Stop Saying "Claude Told Me To"
Last week, in 3 different meetings, from 3 different people, at three different times, I heard the exact same sentence. "Claude said this was the best approach." Not "I think," not "here's my read after checking a few sources." Just an AI's name, dropped into a business conversation like a credential. While that sentence absolutely made me want to scream, my scariest thought was that the comment isn’t more worrisome to everyone else in the room. That silence should be much more concerning to all of us than it apparently is.
What's Actually Happening Here
This isn't really a story about AI. It's a story about a very old human habit finding a new outlet. Researchers call it automation bias: the tendency to trust an automated system's output and stop scrutinizing it, especially once the system has been right a few times in a row.1 It predates generative AI by decades. Pilots have missed instrument failures because autopilot "seemed to have it handled." Doctors have gone along with a flawed algorithm's diagnosis rather than trust their own read of a patient in front of them.2 In 2012, a trading firm lost $440 million in forty-five minutes because a program was allowed to run without anyone questioning its output in real time.3
What's changed is the interface. A dashboard error is easy to treat with suspicion. A chatbot that writes in full sentences, sounds confident, and answers back conversationally is not. The more fluent and human the output feels, the less we tend to check it. That's not a character flaw. It's how our brains are built to save energy, and AI tools are exceptionally good at triggering that shortcut.
What the Data Actually Shows
A widely cited 2025 study out of the Swiss Business School surveyed and interviewed 666 people across age groups and education levels.4 The finding: heavier AI use correlated with lower critical thinking scores, and the mechanism was cognitive offloading, meaning people were handing over not just tasks but the thinking behind the tasks. Younger respondents showed the steepest drop. Education level was the strongest protective factor, which tells you something important: this isn't about avoiding the tools. It's about how much independent reasoning you bring to them before you ask.
MIT Media Lab ran a smaller but sharper study in 2025, wiring up 54 people with an EEG while they wrote essays with and without AI assistance.5 The group that leaned on an LLM showed measurably lower cognitive engagement while writing, and when that same group was later asked to write without any AI help, their performance dropped compared to people who'd never used the tool at all. The researchers called it cognitive debt. You borrow ease now and pay it back later, with interest, in the form of a skill that's quietly atrophied.
None of this means AI makes people worse at thinking, full stop. In a follow-up experiment, the same lead researcher had people write persuasive essays under three conditions: no AI, unstructured ChatGPT use, and ChatGPT use with a defined critical thinking framework layered on top.6 Structure changed everything. Used deliberately, AI can sharpen reasoning. Used as a replacement for it, it flattens it.
Where This Gets Too Easy
I want to push back on my own argument here, because it's tempting to write this post as a clean morality tale: AI bad, human judgment good. That's not true and it's not useful.
Most of the people I hear saying "Claude told me to" aren't lazy or incurious. They're busy executives, operators, and specialists who are drowning in decisions and grateful for anything that reduces the volume. The instinct to offload isn't a moral failure. It's a rational response to too much to do and too little time to do it in. If anything, the tools have gotten good enough that trusting them by default now feels efficient rather than reckless, which is exactly what makes the bias so hard to notice in yourself. You don't feel like you're being careless. You feel like you're being productive.
I'd also push back on the idea that "verify everything" is realistic advice. Nobody has the bandwidth to fact-check every output from every tool for every task. The honest answer isn't universal skepticism. It's knowing which decisions deserve your full scrutiny and which ones genuinely don't matter enough to slow down for.
A Better Way to Think About Where AI Belongs
The useful question isn't "should I trust AI." It's "what kind of task is this."
Low-stakes, reversible, high-volume work is where AI earns its keep with almost no downside. First drafts of routine emails. Summarizing a long document you'll still skim yourself. Reformatting data. Generating options you'll evaluate anyway. If it's wrong, the cost of being wrong is small and easy to fix.
Anything involving judgment calls with real consequences is where you need to stay the decision-maker, not the approver. Strategic direction. People decisions. Anything where the AI is missing context it can't possibly have: the politics in your building, the history with a specific client, the thing your gut is telling you that you haven't put into words yet. An AI model has no stake in the outcome and no accountability for it. You do.
Anywhere you're tempted to skip your own expertise because the answer sounded confident is the actual danger zone. Confidence is not the same as correctness, and language models are trained to produce fluent, plausible-sounding text whether or not the underlying reasoning holds up. Treat a confident-sounding answer the way you'd treat a confident-sounding colleague who's new to the company: worth hearing, not worth outsourcing your judgment to.
A simple gut check I use with my own consulting work: if I wouldn't accept "an AI told me to" as a defensible answer in front of a client or a board, I don't accept it from myself in private either. The tool can generate the option. I still have to own the reasoning.
