The answer arrived quickly. What happened to the person who still has to think?
When a tool gives me the answer faster, do I ever stop to ask what habit the shortcut is building in the person who will face the next question without it?
Learn the source
Ramchal opens zehirus by describing a person who is ‘examining and watching over his deeds and ways: are they good or not?’ His image is not of a single dramatic mistake but of walking in darkness because movement has become automatic.
That order matters. Zehirus does not begin by condemning a tool outside us; it begins by looking at the path our own repeated choices are making. A result can be useful while the habit that produced it still deserves inspection.
What happened
On October 9, UC Berkeley summarized an updated, peer-reviewed study presented that week at the Conference on Language Modeling.
Across three randomized experiments, the researchers recruited 1,222 US-based participants for fraction and reading-comprehension tasks, giving some brief access to an AI assistant before removing it.
AI assistance improved performance while available, but afterward assisted participants performed worse without it; skip rates also rose significantly in experiments one and three, while experiment two’s aggregate skip-rate difference was not statistically significant.
The authors caution against generalizing the result to every task or form of AI use and call for systems that scaffold lasting competence rather than only immediate completion.
UC Berkeley Research · Using AI for just 10 minutes erodes persistence · 9 October 2026 ↗Liu et al. · AI Assistance Reduces Persistence and Hurts Independent Performance · COLM 2026 ↗
Return to the news
AI can save time, clarify difficulty and widen access. The study does not make the tool assur, and Ramchal is not speaking about software. But its most arresting detail is that success while help was present concealed weakness when help disappeared. Zehirus asks me to inspect not only the answer I obtained but the derech by which I obtained it. If every hard paragraph is handed over before I have tried to hold the question, I may be training impatience while calling it efficiency. The avodah is not theatrical rejection of a useful tool. It is to keep one place where my own attention must stay at work before assistance enters.
Stay with the thought
The uncomfortable part of the study is not that the machine kept giving wrong answers. Often it helped participants become more accurate in the moment. The weakness appeared afterward, when the help was gone and the person still had to stay with the difficulty.
That makes the research a sharp mashal for derech. We usually inspect the output: Did I get the email, summary or translation I needed? Ramchal asks for a second inspection: what sort of walker is this repeated way of arriving at the result producing?
AI can be used as a tutor, drafting aid or partner that offers hints, objections and checks after an honest first attempt. The point is not to make struggle holy for its own sake. It is to notice whether assistance is leaving me more capable of thought or merely more accustomed to completion.
Zehirus becomes concrete before the next request: decide which part is mine to carry—the first attempt, the formulation of the question, the judgment of sources—and only then invite the tool. Hashem gave us both useful instruments and the faculties that use them; receiving help need not put those faculties to sleep.
Mesilas Yesharim · chapter 2 · Zehirus ↗Liu et al. · AI Assistance Reduces Persistence and Hurts Independent Performance · COLM 2026 ↗
Take one thing with you
Before one AI-assisted task today, spend three uninterrupted minutes forming your own attempt or question, then ask for help that responds to—not replaces—that work.
A reflection, not a personal halachic ruling.Sources & reporting notes 3 references
- UC Berkeley Research · Using AI for just 10 minutes erodes persistence · 9 October 2026 ↗
University research account published October 9, 2026 after the updated study was peer-reviewed and presented at COLM 2026. It identifies the three experiments, total recruited sample and the bounded tasks; its headline language is not treated as proof of permanent harm.
- Liu et al. · AI Assistance Reduces Persistence and Hurts Independent Performance · COLM 2026 ↗
Primary paper by Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker and Rachit Dubey. First posted April 6, 2026 and updated for the October conference. The authors caution against treating the experiments as broad evidence about every task or every kind of AI help.
- Mesilas Yesharim · chapter 2 · Zehirus ↗
Exact approved corpus excerpt and context. Ramchal discusses conscious examination of one’s own conduct, not technology policy. English is Torah Lens’s rendering; textual verification is not rabbinic approval.
Hebrew passages are quoted sources; English renderings are our own. The event connection and suggested reflection are Torah Lens’s application. Prepared with AI assistance. Rabbinic supervision has not yet been established.