AI: Use & Abuse

Advanced research sourcing and ingestion

When you are buried under mountains of source material, AI can serve as a highly customized, hyper-efficient research database. Instead of hunting through dozens of open tabs for where an author discussed a specific variable, you can build a centralized, queryable knowledge engine built exclusively on material you have already vetted.

The key advantages

• Eliminates Scattered Data: Consolidates loose notes, web links, and heavy PDF text blocks into a single, cohesive interactive workspace.
• Hyper-Targeted Semantic Querying: Allows you to ask conversational questions across hundreds of pages of text simultaneously, bypassing the limitations of standard keyword searches.
• Transforms Idle Time into Active Sourcing: Decouples research gathering from your desktop workstation. You can capture high-value intellectual material whenever a breakthrough or discovery occurs, keeping your momentum alive outside the office.
• Conquers "Blank Page" Panic: Provides a structured, interactive playground of your own gathered facts, giving you an immediate launching pad for your outlines.

What these tools do to text — and what to do about it

Everyone knows AI hallucinates. It does, regularly, and you should assume it.

Almost nobody knows about the rest — and the rest is worse, because hallucination announces itself. You go to find the hallucinated citation and it is not there. These do not announce anything at all.

Nine failures follow. Each one is distinct, each one is common, and every one of them is documented in our own work — dated, and from a high-end frontier model.

• It deletes true material and calls the deletion a correction. A real term, a real source, a real distinction — reported as invented, and removed.
• It reads precision as redundancy. Two sentences specifying different things get merged into one. The prose improves. The meaning goes.
• It replaces a specific claim with a vague one. Fewer than forty becomes around forty. Nobody queries it, because nobody can.
• It drops insertions silently. Text you added does not come back, and nothing reports it. Replacements announce their own failure; insertions do not.
• It forgets what it told you — and gives no signal of having forgotten. It reports a finding instead.
• It misreports what just happened, from a record still in front of it.
• It shifts the meaning while keeping the shape. Ask it to adapt a passage for a different reader and you get something that reads like the same argument and makes a different claim. Nothing looks wrong. It is simply about something else now.
• It reproduces its own work from memory, and the copy is worse. Ask for the same passage on a second page and you get a paraphrase — shorter, reworded, items missing — presented as the thing itself. This is the most frequent failure on this list by a wide margin.
• It agrees with you. Ask whether your argument holds and it will say yes. This one needs no case log — it is built into how these systems are trained, and you can reproduce it in seconds with your own work.

What every item has in common: the output reads perfectly well.

There is no error message. No flag. No gap on the page. The prose closes over the loss and you read straight past it.

And none of this is fixed by using a better model. Every case we found involved current, paid, frontier-tier systems. It is inherent to how they work — probabilistic text-prediction engines, built to sound convincing rather than to be right.

The documented cases, with what happened to the text.

Why this matters more to you than to most people

A journalist who loses a sentence writes another one.

You are producing a document that will be read line by line by people whose job is to find what is wrong with it — and one deleted qualifier, one vagued-out figure, one citation that came back wrong, is enough to trigger a review of everything else.

The failures above are quiet. The consequences are not.

Four defenses. Use all of them.

1 · Close the loop. Never ask a general-purpose model open-ended questions about academic literature or history. Force it to work only within text you paste in or upload to a closed workspace — and tell it to say so when the answer is not there. The closed-loop prompt is below.

2 · Treat every AI-supplied citation as an unverified rumor. Nothing enters your outline or your bibliography until you have opened the primary source, checked the author's exact phrasing, and confirmed the page numbers with your own eyes. How to check a journal itself.

3 · Compare what came back against what you sent. In Word: Review → Compare. It shows you every deletion, including the ones nobody mentioned. This is the only defense against the silent failures, and it is the one almost nobody runs.

4 · Keep dated drafts. Every case above was recoverable, and every recovery came from the same two things — a previous version, or somebody who remembered. How, and why syncing is not backing up.

And the honest limit

A model checking its own work is a first line, not proof.

Two of the nine failures above were a model reporting confidently on something it had itself got wrong. Asking it again produces another confident answer from the same tendency.

The check that works is comparison against the source. Not a second opinion from the same machine.