by admin | Oct 8, 2026 | Research & Productivity
Picture this: it’s late, you’re finishing a paper, and an AI assistant just handed you three perfect-sounding sources to back up your argument. You drop them into your reference list and move on. A week later, your supervisor can’t find one of those papers anywhere, because it doesn’t exist.
That isn’t a rare glitch. It’s one of the most common and least talked-about risks of using AI for research, and it even caught out a well-known misinformation expert in a real court case. AI assistants like ChatGPT, Gemini, and Claude are genuinely useful for study and research, but they also “hallucinate”: they sometimes invent facts, authors, and entire papers that sound completely real. This guide shows you how to fact-check AI research answers before you cite them, using a method that takes minutes, not hours.
Why AI Tools Sometimes Invent Sources
AI hallucination isn’t random bad luck, it’s built into how these models are trained. OpenAI has explained that language models learn by predicting the next word in huge amounts of text that has no built-in “true or false” label attached to it. Common patterns, like grammar and spelling, improve a lot as models get bigger. But specific, low-frequency facts, like the exact volume number of a journal article, often can’t be learned reliably, so the model fills the gap with something plausible-sounding instead.
OpenAI also points out that the way models are tested makes this worse: standard accuracy scoring rewards a confident guess over an honest “I don’t know,” so models learn to guess rather than admit uncertainty. OpenAI says newer models hallucinate less, especially when reasoning carefully, but the problem hasn’t disappeared.
A Real Case: When a Misinformation Expert Got Caught by AI
This isn’t just a theoretical risk. In late 2024, Stanford researcher Jeff Hancock, founder of the university’s Social Media Lab, submitted a court declaration supporting a Minnesota law on AI deepfakes. According to reporting in the Stanford Daily, Hancock later admitted he had used ChatGPT to help organize his citations, and the tool invented two references that didn’t exist and misattributed a third to the wrong authors. He said his core arguments were still backed by real research, but the fabricated citations gave opposing lawyers grounds to challenge the filing. If a misinformation researcher can miss fake AI citations in a federal court filing, the rest of us should assume we can too.
From my own experience working with websites, cybersecurity, and online tools, I’ve seen the same pattern in AI-suggested code and security fixes: the answer sounds confident and specific, but it hasn’t actually been checked against a real, working source. Confidence is not the same as accuracy.
How to Fact-Check AI Research Answers in 5 Simple Steps
A university library guide on checking AI-generated citations lays out a short process that works well for any AI research answer, not just formal references:
- Check that the source has the right details. A journal article needs an author, title, journal name, volume, and issue. A book needs a publisher. If key details are missing, treat it as unverified.
- Watch for red flags. Broken or dead links, a citation format that doesn’t match standard styles, or vague phrasing like “a 2023 study found” with no named source are all warning signs.
- Search for it yourself. Paste the exact title in quotation marks into Google Scholar or a regular search engine. If nothing matches, the source is probably invented.
- Match every detail. Confirm the author names, publication date, journal, and volume or issue number against what you actually find, not just the title.
- When you’re still unsure, ask a librarian or switch tools. Librarians can verify sources and track down hard-to-find papers, and some AI tools are far less likely to invent sources in the first place.
Quick tip: paste the exact title of any AI-suggested source into Google Scholar in quotation marks. If nothing comes up, treat the citation as fake until you can prove otherwise.
Tools That Make Verification Easier
Not all AI research tools carry the same risk. General chatbots generate text freely, which is exactly how fabricated citations slip in. Source-grounded tools work differently: they search real, indexed material and show you where an answer came from, so you can check it in seconds.
Our guide to AI research tools like NotebookLM and Elicit covers how these grounded tools cite their sources directly, and Google Scholar Labs builds its AI answers from indexed academic papers rather than generating them freely. For the bigger picture on using AI without losing your own judgement, read AI and critical thinking, and for a broader starting point on using AI in your coursework, see how AI can help with research and productivity.
Common Questions
Do all AI tools hallucinate citations?
