025). Supports the moderating role of education level.

The whole question is when

  • Dell'Acqua et al. (2023). Supports the dependence of AI's value on the task. The conclusion is the author's own.
The link has been copied!