Should there be AI-free periods at work?

Yes, for a short list of named tasks, and the reason has little to do with discipline. A stretch without the tool is the only way anyone finds out what a person can still do. Nobody has trialled scheduled AI-free periods in knowledge work, so the case rests on measurements of what happens when the tool is taken away and on a regulator that has been asking pilots to do the equivalent since 2013.

Aviation wrote the rule down in 2013 and again in 2017#

On 4 January 2013 the US Federal Aviation Administration issued Safety Alert for Operators 13002, titled Manual Flight Operations. It says that continuous use of autoflight systems could lead to degradation of a pilot's ability to recover quickly from an undesired state, and it asks operators to write policies giving pilots appropriate opportunities to fly by hand, with guidance on when automation is the better choice, high workload being the example. A second alert, 17007, dated 4 May 2017, widens the request to training and line-operations policies, so that manual proficiency is developed and maintained and checked at proficiency checks. Both are guidance. Neither reports a measurement. Graded entry, graded entry.

The working group behind the 2013 flight-path-management report found vulnerabilities in manual handling after automation, and in how those skills are defined, developed and retained. Casner and colleagues then put 16 airline pilots through routine and non-routine scenarios in a Boeing 747-400 simulator. Instrument scanning and manual control were mostly intact even where pilots reported little recent practice. The cognitive tasks were not: tracking position without a map, deciding the next navigational step and recognising instrument failures all showed frequent and significant problems. PARC/CAST, Casner. The hands survived and the judgement did not, which is the part of the aviation record that transfers best to desk work, because desk work is mostly the cognitive part. That transfer is an inference, and the step on this page most in need of a test.

Every measured gap was found by taking the tool away#

Two studies in this evidence base show the pattern. Budzyn and colleagues compared 1,443 colonoscopies performed without AI help, before and after AI was introduced at four Polish centres, by 19 experienced endoscopists. The adenoma detection rate in unassisted procedures fell from 28.4 to 22.4 per cent. Bastani and colleagues ran a field experiment with nearly 1,000 high-school students. While the tool was present, grades rose 48 per cent for the unrestricted GPT-4 group. With access removed, that group scored 17 per cent lower than students who had never had it. A hints-only tutor largely removed the harm. Budzyn, Bastani.

In both, the gap was invisible while the tool was in use. Assisted output looked fine, and in the school trial it looked far better. A team that never works without the tool has no way to see the difference, so the strongest argument for an AI-free period is that it works as an instrument. It is the cheapest version of the removal that every one of these studies used.

What nobody has tested, and what a period costs#

The graded evidence base holds no trial of scheduled AI-free periods in knowledge work. Whether they preserve capability, how long a period must be and how often it must recur are all unmeasured, and the FAA alerts give no schedule in the passages read for this page. The supporting studies each have limits. Casner's sample is 16 pilots in a simulator. Budzyn is observational, covers one procedure in one country, uses detection rate as a proxy for skill, and a published correction exists whose content has not been read here. Bastani concerns school mathematics over a bounded period.

A period also has a price. An hour without the tool costs output that the tool would have produced, and a period that nobody scores becomes a ritual that costs the output and returns nothing. The aviation alerts handle this by being specific about conditions: practice in low-workload settings, and automation preferred when workload is high. A desk equivalent would pick moments when a slower result is cheap, and that too is an inference from the aviation text and has no trial behind it.

Fly by hand where you would be blamed for the outcome#

Rahim Hirji's essay of 1 March 2026, Are You Flying, Or Are You Being Flown?, argued that the people who stand out in the AI era "will be the ones who never forgot how to fly without it." That is a dated position and not a measurement. Read alongside the evidence above, it suggests a selection rule: protect the tasks where you would have to sign off a judgement you could not check without the underlying skill, and let the tool take the rest. A day without AI would be too blunt and an hour a year too thin. A short, named list of tasks, done unaided at an interval and compared with the last time, matches what the aviation alerts ask for and what the studies measured.

A team could treat it like a fire drill. Choose three pieces of work that matter to the team's judgement, run the unaided version at a fixed interval, keep the results, and drop any task where the gap turns out to be nil. The test for an individual is the same one set out in the dependency diagnostic: remove the tool and watch. The encyclical Magnifica Humanitas makes a related point in paragraph 140, that educating people about AI involves teaching them when and for what purpose it ought not to be used. Graded entry. It is a moral argument and carries no measurement.

Key sources

• Federal Aviation Administration (2013). SAFO 13002: Manual Flight Operations. 4 January 2013. Graded entry.
• Federal Aviation Administration (2017). SAFO 17007: Manual Flight Operations Proficiency. 4 May 2017. Graded entry.
• PARC/CAST Flight Deck Automation Working Group (2013). Operational Use of Flight Path Management Systems: Final Report. FAA. Graded entry.
• Casner, S. M., Geven, R. W., Recker, M. P. and Schooler, J. W. (2014). The Retention of Manual Flying Skills in the Automated Cockpit. Human Factors, 56(8). Graded entry.
• Budzyn, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology and Hepatology, 10(10), 896-903. Graded entry.
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning. PNAS, 122(26). Graded entry.
• Leo XIV (2026). Magnifica Humanitas, paragraph 140. The Holy See, 15 May 2026. Graded entry.
• Hirji, R. (2026). Are You Flying, Or Are You Being Flown? Box of Amazing, 1 March 2026.

Related SuperSkills research#

On how aviation handled the same problem, what professions can learn from aviation. On testing yourself, am I becoming dependent on AI and can you regain a skill you have lost. On the practice itself, the missed reps and deliberate practice. On the habit that makes the gap invisible, automation complacency.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The two FAA alerts were read at source on 4 October 2026, and the Hirji essay was re-read at source the same day. The term "AI-free period" is descriptive and not claimed. The interpretation is kept separate from the findings.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.