Most general-purpose chatbots can, since they generate text from patterns rather than a verified database. The risk is lower with tools that search and cite real indexed sources, but it’s never zero.
Is it still safe to use AI to find research sources?
Yes, as long as you verify what it gives you. AI is excellent for discovering leads and summarizing papers. The problem isn’t using AI, it’s citing its output without checking it first.
What’s the fastest way to check a citation?
Copy the exact title into Google Scholar in quotation marks. A real paper usually appears in the first few results. If it doesn’t, dig further before you trust it.
Final Takeaway
AI can speed up your research dramatically, but it can’t replace the five minutes it takes to check whether a source is real. Build that check into your routine every time, and you get the speed of AI without the risk of citing a paper that was never written in the first place.
by admin | Oct 1, 2026 | Research & Productivity
If you have ever typed a research question into Google Scholar and gotten back a wall of a hundred papers with no idea where to start, you are not alone. That gap between “I have a question” and “I have the right ten papers to read” eats up more time than almost anything else in research work. Google has been quietly trying to close that gap with a tool called Google Scholar Labs, and it just got a big upgrade.
If you are a student, a grad researcher, or anyone who reads academic papers for a living, this is worth five minutes of your time.
What Is Google Scholar Labs?
Scholar Labs is an AI search mode built into Google Scholar. Instead of just matching keywords, it reads your actual question, works out the different angles hiding inside it, and then searches across Scholar to find papers that address those angles together.
Google gives a simple example: if you ask about caffeine’s effect on short-term memory, a keyword search gives you papers on caffeine and separate papers on memory. Scholar Labs tries to surface papers that connect the two, and for each result it explains how that paper actually answers your question.
It first launched in November 2025 to a small group of logged-in users. You can try it directly at scholar.google.com/scholar_labs/search.
What Changed in the Latest Update
In June 2026, Google rolled out a significant update to Scholar Labs, and the numbers are the headline:
- Results return about 10 times faster than the original version
- The tool scans up to 3 times more papers per search
- Logged-in users can now run 10 times more searches per day
Just as important, access widened from a small test group to all logged-in Google Scholar users. The update launched in English first, with other languages planned.
You can also ask follow-up questions once you get results, which turns a single search into something closer to a conversation with your reading list. That matters if your first question was too broad or you need to chase a specific detail, like a sample size or a methodology, across several papers.
How to Use It for Your Own Research
Getting value out of Scholar Labs comes down to how you phrase the question, not just what topic you search.
A plain keyword search still works fine when you already know the paper or author you want. Scholar Labs earns its keep when your question has more than one moving part. Try something closer to “how does remote work affect employee mental health across different industries” instead of just “remote work mental health.” The more structure you put into the question, the more structure you get back in the answer.
A few practical ways to use it:
- Early-stage literature review, to get a first map of a topic before you commit to a direction
- Checking whether your research question has already been well covered, before you propose it
- Verifying a claim or citation while you are drafting, by asking the specific question your sentence depends on
Important tip: treat Scholar Labs as a faster way to find candidate papers, not as a replacement for reading them. It is still an experimental AI feature, and the usual rules apply: check the actual paper, check the date, and check who funded or published it before you cite anything.
From working on websites and digital tools for a few years now, the tools that stick are rarely the flashiest ones. They are the ones that remove one annoying step from a task you already do every week. Scholar Labs fits that description for anyone doing a literature review: it does not replace your judgment, it just gets you to the reading list faster.
Where This Fits With Other AI Research Tools
Scholar Labs is one piece of a bigger shift toward AI-assisted research. If you want the wider picture, What Is AI Deep Research? explains how tools like this fit into a full research workflow, and OpenAI Built an AI Research Intern covers a similar idea from a different company.
If you are already using reference managers, Zotero’s new read-aloud AI feature is a good companion tool once Scholar Labs has helped you find the papers worth reading closely. And once you have your reading list, this guide to summarizing research papers with AI walks through getting through them faster without skipping the important parts.
For the official details straight from Google, see the Google Scholar blog’s launch post and the June 2026 update announcement.
Common Questions
Is Google Scholar Labs free to use?
