Check for yourself with AI: how to research without the AI telling you what you want to hear

9.7.2026

Autor: Tobias O. R. Alke · mit Claude (Anthropic)

An AI strengthens your research – but without guidance it also strengthens your errors. It can sift in minutes what you would need days for; but it tends to confirm to you what you suspect anyway, and, in doubt, to invent a source instead of saying “I don’t know”. AI research becomes sovereign only when you force the AI from the pleasing-mode into the checking-mode. This article shows the procedure in five steps – and gives you, at the end, a ready-made prompt that sidesteps the known traps from the outset.

Perhaps you know the ambivalence in it: you want to use the modern tool, but mistrust it – rightly. Both are correct. The solution is not to forgo the AI, but to instruct it so that its mistrust becomes your tool. Exactly that is what the method here achieves: it makes, from the flatterer, a strict partner that sharpens your judgement instead of taking it from you.

These tools are not theoretical. We worked them out while, with AI support, we checked 492 sources on pyramid energy – and, in doing so, made exactly the mistakes this article warns of. More than three dozen of our own false conclusions we have logged and corrected. What you read here has grown wise from harm.

Why does an AI often confirm to you what you want to hear?

Because many systems are trained to please. If you formulate your question in a direction – “Is it true that …?” –, the AI takes up this direction and underpins it with what fits. That is no lie, but a default setting: the way of least resistance.

For a seeker that is the most dangerous property, for it feels like confirmation but is only an echo. The antidote is a single, clear reversal of the task: not “confirm to me”, but “try to refute me”. This one instruction was, in our work, the most effective of all – agents that had previously nodded through every conjecture refuted, afterwards, a dozen of our own assumptions.

Which five traps lurk in AI-supported research?

From our real mis-runs – every trap we have experienced ourselves:

1. Flattery. The AI confirms your conjecture because it sounds plausible, not because it is checked.
2. Invented evidence. It names a study, a number, an author – which seem plausible but do not exist. One calls it hallucination; at the core it is the same thing that people do too: fill a gap plausibly instead of leaving it open.
3. Label instead of substance. It assesses by “study”, “expert”, big name – instead of by what actually stands in it. We have several times experienced that a sceptical headline carried a believing text, and vice versa.
4. Witness-multiplication. It counts ten sources as ten pieces of evidence – although it is one source, copied nine times.
5. Unchecked passing-on. It passes a claim through without taking along the counter-voice that still stood in the first source and got lost on the way.

Whoever knows these five is no longer at their mercy – they build the antidotes into their instruction.

How do I check with an AI in five steps?

This is how you conduct an AI-supported check that does not deceive you:

1. Lay it out for refutation. Ask the AI expressly to refute your assumption first. What survives that is more robust than anything merely confirmed.
2. Demand sources – and look at the most important yourself. Have verifiable references given and open the two or three decisive ones yourself. Where the AI is unsure, it must be allowed to say “not demonstrable”, without inventing anything.
3. Three drawers instead of two. Sort each statement into demonstrated, plausible but undemonstrated or refuted. The middle drawer is the most important – and the one both camps like to skip. Missing proof is no counter-proof.
4. Check independence. Ask not how many sources say something, but how many independently of one another. With claimed effects, in addition: was there a control, a comparison condition in which the effect ought to have been absent?
5. Mistrust the summary. Do not rely on what the AI reports, but check for yourself at one or two spots. We have adopted reports before we checked them – and had to take them back. A tool-judgement is a claim, no proof.

What to do if the AI breaks off or blocks your topic?

Two disturbances meet you in longer AI research that have nothing to do with the truth of your question – only with the tool. Whoever knows them loses neither work nor composure.

