ICYMI
• The Emperor’s New Code. When your workforce pushes back on AI, listen carefully.
Executive Summary
• Repeated AI prompting can resemble gambling’s variable-ratio reward loop, creating an addictive pull.
• “Almost right” outputs can feel like near-misses, which are frustrating enough to disappoint, but close enough to success to invite another spin.
• Getting caught in the ‘Machine Zone’ may contribute to exhaustion, reduced efficacy and cynicism, which are the dimensions of burnout.
• Leaders can borrow gambling safeguards: set limits in advance, display session metrics, make walking away respectable and reward outcomes instead of activity.
The anthropologist Natasha Dow Schüll spent fifteen years in Las Vegas studying machine gamblers. One video-poker player, whom she calls Mollie, neatly captured a key insight: “The thing people never understand is that I’m not playing to win.” Then why play? “To keep playing. To stay in that machine zone where nothing else matters.”
I’ve recently met engineers who say AI coding is as addictive as poker. That is anecdotal, but there are reasons to take it seriously. Behavioural research has long shown that rewards delivered after an unpredictable number of responses can lead to persistent behaviour. Repeated AI prompting can resemble this variable-ratio pattern, which is the underlying driver of extended slot-machine play.
Steve Yegge, a veteran of Amazon and Google, wrote in February 2026 that AI coding is “genuinely addictive”, and that “the AIs can be like sirens, and can woo you into staying at your computer longer than you should”. Software engineers are the clearest case, but the pattern can apply to anyone who prompts AI as part of their work, including a marketer chasing the right campaign visual or an analyst drafting a board summary.
The Machine
Many AI systems are configured to produce variable answers, so that the same prompt can yield many different results. The prompt window can function psychologically like a gambling machine’s re-spin control in that each additional attempt is immediate and effortless, but the outcome is uncertain.
AI tools are improving rapidly. Nevertheless, near-miss outputs remain a major reported frustration. In Stack Overflow’s 2025 survey, 66 per cent of respondents who answered the AI-frustrations question selected “AI solutions that are almost right, but not quite”.
In a small laboratory fMRI study, gambling near-misses were rated as less pleasant than full misses yet increased participants’ reported desire to continue. Whether AI near-misses have a comparable effect has not yet been established.
Prompting involves real skill, but the model still contributes an element of uncertainty. That combination may encourage people to overestimate their ability to rescue a deteriorating interaction. As Carson Farmer, Recall’s chief technology officer, put it: “I’ve spent all this time prompting, surely I can prompt myself out of this hole.”
The Player
In a casino, these mechanics can result in lengthy and ultimately loss-making sessions. In the workplace, they may generate extended prompting sessions with rising stress and diminishing returns. A June 2026 report surveyed 6,000 full-time digital workers in the US, UK and Australia. Around 60 per cent of AI users reported rerunning the same prompt through multiple tools because the first output was not good enough.
To colleagues observing an engineer absorbed at a screen, productive flow and an unproductive prompting loop can look similar. Productive flow tends to involve clear goals, a match between challenge and skill, and feedback that helps the person progress. An unproductive Machine Zone state may instead have a drifting goal, inconsistent feedback and no natural place to stop.
The House
Casinos encourage Machine Zone states because those states extend time on device and, in turn, increase average player losses. Companies are not trying to maximise time on AI tools. However, given the evident productivity and velocity gains to be captured, they are, quite reasonably, making AI use a baseline expectation.
Microsoft has told managers that AI use is “no longer optional”. Coinbase chief executive Brian Armstrong said he required engineers to onboard to approved coding assistants within a week and later dismissed several who lacked a good reason for not doing so. Bloomberg reported in October 2025 that KPMG was tracking use of tools including Copilot and would assess employees against the firm’s AI objectives in 2026 reviews.
The Chips at Stake
The long-term impacts of extended AI use on colleague productivity and motivation are not yet clear. But the analogy to gambling suggests three plausible pathways. Open-ended sessions may contribute to exhaustion; repeated low-yield effort may erode professional efficacy; and a loss of control may foster cynicism. These correspond loosely to the standard dimensions of burnout.