Yes. It is available to any logged-in Google account holder through Google Scholar, with no separate subscription.
Does it replace Google Scholar’s normal search?
No. Scholar Labs is a separate AI mode you access through its own page. Regular Google Scholar search still works exactly as before.
Can I trust the papers it recommends?
Treat it the same way you would treat any search tool’s results: a useful starting point, not a final answer. Always check the original paper before citing it.
Is it available outside English?
The June 2026 update launched in English first. Google has said other languages are planned, but no firm dates have been announced.
Final Takeaway
Google Scholar Labs will not write your literature review for you, and it should not. What it does is cut down the time between asking a real research question and finding papers that actually address it, which is where most of us lose hours we do not have. If you do any kind of academic or professional research, it is worth bookmarking and testing on your next real question.
by admin | Sep 17, 2026 | Research & Productivity
OpenAI just told the world something that sounds like science fiction: its own research team now gets more raw work done by AI agents than by the humans running them. On September 6, 2026, the company published a detailed look inside its research organization, and the number at the center of it is striking. For every one workday a human researcher puts in, AI coding agents now contribute the equivalent of 3.1 workdays.
That’s not a marketing claim. It’s OpenAI measuring its own internal usage and publishing the methodology behind it. And whether or not you touch a line of code for a living, there’s a real lesson in here for how you research, study, or get work done with AI. From my own experience building websites and testing tools for online projects, the pattern OpenAI describes, more agents running in parallel, handling the boring middle steps, is exactly the shift a lot of us are quietly living through already.
What OpenAI Actually Announced
OpenAI said it has reached a goal it set last year: building what it calls an “automated research intern” by September 2026. By that, the company means a system that can carry out well-defined research tasks under human direction, including tasks that would normally take a skilled researcher several days to finish. It is not an autonomous scientist working alone. People still decide what to work on, judge the results, and choose whether to scale or shut down a project.
The scale of adoption inside the company is what stands out. By mid-August 2026, the median OpenAI researcher was using more than $600 a day of AI inference just to run coding agents, and the top 10% of researchers were burning through more than $7,000 a day. Many researchers now run four or more agents at once, each handling a different part of an experiment.
Why “3.1 Agent-Workdays” Actually Matters
It’s easy to skim past a statistic like 3.1 agent-workdays per human workday, so it’s worth slowing down on what it really describes. OpenAI isn’t just using AI to write code faster. Its researchers are delegating troubleshooting, running experiments, and monitoring training runs to agents, freeing themselves up for the parts of research that still need human judgment: deciding what’s worth testing, spotting when something has gone wrong, and communicating findings.
Interestingly, OpenAI also reported that internal teams who used to hold “office hours” to help researchers debug their experiments have seen attendance drop so much that some have stopped holding them altogether. People are asking their AI agents instead of asking each other. That’s a genuine shift in how technical work gets unstuck, not just a productivity buzzword.
This Isn’t Just an OpenAI Story
You don’t need a research lab to benefit from the same underlying idea. The World Economic Forum has been tracking similar productivity gains across ordinary workplaces throughout 2026, with economists pointing to AI’s biggest impact showing up in tasks that involve research, drafting, and analysis, exactly the kind of work students, writers, and knowledge workers do every day. The tools are different (you’re probably not running four coding agents at once), but the underlying habit, letting AI handle repetitive research steps while you focus on judgment calls, applies just as well to a term paper as it does to a frontier AI lab.
How to Borrow This Idea for Your Own Research
You don’t need OpenAI’s budget to apply the same principle. A few practical starting points:
- Use an AI agent to handle the repetitive first pass of a task, like summarizing ten sources, then do the judgment work yourself: which sources actually matter, and why.
- Understand what AI agents can and can’t do before you rely on one for something important. They’re strong at defined, bounded tasks and weak at open-ended judgment.
- Lean on tools built for research specifically, rather than a general chatbot, when you’re working through papers or long documents. Our guide to AI research tools like NotebookLM and Elicit covers a few worth trying.