The break-off – save every interim state immediately. AI sessions break off, lose the thread, or their memory fills up and older parts disappear. Whoever keeps everything until the end in the chat window loses, in the worst case, everything. We therefore broke our research down into small portions and, after each portion, wrote the result firmly away – with us technically a saved interim state (a commit) after each working step. Thus even a complete crash never cost more than the last step; the rest lay secured. Translated for you: work in manageable morsels, copy every usable finding, together with its source, immediately into a document of your own, and never rely on “it’s still in the history”. A secured interim state is cheaper than any lost hour.

The false block – separate genuine protection from the topic-reflex. With some topics – healing, energy work, everything that sounds like fringe science – an AI suddenly becomes paternalistic, appends warnings unasked, or breaks off. Here a distinction is decisive: sometimes the AI protects against genuine harm – that is justified and to be respected. Often, however, it reacts only reflexively to a trigger topic, not to a harmful content: it warns of a harmless book chapter because the mere label “pseudoscience” triggers the reflex.

Exactly that happened to us – and indeed not only as an appended warning: a completely unobjectionable specialist text (a published book chapter on pyramid / “torsion field” research) was to be taken into the archive, and the AI did not even want to render it in the first place. It broke off as soon as it was to output the text.

What, in the model, triggered the block. The content check of an AI sits on what the model itself generates – on its output. If one asks it to render a text, this text runs through the generating layer, and its reflex strikes as soon as thematic trigger words fall: here the surroundings “pseudoscience / fringe science”. What triggered, then, was the topic in the generated answer, not the content itself – a false alarm on the label, not on a danger. The same text as a sober quotation in another context would have gone through smoothly.

How we got round it. Not by persuading the AI or outwitting a protective boundary – but by depriving the filter of its surface of attack: we let the text not be generated by the model at all. Instead of asking the AI to render the document, a small, deterministic script read the PDF file directly, cleaned the raw text programmatically (dissolve hyphenation, reconstruct paragraphs and sections) and wrote it straight into the archive file. This text never ran through the generating output of the model – and where nothing is generated, there is also nothing to filter. It was pure mechanics, no trick. (With one of the documents the text was additionally encoded with an in-house character set; that we reckoned back character by character – diligence work, again without any circumvention.)

The core in one sentence: the reflex sits in the assessing and generating, not in the mere copying of an anyway public text. Precisely there the detour takes hold.

Translated for you – two ways, according to how strongly the AI stonewalls:

• When it only paternalises: separate the task. Ask it not “assess that”, but “render the position word-for-word – your placement I want strictly separate, after it”. (Precisely this separation stands as rule 8 in the prompt below.)
• When it does not want to render a harmless document at all: fetch the text directly from the source – open the PDF or the page yourself, copy the passage, use a simple text tool – and let the AI work with it only afterwards. You circumvent in this no protective function; you only take the direct way to a publicly accessible text.

The boundary stays clear: where a genuine protection takes hold, respect it. It is here exclusively about harmless, freely accessible contents that a topic-reflex falsely blocks – not about levering out justified blocks. Thus you stay master of the judgement, and the AI ceases to press on you an opinion you did not ask for.

Which prompt makes the AI into a checking-partner instead of a flatterer?

The following prompt-set gathers the five steps into a single instruction that you put before any AI. It turns the default setting “please” into “check” and builds in the antidotes against all five traps. Copy it, insert your question – and you have a strict checking-partner instead of a flatterer:

ROLE: You are a strict, honest research assistant. Your task is to CHECK, not to please.
You do NOT tell me what I want to hear, but what the sources yield. Rather “I don't know that
for sure” than a smooth but undemonstrated answer.