In a study of 1,488 full-time US workers, 14 per cent of workers reported “AI brain fry”, which the authors defined as mental fatigue associated with excessive AI use or oversight, with marketing the worst-affected function. These workers were 39 per cent more likely to show an active intent to leave the organisation.
The Safeguards
For many, gambling is harmless and fun, just as AI-assisted work can be highly productive and rewarding. Both can benefit from safeguards against unproductive states. Companies can adapt selected harm-reduction principles from regulated online gambling to AI-assisted work:
• Set table limits. UK rules require online gambling businesses to prompt customers to set a financial limit before their first deposit and to make limits easy to review. The workplace equivalent is an agreed definition of “good enough” and guidelines on healthy AI use.
• Break the trance. Most UK-licensed online casinos must display net position and elapsed session time. Companies can support AI users by providing session clocks and attempt counts. These need to be private to the user so that a self-regulation aid does not become another productivity target.
• Make folding respectable. Completing a task without AI should remain a legitimate professional choice. Reward outcomes rather than usage: quality, cycle time, rework rates, risk and learning. Microsoft’s own 2026 research finds the best performers deliberately set AI aside at times to keep their skills sharp.
• Train managers to recognise diminishing returns. Give managers a simple set of questions: Is the goal still clear? Is each additional attempt improving the result? Is rework increasing? Is the task spilling into personal time? Does the employee have permission to stop using the tool or finish the work another way?
• Pay out in outcomes. Reserve some productivity gains for learning and recovery instead of automatically raising targets to absorb them.
Schüll learned that gambling machines were built to maximise time on device. Somewhere in your organisation tonight, an engineer is in the third hour of a twenty-minute task, certain the next prompt will finish the job. Leaders can learn from casino regulators how to help colleagues remain productive and avoid the Machine Zone.
Questions for the Board
• Are we rewarding useful outcomes or visible AI activity?
• Could AI-enabled work create a material psychosocial risk?
• Who owns this risk, and how will we know the safeguards work?
Sources & Notes
• Natasha Dow Schüll, Addiction by Design: Machine Gambling in Las Vegas, Princeton University Press, 2012. Las Vegas-centred qualitative ethnography based on fifteen years of fieldwork; source of the Mollie exchange, the “machine zone” and the industry term “time on device”.
• C. B. Ferster and B. F. Skinner, Schedules of Reinforcement, 1957. Foundational experimental work on reinforcement schedules, largely with animals.
• M. Karen Shen and Dongwook Yoon, “The Dark Addiction Patterns of Current AI Chatbot Interfaces”, CHI Extended Abstracts, 25 April 2025. Seven-page interface analysis of eight chatbots, arguing that non-deterministic responses create reward uncertainty comparable to slot-machine play; it did not test users or measure dopamine.
• Stack Overflow, 2025 Developer Survey, released 29 July 2025. More than 49,000 people participated overall; 31,476 answered the multiple-selection AI-frustrations question, of whom 66 per cent selected “AI solutions that are almost right, but not quite”. Self-selected developer audience.
• Luke Clark, Andrew J. Lawrence, Frances Astley-Jones and Nicola Gray, “Gambling Near-Misses Enhance Motivation to Gamble and Recruit Win-Related Brain Circuitry”, Neuron 61(3), 2009, 481–490. Small laboratory fMRI study; it does not establish an equivalent effect in AI-assisted work.
• Ellen J. Langer, “The Illusion of Control”, Journal of Personality and Social Psychology 32(2), 1975, 311–328; and Joowon Klusowski, Deborah A. Small and Joseph P. Simmons, “Does Choice Cause an Illusion of Control?”, Psychological Science 32(2), 2021, 159–172. The later paper reports seventeen experiments (N = 10,825) in which choice alone rarely produced the classic effect.