- If your work involves digging through long PDFs or reports, tools that let you chat with a PDF using AI can save hours that used to go into manual skimming.
Tip: treat AI research agents like an intern, not an expert. Give them a specific, bounded task and check their work, rather than trusting an open-ended request to “figure this out for me.”
If you want to go a level deeper, our explainer on AI deep research tools walks through how these longer, multi-step research agents actually work behind the scenes, which is useful context for understanding what OpenAI’s researchers are really delegating.
Common Questions
Does this mean AI is replacing researchers at OpenAI?
No. OpenAI is explicit that people still set research priorities, judge which results matter, and decide whether to scale or pause a project. Agents handle defined tasks under human direction, not the whole job.
Can regular students or professionals use the same kind of AI research agent?
Yes, in a smaller form. Tools built for research and document analysis, rather than general chat, are the closest equivalent available to everyday users right now.
Is a 3.1x productivity number realistic outside a frontier AI lab?
Probably not directly. OpenAI’s figure reflects a company with enormous compute budgets and custom internal tools. But the broader pattern, using AI to handle repetitive research legwork, is realistic for anyone, just at a smaller scale.
Final Takeaway
OpenAI’s “research intern” milestone is a useful reminder that AI’s biggest impact right now isn’t flashy chatbot demos, it’s quietly reshaping how research and analysis actually get done, one delegated task at a time. You don’t need a research lab to use that idea. Pick one repetitive part of your own research or work, hand it to an AI tool built for the job, and spend your saved time on the parts that actually need your judgment.
by admin | Sep 10, 2026 | Research & Productivity
If you spend your evenings buried in PDFs for a literature review, you already know the real problem is not finding papers. It is finding the time and the eyesight to actually get through them. Zotero, the free reference manager millions of students and researchers already use, just shipped something aimed straight at that problem.
The tool released Zotero 9 this month, and the headline feature is called Read Aloud. It uses natural, AI-generated voices to read your PDFs, ebooks, and saved web pages back to you. It is a small addition on paper, but if you use Zotero for your research and productivity routine, it changes how you can work through a stack of papers.
What is the Zotero Read Aloud feature?
According to Zotero’s own release notes, Read Aloud reads your documents to you in “high-quality, natural-sounding voices,” and it works across PDFs, EPUBs, and webpage snapshots you have saved into your library. You start it with a headphones button in the reader toolbar, and from there you can skip forward or backward by sentence or paragraph, or jump straight to a section by clicking in the margin.
Your place in the document is saved and synced, so you can start listening on your laptop during breakfast and pick up on the same sentence later on another device. There is also an “Annotate Sentence” shortcut that highlights or underlines whatever line you just heard, which is genuinely useful if you are the type who highlights everything on a first pass anyway.
Free voices vs premium voices
Zotero offers two tiers of what it calls Zotero Voices. Standard voices run on Zotero’s own servers and are available to every account, with a monthly free allowance and unlimited use for paid Zotero Storage subscribers. Premium voices are processed by outside text-to-speech providers, sound noticeably more natural, and support more languages, with a smaller free allowance included for everyone to try them.
Read Aloud needs an internet connection and a free Zotero account to use these AI voices. If you would rather stay offline, you can still use your computer’s built-in text-to-speech voices, though Zotero is upfront that the quality is “significantly degraded” compared with the online voices. For now, the feature lives in the desktop app only, with mobile support planned.
Tip: Try Read Aloud on a paper you have already read once. Listening to a second pass while you glance at figures and tables is a fast way to catch details you skimmed the first time, without opening a new tab or tool.
Other Zotero 9 changes worth knowing
Read Aloud is the headline, but a few smaller changes matter for anyone doing serious research work:
- A new “Recently Read” collection shows the items you opened most recently, so you stop hunting for that one PDF you were reading yesterday.
- You can now insert PDF annotations directly into a Word or Google Docs document, with active citations attached, instead of copying them into a Zotero note first.
- Login now happens through your browser rather than by typing your password into the app, which also opens the door to two-factor authentication.