"Claude told me to" is not a thought. It's the absence of one.
The concerning part that we need to name, because it's the actual habit we need to break, and the sentiment we need to collectively challenge. "Claude told me to" is not a stopping point in a decision. It's not a conclusion. It's a citation with no argument behind it. The moment that sentence shows up in your head, or worse, out loud in a meeting, it means you've stopped reasoning and started reciting.
A model has no idea who's sitting across the table from you. It doesn't know that your CFO reads every recommendation looking for the catch, or that your board has been burned by a vendor before and will read any AI-flavored language as a red flag, or that the employee you're about to have a hard conversation with is dealing with something at home that changes how you should open it. It cannot stand in someone else's shoes, because it has no shoes and it has never met the person. You can. That's the entire value you bring that the tool doesn't.
So the discipline isn't "don't use AI." It's: before you carry a recommendation into a room, ask whose lens is missing from it. Whose incentives, history, or context didn't make it into the prompt. An AI's answer is a starting draft of a perspective, usually your own, reflected back at you with more confidence than you started with. It isn't a second opinion until you've actually gone and gotten one, from a person who sees the situation differently than you do.
A Small Framework for the "Because Claude Said So" Moment
The next time you catch yourself about to lean on an AI answer as if it settles something, run it through four questions before you act on it or repeat it to anyone else.
Whose lens is missing? Name a specific person or group who would read this differently than the AI did, and ask what they'd flag. Not "someone might disagree." A specific stakeholder, with a specific likely objection.
What did I not tell it? AI answers are only as good as the context you fed in. List the things you know that never made it into the prompt: office politics, history with this client, a constraint you didn't think to mention. If that list is longer than a sentence, the answer is incomplete by definition.
Would I defend this without the AI's name attached? Strip "Claude said" from the sentence and see if the reasoning still stands on its own. If the only support for the idea is where it came from, you don't have an idea yet. You have an unexamined output.
What's the cost of being wrong here, and who pays it? This sorts the stakes. Low-cost, reversible, your call alone to live with: move fast, the AI draft is probably fine as a starting point. High-cost, hard to reverse, or someone else bears the consequence: that's exactly when you slow down and do the human work the tool can't do for you.
To test this theory, I ran this piece through its own checklist before sending it to you. Whose lens is missing: the person who's genuinely time-starved and using AI defensively, not lazily, which is most of the executives I know, including me some weeks. What didn't make it into my own thinking at first: that "verify everything" is advice from someone with the luxury of time, not a realistic standard for someone running a team. Would the argument hold without citing the studies: mostly, yes, because the pattern matches what most people already sense in themselves, the research just confirms it. What's the cost of getting this wrong: low. Worst case, someone reads a blog post and disagrees with me. That's exactly the kind of low-stakes work where I was comfortable using AI to help me draft and organize, while keeping the argument itself mine.
The Actual Skill to Build
The people who will do the best work over the next decade won't be the ones who avoid AI, and they won't be the ones who defer to it either. They'll be the ones who get sharper at asking it good questions, checking its blind spots, and knowing exactly which 20 percent of their thinking they're not willing to hand over.
That's not a technical skill. It's the same critical thinking that's always separated strong operators from average ones. AI just raises the stakes on whether you actually have it.
Joy Wilson, Jelena Partners
Footnotes
Automation bias is documented across aviation, medicine, and finance as the tendency to over-trust automated outputs and discount contradictory information. See CSET, "AI Safety and Automation Bias" (Georgetown, 2024) and TechTarget, "What is Automation Bias?" (Informa, 2025). ↩
On automation bias in medical diagnostics specifically, see the International AI Safety Report 2026, which cites evidence of automation bias in AI-assisted clinical decision-making. ↩
Knight Capital's 2012 trading loss of $440 million in 45 minutes due to an automated system running without adequate human oversight, as cited in Forbes, "Automation Bias: What It Is And How To Overcome It" (2024). ↩
Gerlich, M. (2025). "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking." Societies, 15(1), 6. Mixed-methods study of 666 participants; found a significant negative correlation between frequent AI tool use and critical thinking scores, mediated by cognitive offloading. ↩
Kosmyna, N. et al. (2025), MIT Media Lab. EEG study (n=54) on essay writing with and without LLM assistance, finding reduced cognitive engagement during AI-assisted writing and reduced performance on a subsequent unaided task, termed "cognitive debt." ↩
Follow-up experiment by Gerlich, reported in APA Monitor, "How AI is reshaping human skills and thinking" (2026), comparing essay writing under no-AI, unstructured-AI, and structured-AI-with-critical-thinking-framework conditions. ↩