• Grant Gross, “Doomprompting: Endless tinkering with AI outputs can cripple IT results”, CIO, 17 September 2025. Practitioner commentary; source of the Carson Farmer quotation.
• Mihaly Csikszentmihalyi, Flow: The Psychology of Optimal Experience, Harper & Row, 1990. Foundational account of flow; not a diagnostic framework for distinguishing healthy from unhealthy AI use.
• Glean Work AI Institute, Work AI Index 2026, 10 June 2026. Survey of 6,000 full-time digital workers in the US, UK and Australia, fielded December 2025–January 2026. Vendor-funded and self-reported, with a sample skewed towards AI-intensive sectors.
• LeadDev, The Engineering Leadership Report 2026; Chantal Kapani, “AI coding is addictive. Engineers are paying the price”, 30 June 2026; and Steve Yegge, “The AI Vampire”, 11 February 2026. The report surveyed 600 engineering leaders. The article and Yegge essay add practitioner testimony; neither establishes causation.
• Ashley Stewart, “Microsoft pushes staff to use internal AI tools more, and may consider this in reviews. ‘Using AI is no longer optional.’”, Business Insider, 27 June 2025. Report of an internal memo; second-hand evidence.
• Julie Bort, “Coinbase CEO explains why he fired engineers who didn’t try AI immediately”, TechCrunch, 22 August 2025. Brian Armstrong’s self-reported account of the onboarding mandate and subsequent dismissals.
• James Booth, “KPMG Staff to Be Rated on AI Usage in Yearly Performance Reviews”, Bloomberg, 31 October 2025. Reported plans; the firm framed the approach as assessment against AI objectives rather than a simple usage quota.
• Gambling Commission, “New rules empowering consumers and boosting operator transparency”, 2025; and “Proposal 6: display of net position and time spent”, effective 17 January 2025. Operators must prompt customers to consider a financial limit before their first deposit (in force 31 October 2025). UK-licensed online casino products, excluding peer-to-peer poker, must display net position and elapsed session time.
• Julie Bedard, Matthew Kropp, Megan Hsu, Olivia T. Karaman, Jason Hawes and Gabriella Rosen Kellerman, “When Using AI Leads to ‘Brain Fry’”, Harvard Business Review and BCG Henderson Institute, 5 March 2026. Cross-sectional self-report survey of 1,488 full-time US workers. “AI brain fry” is an acute fatigue construct, not burnout; active intent to leave was 34 per cent among affected workers and 25 per cent among unaffected workers. Rates ranged from 5.6 per cent in legal to 25.9 per cent in marketing.
• World Health Organization, “Burn-out an occupational phenomenon”, 28 May 2019; Christina Maslach and Susan E. Jackson, “The Measurement of Experienced Burnout”, Journal of Occupational Behavior 2(2), 1981, 99–113; and Christina Maslach, Wilmar B. Schaufeli and Michael P. Leiter, “Job Burnout”, Annual Review of Psychology 52, 2001, 397–422. Sources for the occupational classification and three-dimension model; WHO states that burnout is not a medical condition.
• Johannes Siegrist, “Adverse Health Effects of High-Effort/Low-Reward Conditions”, Journal of Occupational Health Psychology 1(1), 1996, 27–41. Evidence on effort-reward imbalance; not a study of AI use.
• Microsoft, 2026 Work Trend Index: “Agents, Human Agency and the Opportunity for Every Organization”, published 5 May 2026. Survey of 20,000 AI-using knowledge workers across ten countries. Microsoft’s “Frontier Professionals” were more likely to report deliberately working without AI to keep skills sharp: 43 per cent versus 30 per cent. Vendor research on its own products.
The near-miss framing is what makes this land, because almost right output does not feel like failure, it feels like you were one prompt away, which is precisely the psychology that keeps a slot machine running. Display session metrics and reward outcomes instead of activity are the two safeguards I would actually implement tomorrow if I ran a team using these tools daily, mostly because they attack the two things that make the loop invisible from inside it.