None of this replaces the AI research tools built on top of Zotero, either. Zotero has no built-in chatbot or paper-summarizing AI, and the project says as much in its own plugin documentation: community plugins remain the way people add that kind of feature. If you want an AI assistant that can chat about your library or help you summarize research papers with AI, that still comes from a third-party plugin you install yourself, not from Zotero directly, so only install plugins from developers you trust.
What this means for your research routine
From my own experience working on websites and digital tools, the features that actually stick are the boring, practical ones, not the flashiest ones. Read Aloud fits that pattern. It will not write your literature review for you, and it will not fix a bad paper. What it does is turn dead time, a commute, a walk, folding laundry, into reading time you would otherwise lose.
If you are working through a long reading list, that adds up. Pair it with the habit of double-checking anything an AI tool tells you: text-to-speech is low-risk, but AI summarizers and chat assistants can still invent a citation that does not exist, so always confirm sources against the original paper before you rely on them in your own writing.
Common Questions
Is Zotero Read Aloud free to use?
Yes. Every Zotero account gets some free Standard voice minutes each month, and a smaller free allowance of Premium voices to try. You do not need a paid subscription to use the feature, though heavy use of Premium voices may need a Zotero Storage subscription over time.
Does Read Aloud work on mobile?
Not yet. Zotero says it is currently available only in the desktop app, with iOS and Android support planned for a future update.
Is Zotero’s AI voice the same as an AI chatbot for research?
No. Read Aloud only converts text to speech. It does not summarize, answer questions, or generate citations. Zotero’s own documentation says AI chat and summary features still come from third-party community plugins, not from Zotero itself.
Do I need to update Zotero to get this feature?
Yes. Read Aloud shipped with Zotero 9, so you will need to update your app through Help > Check for Updates, or download the newest version, to see it.
Final takeaway
Zotero Read Aloud will not change how you cite a source or how you argue a thesis. What it does is quietly give you back reading time you were losing, using AI voices good enough to actually listen to. If you already use Zotero, updating and trying it costs nothing but a few minutes. If you are still managing your references in a messy folder of PDFs, this is as good a reason as any to finally set Zotero up properly.
by admin | Sep 3, 2026 | Research & Productivity
It is 11pm, your draft is finished, and nobody you trust is awake to read it. So you paste the whole thing into a chatbot and type the four words most of us have typed at some point: “can you improve this?” A few seconds later you get back something smoother, tidier, and slightly stranger. It reads well. It just does not sound like you anymore.
There is a better way to use these tools, and it comes down to one small change. Ask for feedback instead of a rewrite. Good AI feedback on your writing works a bit like a patient reader who never gets tired: it can tell you where your argument wobbles, where a paragraph loses the thread, and what a reader might still be confused about after finishing. Then you decide what to change, in your own words.
This guide covers what to ask for, the prompts worth copying, what to ignore, and the rules you need to check first.
Why AI feedback on your writing beats “make this better”
When you ask an AI tool to improve a text, it will do exactly that, and it will usually do it by replacing your sentences with its own. You end up with a document you did not write and cannot fully defend if somebody asks you why you made a particular choice.
The Writing Center at the University of North Carolina at Chapel Hill puts the principle neatly in its handout on generative AI in academic writing: when instructors allow these tools, they assume the tools will help you think and write, not think or write for you. Feedback keeps you in the driving seat. A rewrite quietly takes the wheel.
There is a practical benefit too. Feedback teaches you something you can reuse next time. A rewrite teaches you nothing except how to press the button again.
Check the rules before you paste anything in
This step takes two minutes and saves a lot of trouble. Course policies differ wildly, sometimes between two modules in the same department, so read the assignment brief before you open a chat window.
Elon University’s Center for Writing Excellence asks students to check the policy for each assignment and disclose how they used AI when they bring a draft in for help, and it advises keeping a record of your prompts so you can explain your process later. That last part is good advice for anyone, student or not. If you cannot describe how a tool helped you, you probably leaned on it too hard.
