AI has made research faster and self-deception more efficient. The difference between the two is a set of habits, and none of them are complicated.
I use AI every working day: for research, building systems, organizing information, and developing the infrastructure behind this site. It has collapsed the distance between a question and a plausible first answer. That is the gift and the trap: the first answer often arrives with the tone and structure of a final one.
A model can produce a clean explanation before I have done the work required to know whether the explanation is true. It can make weak evidence feel settled, turn a vague memory into a confident claim, and furnish my assumptions with footnotes that may not support them. Research becomes faster, but so does rationalization.
The solution is not to avoid the tool. It is to assign the tool a role it can perform and keep it out of roles that require responsibility.
Decide the role before opening the tool
I get better results when I decide what kind of help I want before I prompt. AI can be an explorer that identifies avenues of inquiry, a librarian that suggests sources, a clerk that organizes notes, a critic that finds gaps, or an editor that improves structure. Those are useful roles, and they are different jobs with different failure modes.
The explorer can invent avenues that do not exist, so its output is a list of hypotheses, not findings. The librarian can recommend books that were never written, so every citation gets checked against a catalog before it enters my notes. The clerk is the most trustworthy of the group, because sorting and formatting are verifiable at a glance. The critic is only as sharp as the questions I give it. The editor is excellent at structure and rhythm and completely indifferent to whether the argument is true.
What the model is not: a witness. It is not a primary source. It is not an accountable author. And it should not become the final judge of a question simply because it can state a conclusion smoothly.
AI drafts the map. It does not become the territory.
Treat every answer as a set of leads
When a model says that a report found something, the next step is opening the report. When it summarizes a court decision, the next step is reading the decision or a reliable legal analysis. When it describes a debate, the next step is finding the strongest serious voice on each side, not merely the voices the model happened to retrieve.
The legal profession learned this in public. In 2023, two New York attorneys were sanctioned after filing a brief containing judicial opinions that a chatbot had invented, complete with case names, docket numbers, and internal citations. The embarrassing part was not that the tool fabricated cases. Tools fail. The embarrassing part was that checking the citations would have taken minutes in any legal database, and no one spent the minutes.
But the fabricated citation is actually the easy failure to catch, because it dissolves on contact with a database. The subtle errors are more dangerous than the absurd ones. A model may identify the right study, author, and date, then overstate what the study established. A survey of what workers say about productivity becomes evidence of measured productivity. A correlation becomes a cause. A recommendation in the discussion section becomes a finding. The citation looks real because it is real; the claim attached to it is the part that drifted.
That is why source verification means more than confirming that a document exists. I need to confirm that the document says what I am about to say it says.
Use a verification ladder
Not every sentence deserves the same research burden. A passing illustration may need a quick check. A claim involving law, health, money, theology, reputation, or current events needs more. For important claims, I use a simple ladder:
- Find the primary source. Locate the paper, filing, transcript, dataset, statute, product documentation, or original statement whenever possible.
- Match the claim to the source. Read enough context to know whether the evidence supports the wording, not just the topic.
- Check the date and version. AI tools, regulations, product features, and company facts change quickly. Yesterday’s accurate answer may now be stale.
- Look for a credible counter-case. Search for contrary evidence and the strongest methodological criticism.
- Keep the source. Save the document or a durable link in my notes so the published claim can be checked later.
Here is what the ladder looks like in practice. Suppose a model tells me that a federal report concluded a certain category of UAP sightings was explained by drone activity. Step one: find the actual report, not an article about the report. Step two: read the section. Often the report says something narrower, such as that drones were the assessed explanation for a subset of cases with sufficient data. Step three: check whether a newer report supersedes it. Step four: search for informed criticism of the report’s methods. Step five: file the PDF and the page number. The published sentence that survives this process is usually more modest than the one the model offered. It is also one I can defend a year later.
The ladder is slower than copying the model’s paragraph. It is still much faster than doing the entire search manually, and it preserves the part of research that matters: contact with the evidence.
