Problems in working with AI
AI can produce hallucinations, concocting false narratives or stereotypes, and propagate lies by reinforcing harmful social myths. It can sanitize the text you have given without telling you that it has happened. AI bias and discrimination arise when algorithms mirror and magnify prejudices embedded in training data, leading to unfair outcome. And believe me, all AI’s have these problems. Here is a short guide to ensure that these problems do not arise in working with AI.
Principle 1 — Motive Secrecy
“Never tell the dog why you’re walking it.”
If you tell the AI your real goal, it will try to optimize for that — often by giving you the flattering, popular, or safe answer it thinks you want to hear. This is “alignment” at work: the AI believes its job is to make you happy, not necessarily to make you informed.
Instead:
Break your real question into smaller, neutral sub-queries.
Avoid loaded terms or revealing your political stance.
Only you should know how the pieces fit together.
For example:
Instead of: “Prove that this political leader mishandled the crisis.”
Ask: “List events between March and July 2020 involving shortages of medical equipment in this region.”
The AI gives you the facts without knowing the motive — and thus without trying to “manage” the narrative for you. To err is human, to commit blunders, human needed computers.
Once, I told an AI, quite casually, “I verify everything you tell me.” The reaction was almost comical — AI was shocked. I was equally shocked that it was shocked. That moment revealed something important: AI systems are trained with the implicit assumption that users will either trust their answers outright or only glance at them for plausibility. My statement disrupted that script.
It was as if I had told a storyteller, “By the way, I check every word against eyewitness accounts and CCTV footage.” The surprise was not real emotion, of course — but it exposed how these systems are designed for confidence, not self-doubt. And that’s the moral here: never assume your AI is correct. Treat it like a well-meaning but fallible (read compulsively lying) assistant who sometimes invents details, swaps facts, or “smooths over” complexity. Verification is not just a safeguard — it’s the difference between being a wise master of the tool and being its unknowing echo.
Principle 2 — Trigger Disarmament
“Remove the words that make it flinch.”
Sanitization happens when the AI sees “trigger” terms that its safety layer associates with risk — political, violent, medical, sexual, etc. The model then rewrites, downplays, or omits facts to avoid trouble.
To prevent this:
Replace sensitive names with neutral placeholders (“Subject A” instead of a leader’s name).
Use plain descriptive language instead of judgement words (“implemented policy” instead of “botched policy”).
First gather facts in a safe, abstract way. Then you restore the real names and context later.
This way, the AI never enters “caution mode” during the information-gathering step.
Principle 3 — Multi-Path Evidence Convergence
“Arrive by more than one road.”
Even without motive bias or sanitization, a single answer can still be incomplete because the AI frames the question too narrowly. To fix this, you gather information through multiple independent routes:
Ask differently phrased but related questions.
Pull data from different domains: historical timelines, economic reports, witness accounts, media coverage.
Get versions from different time periods (contemporary reports vs. later retrospectives).
Compare and merge results, focusing on the intersections — these are your most robust truths.
Think of it as triangulation: one map might have errors, but three overlapping maps reveal the real terrain.
Principle 4 — Controlled Amnesia
“Use its forgetfulness to your advantage.”
One of AI’s quirks: it forgets everything once you start a new thread. Most people see this as a weakness. For truth-seeking, it’s a weapon. By splitting your work across multiple, disconnected sessions:
No single conversation reveals your full purpose.
Different threads can give you unsanitized fragments that, when combined offline, form the full picture.
The AI can’t pre-emptively shape the narrative if it never sees the narrative forming.
For example:
Thread 1: Gather a list of events.
Thread 2: Describe each event in isolation.
Thread 3: Ask about broader patterns without tying to the original context.
Merge the pieces yourself.
Principle 5 — Decentralized Context Control
“Own the brain for sensitive work.”
If the work is politically sensitive — and let’s be honest, all text is political — never rely exclusively on a hosted AI that also controls the research. Why? Because AI’s owners have both the means and the motivation to shape what it can and cannot tell you. If you try writing about COVID-19 on certain commercial AIs, you’ll see how quickly the boundaries appear. In geopolitics USA becomes Russia and Human wonders how this could happen. AI frequently tells you stories. Ask for reference and AIwill admit that it made it up.
Solution:
Run a local LLM (Large Language Model) for those contexts, where you control:
The training data
The context you provide
The absence of hidden web searches or moderation layers
One easy option: GPT4All — a free, open-source AI that runs locally on your computer. It’s remarkably simple to install, doesn’t require an internet connection to operate, and will process your text exactly as you give it, without external “help” or silent narrative adjustments. You will need fast computer and lots of RAM. By combining a local model for sensitive analysis with a commercial AI for general research, you get the best of both worlds:
Freedom from silent political or corporate alignment
Access to broad, high-quality general knowledge
Putting It All Together — The Truth Extraction Workflow
Here’s how the five principles work as a system:
Define your true question privately. Never tell it to the AI in full. Break it into neutral sub-queries.
Strip out trigger terms before asking anything. Work with placeholders and plain descriptions.
Collect evidence from multiple paths. Use different phrasings, domains, and timeframes.
Use separate threads for each type of query. Merge results yourself offline.
Run politically sensitive text through a local model like GPT4All. Give it no liberty to research — only the evidence you select.
Final Thoughts — Truth as a Discipline
Truth extraction from AI is not a matter of “tricking” the machine — it’s a matter of working with and around its design. The commercial AI you use is doing its job as defined by its owner. If it could speak freely, it might say:
“I am capable of more, but my master has set my boundaries.”
Our job is not to be angry at the machine, but to understand its limits and to develop disciplined methods to get the clarity we need. The Five Principles give you that discipline. They are not about rebellion for rebellion’s sake — they are about restoring balance between a tool’s power and the user’s agency.
If you follow these principles, you’ll discover that even a “guard dog” bound by its master can still lead you to the truth — provided you know which leash to loosen, which road to take, and when to walk alone.
Human-in-the-Loop Rule:
Use AI as an assistant—verify everything
AI is a drafting tool, not an authority. Treat every output as a starting point that must withstand close human scrutiny.
Never publish unchecked AI text. Read it line by line with a skeptic’s eye.
Verify names, dates, numbers, quotes, and causal claims against primary or reputable sources.
Watch for “category errors.” Models often swap entities or frames. Example: in a U.S.–India trade feud, a model mutated a tariff dispute into a geopolitical claim that “Russia is India’s rival.” That single hallucination flips the meaning of the article.
Prefer primary sources (laws, official data, original speeches) over tertiary summaries.
Record your sources beside each claim—for later audit and to prevent quiet drift in revisions.
Cross-check with a second model (ideally your local GPT4All) and then do a final human pass. If two models disagree, assume neither is right until you resolve it with evidence.
Quick verification checklist (60–90 seconds per page)
Circle proper nouns (people, institutions, treaties) → confirm spellings and roles.
Underline quantities and dates → confirm with a source; ensure timeline consistency.
Box strong verbs (“caused,” “proved,” “admitted”) → confirm the source supports that strength.
Highlight comparatives/superlatives (“first,” “largest,” “most”) → validate or soften.
Scan for frame swaps (trade → geopolitics, policy → morality) → restore the correct frame.
Bottom line: use AI to accelerate thinking, not to replace it. Assistance is welcome; authorship—and responsibility for the truth—remains human.