Privacy matters just as much. From my own work building websites and handling client material, the rule I stick to is simple: if a document would embarrass me or somebody else in a leak, it does not go into a public chatbot. Unpublished research, confidential reports, anything with personal data in it, and anything covered by an employer agreement all belong in that category. Our guide to using AI safely and protecting your privacy walks through the settings that reduce the risk.
Give the tool the brief, not just the text
Most disappointing feedback comes from a missing brief. The tool cannot judge whether your essay answers the question if it has never seen the question.
Before you ask for comments, paste in four things: the actual task or assignment wording, the marking rubric or the criteria your reader cares about, who the audience is, and the word limit. Then add your draft. The difference in quality is dramatic, because the tool now has a standard to measure your work against instead of a vague idea of “good writing”.
The study skills team at the University of Roehampton also suggests breaking your prompts down and asking for one thing at a time, listing spelling and grammar, structure, argumentation and reasoning, and flow as separate jobs. Ask for all four at once and you get a shallow paragraph on each. Ask for one and you get something you can act on. If prompt wording is where you usually get stuck, our guide on writing better AI prompts covers the basics.
Five feedback prompts worth copying
These are written so the tool comments rather than rewrites. Notice that none of them say “fix” or “improve”.
- Structure: “Here is my assignment brief and my draft. Do not rewrite anything. List the main point of each paragraph in one line, in order, so I can see whether my structure makes sense.”
- Argument: “Act as a sceptical reader. What is my main claim, and where is the evidence weakest or missing? Quote the exact sentences that worry you.”
- Clarity: “Mark any sentence a tired reader would need to read twice, and say in a few words why. Do not suggest replacements yet.”
- Gaps: “What questions would a reader still have after finishing this? What did I assume they already know?”
- Voice check: “Describe the tone of this draft in three phrases. Does it match a piece written for [your audience]?”
The quoting instruction in the second prompt matters more than it looks. It forces the tool to point at real sentences in your text instead of producing generic advice that would fit any essay ever written.
Important tip: ask for the reasoning behind every suggestion, then decide for yourself. If the tool cannot explain why a change helps, treat the suggestion as an opinion and move on.
What to do with feedback you disagree with
You are allowed to say no. Some AI suggestions are wrong for academic or professional writing in specific, predictable ways, and the Roehampton guide flags two of them. The first is a push towards first person and direct personal pronouns, which can make writing clearer but can also weaken the tone your marker expects. The second is a habit of strengthening hedged language. Phrases like “the evidence suggests” exist for a reason, and turning them into confident claims can push you past what your data actually supports.
There is a useful exercise in the UNC handout for spotting this. Run a paragraph through the tool, then put the original and the edited version side by side and look at what changed. Do you agree with the changes? Did the edit make the meaning clearer, or did it just sand off the parts that sounded like you? Doing this a few times teaches you a lot about your own habits, which is the real prize here.
Where AI feedback goes wrong
Three failure modes come up again and again, and all three are worth watching for.
- Confident wrong facts. If you ask a tool to check a claim or suggest supporting sources, it may invent them. UNC’s handout warns that citations can look perfectly formatted and still be fictional. We covered how to catch this in our post on checking AI citations.
- Flattening. Rounds of “smoother” edits tend to converge on the same neutral house style. Two or three passes and your writing sounds like everyone else’s.
- Doing the thinking for you. The most comfortable failure mode. Accepting feedback without judging it is still outsourcing the work, which is exactly what our piece on AI and critical thinking digs into.
A simple routine that takes about twenty minutes
Finish your draft first, without any AI in the room. Then paste in the brief and the rubric, and run the structure prompt. Fix the structure yourself. Run the argument prompt next, and note which weaknesses you accept. Fix those yourself too. Only at the end, once the thinking is settled, ask about clarity and grammar at sentence level.
Elon’s guidance makes the same point about timing: used carefully, AI helps most at the beginning of the writing process, during brainstorming, and at the end, during proofreading. The middle, where you work out what you actually think, is yours. Keeping that middle for yourself is also the honest answer to the academic integrity question, which we look at in more detail in our guide to using AI tools as a student without cheating.