Make the model argue against you
Models often take the shape of the question. If I ask for evidence supporting my idea, they are good at supplying it. If I present a theory with enthusiasm, they tend to join the enthusiasm. Agreeable assistance is comfortable, but it can turn a hunch into a closed loop.
One of the most useful prompts in my toolkit is some version of: What is the strongest case that I am wrong? I also ask which assumptions are doing the most work, what evidence would change the conclusion, and how a careful critic would describe the weakness in my sources. When a draft matters, I will run it through the model in a fresh conversation with no context about my intent, and ask it to review the piece as a hostile fact-checker. The fresh conversation matters: a model that has spent an hour helping me build the argument has, in effect, joined my side.
The answer is not automatically correct merely because it is contrary. The value is that it generates tests I might not have invented while defending my original view. Some of the counterarguments will be weak. The one that is not weak is worth the whole exercise.
Separate retrieval, summary, and judgment
These activities can feel like one continuous conversation, but they are different jobs. Retrieval asks what material exists. Summary asks what that material says. Judgment asks what weight it deserves and what conclusion follows. AI can help with all three, but its reliability and authority are not equal across them.
I am comfortable letting a model sort a stack of documents, propose themes, compare definitions, or flag contradictions. I am less comfortable letting it decide which source is credible without checking the reasons: credibility judgments encode assumptions about institutions, incentives, and track records that I want visible, not automated. And I do not delegate the final conclusion, because a conclusion is not only a sentence. It is a commitment to defend the reasoning behind it.
Ask for provenance, not decorative citations
A long list of links can create the appearance of research while hiding the absence of it. I would rather have three sources that directly support three specific claims than thirty links loosely related to the subject. For each important point, I want to know: Where did this come from? Is it a primary or secondary source? What exact passage supports the wording? What would weaken it?
This is also a prompting technique. Asking a model for “sources” produces a bibliography. Asking it to attach one source to one sentence, and to quote the passage it believes supports that sentence, produces something I can actually check, and frequently reveals that the support was thinner than the confident summary implied. The discipline is not more links. It is a tighter chain between each claim and its evidence.
If I cannot answer those questions, the claim is not ready to publish. The model may have helped me discover it, but discovery is not verification.
Protect what should not be uploaded
Research discipline also includes data discipline. Private client information, confidential business records, unpublished personal material, and credentials do not belong in a tool merely because uploading them would be convenient. Before I use AI on a document, I ask whether I have the right to share it with that system and whether the system’s data controls fit the sensitivity of the material.
In my real estate work this is not theoretical. Transaction files contain financial details, personal circumstances, and negotiating positions that belong to clients, not to me. The question “would this document’s owner be comfortable with where I am about to send it?” is a better filter than any feature list. Convenience can make boundaries feel old-fashioned. The consequences of ignoring them are not.
Keep the weight of authorship
AI assists with outlining, organization, transcription, comparison, editing, and source discovery on this site. It may help me see a question faster or express an answer more clearly. The conclusions (what I believe, what I publish, and what I am willing to correct) remain human work.
Responsibility cannot be delegated to a system that cannot be held responsible.
The tools will improve. Some of today’s failure modes will become less common, and new ones will appear, likely subtler, because the surface polish improves faster than the underlying reliability. The durable advantage is not memorizing a list of model weaknesses. It is maintaining habits that survive changes in the technology: check the source, match the claim, invite the counterargument, protect sensitive material, and keep judgment where accountability lives.
Used that way, AI does not replace thinking. It clears away some of the mechanical work so more time can be spent doing it.
Books & further reading
Affiliate disclosure: As an Amazon Associate, I earn from qualifying purchases. I recommend these books because they are relevant to the subject, not because of the commission.
- Co-Intelligence: Living and Working with AI — Ethan Mollick. A practical account of working with AI as a collaborator while remaining attentive to its uneven capabilities and boundaries.
- AI Snake Oil — Arvind Narayanan and Sayash Kapoor. A skeptical guide to separating useful AI systems from inflated claims, weak evaluations, and products that promise more than they can deliver.
