Psychology Today62%
When AI Finishes Your Thoughts, What Happens to Them? 37%
By Grant Hilary Brenner MD45% DFAPA45%
8/2/2026, 12:55:17 PM
BS Summary: This article contains 22 faulty reasoning types, including Biased Writer Voice, Negativity Bias, and Confirmation Bias, with Appeal to Authority as the most egregious example at 18.5% saturation with 197 hits. Analysis detected 1,473 faulty-reasoning hits from 1,062 analyzed words, generating a BS Score of 38.1% and a BS Rank of 37% (16,495 of 26,004 articles). This article is better (less manipulative) than 63.40% of the article peer group.
A thought, a plan, begins forming, and before it is formulated, something else finishes it.
I keep running into it, in my work and others': the interval between a partly-formed thought and an AI's optimized response has all but closed.
It's easy to click go and harder to say no, to slow down and think for oneself—particularly when the AI is good at it.
The question is not only what these systems can do, but what happens in us when we let them do the part of thinking that until recently required staying with a problem long enough for something to form.
When Is Easy "Too Easy"?
What comes to mind here is the need to manage one's own sense of being a follower or a leader, and the desire to succumb to seduction.
Self-monitoring and restraint are required to hold steady in such currents.
With AI there is an experience, but not another person—not in the ordinary sense of someone with a body, needs, vulnerability, a history, and a life we can affect in return.
What overlaps is a working surface—language, concepts, alternatives, actions and feedback—not a shared mind.
Large language models appear person-like for understandable reasons: they use language, take turns, answer quickly, carry context forward, and arrive through interfaces built to feel socially present, even warm and friendly.
Selling tokens.
Some of that is deliberate and, in engineering terms, a dependency: to be usable, they must conform to human limits—our tempo, our narrow channel, our readiness to find a face in things.
Design is not necessarily manipulation; the pull can come from affordance, incentive, how easily we read a mind into fluent language, or how readily we stop thinking when tired or conditioned.
The pull is measurable, and it adds up.
Across experiments with 1,401 participants, a biased AI amplified perceptual, emotional, and social judgments in a direction human interaction did not, and the effect grew over rounds (Glickman & Sharot, 2025).
The drift appeared even when people merely believed an AI was answering, and they underestimated how far they had moved.
Score!
There is a real gain on the other side.
We have a slow bitrate and a narrow aperture on the world; attention and working memory are limited; we cannot weigh every alternative or hold every possible future in mind while deciding.
These systems compare and generate at a scale we cannot take in directly, making unmanageable complexity available at the tempo of thought—powerful tools, in the right hands.
The difficulty begins when speed is mistaken for development.
Letting an AI finish, or even supply, one's thoughts can interrupt important learning and developmental chains: a person may arrive at a polished sentence, plan, or interpretation without having formulated the question, tolerated the uncertainty, found the error, revised, and decided what mattered.
The output may look better; the route may be thinner.
The self-monitoring question: am I extending thought or escaping effort—does the system offer something to evaluate, or supply the framing?
We may think we are getting more done than we are—“Accomplishment Hallucination,” as I've called it (Brenner, 2026).
The closest thing to a direct test is more pointed than I expected.
In a randomized trial with roughly 1,000 high schoolers (Bastani et al., 2025), an unrestricted GPT-4 assistant sharply raised math practice scores, then lowered later exam scores once withdrawn, relative to controls.
A second version, whose hints withheld answers, produced larger practice gains and no penalty afterward.
Same model, different mode of help, opposite effect on what a student could do alone.
Mainly, we still do not know how these systems work, though we built them, or how to use them well for learning.
Research with adults overlaps.
Among 319 knowledge workers using generative AI at work (Lee et al., 2025), higher confidence in the tool tracked with less critical thinking; higher confidence in themselves, with more.
One trial and one survey settle little, and show nothing about injury or atrophy.
They point to performance and learning moving in opposite directions (Yan et al., 2025), unnoticed—as the contested handwriting-versus-typing research suggests (Van der Weel & Van der Meer, 2024; Pinet & Longcamp, 2025).
How we do something may matter for what is learned and carried forward, beyond embodied cognition while including it.
Medical education has a vocabulary for this—deskilling, mis-skilling, never-skilling (Abdulnour et al., 2025)—the third matters most: not losing a capacity but never building it.
With incorrect AI readings, accuracy among radiologists of fifteen years' experience fell from 82% to 45.5% (Dratsch et al., 2023); expertise is no protection against automation bias.
Frustration Maxxing
Research on effort points the same way without settling anything.
Conditions that slow visible progress—spacing, retrieval practice, difficulty worked through rather than around—improve retention; current performance is an unreliable index of learning (Bjork & Bjork, 2023).
Pain isn't always gain: unnecessary friction exhausts and excludes, and someone with an enduring cognitive limitation may need continuing support, not a lecture on doing everything unaided.
The question is not whether assistance is permitted, but what kind, for whom, to what end.
The robotic exoskeleton is a useful comparison, bounded.
Gait-training robots after stroke supplement therapy rather than replace it (Mehrholz et al., 2025); if the machine carries the load, the muscle does not get strong.
The analogy is not literal muscle, nor evidence of injury.
Three easily confused use-cases need differentiating: compensation, which lets a person function despite an enduring limitation; scaffolding, which preserves participation and correction and can fade; and substitution, in which the system supplies framing or judgment so completely that the human becomes deskilled.
Augmentation is a fourth possibility, less relevant here.
The brain is use-dependent.
As the neuroscience-famous line from Hebb goes: "Neurons that fire together wire together."
Conversely, "use it or lose it" says it more colloquially.
Developing systems are shaped by what they are given to do, and by what they never get to do.
This is where self-governance is critical, especially since frontier systems vary in how they influence us.
A useful system need not make us avoidant or suspicious; it should keep us authors of our own development, make its contribution visible to self-appraisal, leave enough friction for us to think, and offer guardrails steering us back toward effortful work, relationship, creativity, repair, and action—so we don't lose essential skills, and might develop new ones.
Analysis
Hover over highlighted words in the article to view the associated bias or fallacy analysis.