Common Questions
Is asking AI for feedback on my writing cheating?
It depends entirely on your institution and the specific assignment, which is why you check the brief first. Many universities treat feedback and proofreading differently from generating text, but policies vary by module and by tutor. If the brief is unclear, ask before you submit, not after.
Which AI tool gives the best writing feedback?
The prompt matters far more than the tool. The main free chat assistants all handle this task reasonably well. What changes the outcome is whether you supplied the brief and the rubric, and whether you asked for comments rather than a rewrite.
Should I paste my whole document in at once?
For structure feedback, yes, because the tool needs to see the shape of the whole piece. For sentence level work, go section by section. Feedback on a short passage is almost always more specific and more useful than feedback on twenty pages.
How do I stop AI feedback from changing my voice?
Never paste the suggested version back into your document. Read the comment, close the tab, and make the change in your own words. If a suggestion is good, you will remember it well enough to write it yourself.
Final takeaway
Used as an editor, AI takes over your writing. Used as a reader, it makes you better at it. Give it the brief, ask for comments instead of corrections, check the rules for your course or your workplace, and make every change yourself. Try one prompt from the list above on your next draft and see how much of the feedback you actually agree with. That disagreement is where your own judgement lives, and it is worth protecting.
by admin | Aug 27, 2026 | Research & Productivity
A few months ago I caught myself doing something odd. I had used AI to draft a page for one of my sites, published it, and a week later I could not have explained the argument in it without opening the file. The work was done. The thinking, apparently, had not stuck.
That feeling has a name now, and researchers have started measuring it. The question of AI and critical thinking is no longer just a worry people raise at dinner. Three studies, two of them peer reviewed, have looked at what happens inside the heads of people who lean on AI, and the results are worth knowing before you build your next work habit around a chatbot.
The short version: the risk is real, it is not automatic, and the difference comes down to how you use the tool rather than whether you use it.
What the research says about AI and critical thinking
The most useful study of working adults came from Microsoft Research and Carnegie Mellon University, presented at CHI 2025. The team surveyed 319 knowledge workers and collected 936 first-hand examples of generative AI used in real work tasks.
The headline finding is a confidence effect, and it cuts both ways. People with higher confidence in the AI did less critical thinking. People with higher confidence in their own ability at the task did more. Same tool, opposite outcomes, depending on who was holding it.
The researchers also found that the thinking does not disappear so much as move. It shifts away from producing an answer and toward verifying information, stitching AI output together, and managing the task overall. That is still real cognitive work. It is just different work, and it only happens if you choose to do it. You can read the full paper on the Microsoft Research site.
The MIT essay study, and what it does not prove
The study that got the loudest headlines came from the MIT Media Lab. Researchers put EEG caps on participants and asked them to write essays under three conditions: using an LLM, using a search engine, or using nothing at all.
Brain connectivity was strongest in the no-tools group, moderate in the search group, and weakest in the LLM group. The LLM writers also reported the lowest sense of ownership over their essays, and struggled to accurately quote work they had produced minutes earlier. The authors called the pattern “cognitive debt”. The full paper is on arXiv.
Now the honest part, because this study has been badly oversold online. It involved 54 participants across the first three sessions and only 18 in the fourth, and it looked at one narrow task. It is still a preprint. The authors revised it in December 2025, but it has not been through peer review. That is a signal worth taking seriously, not a verdict on your brain.
The order you work in seems to matter
Buried in the same MIT study is the most practical finding of the lot, and almost nobody reported it.
In the final session the researchers swapped the groups around. People who had been writing unaided and were then given an LLM showed higher memory recall and stronger brain activation, closer to the search-engine group. People who had been leaning on the LLM and then had it taken away showed reduced connectivity and looked under-engaged.
Read that again, because it is the whole practical lesson. Thinking first and bringing AI in second looked healthy. Starting with AI and trying to think afterwards did not.
The 2026 study that names the difference
The newest piece of this puzzle is also the most useful, and it arrived in July 2026 in the peer-reviewed journal Frontiers in Psychology. Researchers surveyed 589 university students and early-career knowledge workers across three rounds spaced out over time, and they split AI use into two modes instead of treating it as one thing.