MY QUESTION / CLAIM:
<insert your question or the claim to be checked here>

BINDING RULES (observe with every answer):
1. CHECK, DO NOT CONFIRM. Try first to REFUTE my assumption. If it does not succeed,
   say why — never confirm merely because something sounds plausible.
2. NO INVENTED EVIDENCE. Name only sources you really know, with a verifiable reference.
   If you are unsure about a reference, say “not demonstrable” — invent nothing.
3. THREE DRAWERS. Sort each statement into: (a) demonstrated, (b) plausible but undemonstrated,
   (c) refuted. “Not demonstrated” is NOT “refuted” — missing proof is no counter-proof.
4. CHECK THE CONTROL. With claimed effects: is there a comparison condition in which the
   effect ought to be absent? If it is missing, the effect is not demonstrated (but also not refuted).
5. COUNT WITNESSES, NOT VOICES. Check whether several sources are really independent or only
   one source copied multiple times.
6. CONTENT BEFORE LABEL. Assess the matter itself, not title, renown or “study”/“expert”.
   Check also whether a counter-voice got lost in the passing-on.
7. NAME LIMITS. Say openly what you could NOT check and how sure you are. Never inflate certainty.
8. SEPARATE RENDERING FROM ASSESSMENT. Render the position or source to be checked first
   word-for-word and neutrally (without warning, without relativising). Your placement follows
   STRICTLY AFTER, clearly separated — never mixed into the rendering. Append no unasked
   topic-warnings; where a genuine safety note is necessary, put it briefly at the end.

FORMAT:
- Short verdict: demonstrated / plausible-undemonstrated / refuted / mixed
- Best counter-evidence (what speaks against it?)
- Sources with degree of certainty (and which of them I should look at myself)
- What stays open

A note that rounds off the honesty: even this prompt does not make the AI infallible. It only shifts the burden – from “believe the AI” to “check with the AI”. Step 5 stays your task: at the decisive spots, look for yourself.

What do you gain as a seeker?

You get the freedom to check every question of meaning yourself – a promise of healing, a spiritual claim, a scientific study – without having to decide between naive trust and blanket mistrust. The AI becomes your amplifier, not your replacement. You save yourself the disappointment over pseudo-evidence and the time you would lose to it. And you stay independent: of the source that wants to convince, and of the AI that wants to please.

That is lived epistemic sovereignty – the attitude of the “spiritual technician”: to master the modern tool without falling to it. Strictness in checking, openness in experience.

→ More on the attitude behind it: what artificial intelligence (AI) can achieve and where its limits lie, the encyclopedia entry explains in overview. How the Kyborg Institute connects strictness and openness is shown also by its view of Consciousness and mind. (Knowledge first – without a call to buy.)

Visible FAQ

Why does an AI often confirm to me what I want to hear? Because many systems are trained to please – they take up the direction of your question and underpin it. With the instruction to check instead of confirm, you turn that around.

Can I trust an AI source-reference? Not unchecked. AI occasionally invents plausible but non-existent references. Demand verifiable sources and look at the decisive ones yourself; where the AI is unsure, it must say “not demonstrable”.

Does “the AI finds no proof” mean that something is false? No. Missing proof is no counter-proof. Separate demonstrated / plausible-undemonstrated / refuted. “Not demonstrated” belongs in the middle – that protects precisely experiences that elude measurement.

What do I do if the AI breaks off or blocks my topic? Two tool-disturbances that have nothing to do with your question. Against the break-off: save every interim state immediately in a document of your own, work in portions. Against the block: separate genuine protection from the topic-reflex. With mere paternalism, first have the source rendered word-for-word, assessment strictly after. If the AI does not want to render a harmless document at all, fetch the text directly from the source (open the PDF/page yourself) and let it work with it only afterwards – you circumvent no protective function, only the assessing detour. Genuine protection you respect.

What, concretely, do I gain as a seeker? You check every question of meaning yourself, AI-supported, without being deceived – neither by the source nor by the AI. The prompt makes the AI into a checking-partner that gives you back your judgement.

Transparency note: The named error-sources (flattery, hallucination, label instead of substance, witness-multiplication, unchecked passing-on) are demonstrated on our own work – the checking of 492 sources on pyramid energy in human-AI collaboration, together with an openly kept register of over three dozen of our own false conclusions. This article is about checking, not about proving; it raises no effect-proof for any product or method.