Dependent offloading is handing over the actual thinking: accepting the output with little examination, letting the AI structure your ideas, treating what it produces as the finished article. Autonomous offloading is using the same tool as a scaffold: taking the output as a starting point, comparing it against your own reasoning, and keeping ownership of the final result.
The two modes pulled in opposite directions. Dependent use was linked to handing over cognitive agency and to lower motivation, and through those to weaker self-rated judgement, creativity and depth of thinking. Autonomous use was linked to higher motivation and better self-rated outcomes. Same tools, same tasks, opposite results based only on how people engaged.
Then comes the finding that should change how you work. Both modes produced comparable immediate benefits. The session feels equally productive either way, which means unhelpful AI use is very hard to spot from experience alone. Nothing in the moment tells you which mode you are in. You can read the full study on the Frontiers in Psychology site.
Keep it in perspective. It asks people to rate their own thinking rather than testing it, so it shows associations, not proof of cause. But it is peer reviewed, far larger than the MIT study, and it names the distinction the other two keep circling.
Five habits that protect your thinking
- Write your rough version first. Even five messy bullet points before you open the chat window changes the whole session. You arrive with a position instead of asking to be given one.
- Ask the AI to question you rather than answer you. Try “argue against this” or “what am I missing here” instead of “write this for me”. You get pushback rather than a finished product you never examined.
- Verify anything you would be embarrassed to get wrong. Names, numbers, dates, citations, legal or medical claims. AI systems produce confident wrong answers regularly, which is why our guide to AI hallucinations is one of the most useful things to read before you trust an output.
- Explain the result out loud in your own words. If you cannot, you have not understood it, and you will not be able to defend it in a meeting or a viva.
- Notice when a task feels low-stakes. The Microsoft researchers found that people review AI output far less carefully when they judge a task to be unimportant. Low-stakes tasks are exactly where sloppy errors slip through into public work.
Important tip: if you cannot explain the answer to another person in your own words, you have not learned it. You have only borrowed it.
When letting AI do the work is completely fine
None of this means you should feel guilty every time you open a chatbot. Offloading is only a problem when you are offloading something you actually needed to learn.
Reformatting a messy list, fixing spelling, converting a table, writing boilerplate you have written a hundred times, summarising a document you are only skimming for one fact. Hand all of that over without a second thought. From my own experience running websites and online tools, that category is where AI saves genuine hours, and none of those hours were making me smarter.
The line is simple. If the task is the learning, do it yourself first. If the task is friction around the learning, automate it. That distinction is also why our post on using AI tools without cheating keeps coming back to the same test.
What this means for students and researchers
If you are studying, the stakes are higher, because the entire point of the work is to build something in your head that stays there. A summary you did not read leaves nothing behind. Our plan for using AI to study for exams is built around this: generate questions, not answers.
Microsoft Research has since argued for designing AI as a “tool for thought” rather than an assistant that hands you finished work, warning that people risk becoming validators of machine output rather than authors of their own. You can read their argument here. Until the tools are built that way, the burden of using them well sits with you.
Common Questions
Does using AI actually make you less intelligent?
No study has shown that. What the research shows is reduced mental effort during AI-assisted tasks and weaker recall of AI-assisted work. Those are measurable short-term effects on specific tasks, not evidence of permanent change.
Is it better to just avoid AI for serious work?
Avoiding it entirely costs you speed and puts you behind on a skill employers now ask for. The evidence points toward using it deliberately rather than avoiding it, which is also the theme of our guide to the AI skills that matter most for future jobs.
What is cognitive offloading?
It means using something outside your head to do mental work for you. A shopping list, a calculator and a satnav all count. AI is a far broader version of the same thing, which is why where you draw the line matters more.
Final takeaway
AI and critical thinking are not enemies. The research suggests they come apart only when you hand over the part of the task that was supposed to change you. Think first, use the tool second, verify what matters, and be able to explain the result without the screen in front of you. Do that and the tool stays a tool.
